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 inline uint8_t
 522 bbr_state_val(struct tcp_bbr *bbr)
 523 {
 524         return(bbr->rc_bbr_substate);
 525 }
 526
 527 static inline uint32_t
 528 get_min_cwnd(struct tcp_bbr *bbr)
 529 {
 530         int mss;
 531
 532         mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
 533         if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED)
 534                 return (bbr_cwnd_min_val_hs * mss);
 535         else
 536                 return (bbr_cwnd_min_val * mss);
 537 }
 538
 539 static uint32_t
 540 bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr)
 541 {
 542         uint64_t srtt, var;
 543         uint64_t ret_val;
 544
 545         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
 546         if (tp->t_srtt == 0) {
 547                 srtt = (uint64_t)BBR_INITIAL_RTO;
 548                 var = 0;
 549         } else {
 550                 srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
 551                 var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT);
 552         }
 553         TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]),
 554             bbr_persist_min, bbr_persist_max);
 555         return ((uint32_t)ret_val);
 556 }
 557
 558 static uint32_t
 559 bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
 560 {
 561         /*
 562          * Start the FR timer, we do this based on getting the first one in
 563          * the rc_tmap. Note that if its NULL we must stop the timer. in all
 564          * events we need to stop the running timer (if its running) before
 565          * starting the new one.
 566          */
 567         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
 568         int32_t idx;
 569         int32_t is_tlp_timer = 0;
 570         struct bbr_sendmap *rsm;
 571
 572         if (bbr->rc_all_timers_stopped) {
 573                 /* All timers have been stopped none are to run */
 574                 return (0);
 575         }
 576         if (bbr->rc_in_persist) {
 577                 /* We can't start any timer in persists */
 578                 return (bbr_get_persists_timer_val(tp, bbr));
 579         }
 580         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 581         if ((rsm == NULL) ||
 582             ((tp->t_flags & TF_SACK_PERMIT) == 0) ||
 583             (tp->t_state < TCPS_ESTABLISHED)) {
 584                 /* Nothing on the send map */
 585 activate_rxt:
 586                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
 587                         uint64_t tov;
 588
 589                         time_since_sent = 0;
 590                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 591                         if (rsm) {
 592                                 idx = rsm->r_rtr_cnt - 1;
 593                                 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
 594                                         tstmp_touse = rsm->r_tim_lastsent[idx];
 595                                 else
 596                                         tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
 597                                 if (TSTMP_GT(tstmp_touse, cts))
 598                                     time_since_sent = cts - tstmp_touse;
 599                         }
 600                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
 601                         if (tp->t_srtt == 0)
 602                                 tov = BBR_INITIAL_RTO;
 603                         else
 604                                 tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) +
 605                                     ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT);
 606                         if (tp->t_rxtshift)
 607                                 tov *= tcp_backoff[tp->t_rxtshift];
 608                         if (tov > time_since_sent)
 609                                 tov -= time_since_sent;
 610                         else
 611                                 tov = bbr->r_ctl.rc_min_to;
 612                         TCPT_RANGESET_NOSLOP(to, tov,
 613                             (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC),
 614                             (bbr->rc_max_rto_sec * USECS_IN_SECOND));
 615                         bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to);
 616                         return (to);
 617                 }
 618                 return (0);
 619         }
 620         if (rsm->r_flags & BBR_ACKED) {
 621                 rsm = bbr_find_lowest_rsm(bbr);
 622                 if (rsm == NULL) {
 623                         /* No lowest? */
 624                         goto activate_rxt;
 625                 }
 626         }
 627         /* Convert from ms to usecs */
 628         if (rsm->r_flags & BBR_SACK_PASSED) {
 629                 if ((tp->t_flags & TF_SENTFIN) &&
 630                     ((tp->snd_max - tp->snd_una) == 1) &&
 631                     (rsm->r_flags & BBR_HAS_FIN)) {
 632                         /*
 633                          * We don't start a bbr rack timer if all we have is
 634                          * a FIN outstanding.
 635                          */
 636                         goto activate_rxt;
 637                 }
 638                 srtt = bbr_get_rtt(bbr, BBR_RTT_RACK);
 639                 thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm);
 640                 idx = rsm->r_rtr_cnt - 1;
 641                 exp = rsm->r_tim_lastsent[idx] + thresh;
 642                 if (SEQ_GEQ(exp, cts)) {
 643                         to = exp - cts;
 644                         if (to < bbr->r_ctl.rc_min_to) {
 645                                 to = bbr->r_ctl.rc_min_to;
 646                         }
 647                 } else {
 648                         to = bbr->r_ctl.rc_min_to;
 649                 }
 650         } else {
 651                 /* Ok we need to do a TLP not RACK */
 652                 if (bbr->rc_tlp_in_progress != 0) {
 653                         /*
 654                          * The previous send was a TLP.
 655                          */
 656                         goto activate_rxt;
 657                 }
 658                 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
 659                 if (rsm == NULL) {
 660                         /* We found no rsm to TLP with. */
 661                         goto activate_rxt;
 662                 }
 663                 if (rsm->r_flags & BBR_HAS_FIN) {
 664                         /* If its a FIN we don't do TLP */
 665                         rsm = NULL;
 666                         goto activate_rxt;
 667                 }
 668                 time_since_sent = 0;
 669                 idx = rsm->r_rtr_cnt - 1;
 670                 if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time))
 671                         tstmp_touse = rsm->r_tim_lastsent[idx];
 672                 else
 673                         tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time;
 674                 if (TSTMP_GT(tstmp_touse, cts))
 675                     time_since_sent = cts - tstmp_touse;
 676                 is_tlp_timer = 1;
 677                 srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use);
 678                 thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts);
 679                 if (thresh > time_since_sent)
 680                         to = thresh - time_since_sent;
 681                 else
 682                         to = bbr->r_ctl.rc_min_to;
 683                 if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
 684                         /*
 685                          * If the TLP time works out to larger than the max
 686                          * RTO lets not do TLP.. just RTO.
 687                          */
 688                         goto activate_rxt;
 689                 }
 690                 if ((bbr->rc_tlp_rtx_out == 1) &&
 691                     (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) {
 692                         /*
 693                          * Second retransmit of the same TLP
 694                          * lets not.
 695                          */
 696                         bbr->rc_tlp_rtx_out = 0;
 697                         goto activate_rxt;
 698                 }
 699                 if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) {
 700                         /*
 701                          * The tail is no longer the last one I did a probe
 702                          * on
 703                          */
 704                         bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
 705                         bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
 706                 }
 707         }
 708         if (is_tlp_timer == 0) {
 709                 BBR_STAT_INC(bbr_to_arm_rack);
 710                 bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
 711         } else {
 712                 bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to);
 713                 if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
 714                         /*
 715                          * We have exceeded how many times we can retran the
 716                          * current TLP timer, switch to the RTO timer.
 717                          */
 718                         goto activate_rxt;
 719                 } else {
 720                         BBR_STAT_INC(bbr_to_arm_tlp);
 721                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
 722                 }
 723         }
 724         return (to);
 725 }
 726
 727 static inline int32_t
 728 bbr_minseg(struct tcp_bbr *bbr)
 729 {
 730         return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options);
 731 }
 732
 733 static void
 734 bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len)
 735 {
 736         struct inpcb *inp;
 737         struct hpts_diag diag;
 738         uint32_t delayed_ack = 0;
 739         uint32_t left = 0;
 740         uint32_t hpts_timeout;
 741         uint8_t stopped;
 742         int32_t delay_calc = 0;
 743         uint32_t prev_delay = 0;
 744
 745         inp = tp->t_inpcb;
 746         if (inp->inp_in_hpts) {
 747                 /* A previous call is already set up */
 748                 return;
 749         }
 750         if ((tp->t_state == TCPS_CLOSED) ||
 751             (tp->t_state == TCPS_LISTEN)) {
 752                 return;
 753         }
 754         stopped = bbr->rc_tmr_stopped;
 755         if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) {
 756                 left = bbr->r_ctl.rc_timer_exp - cts;
 757         }
 758         bbr->r_ctl.rc_hpts_flags = 0;
 759         bbr->r_ctl.rc_timer_exp = 0;
 760         prev_delay = bbr->r_ctl.rc_last_delay_val;
 761         if (bbr->r_ctl.rc_last_delay_val &&
 762             (slot == 0)) {
 763                 /*
 764                  * If a previous pacer delay was in place we
 765                  * are not coming from the output side (where
 766                  * we calculate a delay, more likely a timer).
 767                  */
 768                 slot = bbr->r_ctl.rc_last_delay_val;
 769                 if (TSTMP_GT(cts, bbr->rc_pacer_started)) {
 770                         /* Compensate for time passed  */
 771                         delay_calc = cts - bbr->rc_pacer_started;
 772                         if (delay_calc <= slot)
 773                                 slot -= delay_calc;
 774                 }
 775         }
 776         /* Do we have early to make up for by pushing out the pacing time? */
 777         if (bbr->r_agg_early_set) {
 778                 bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2);
 779                 slot += bbr->r_ctl.rc_agg_early;
 780                 bbr->r_ctl.rc_agg_early = 0;
 781                 bbr->r_agg_early_set = 0;
 782         }
 783         /* Are we running a total debt that needs to be compensated for? */
 784         if (bbr->r_ctl.rc_hptsi_agg_delay) {
 785                 if (slot > bbr->r_ctl.rc_hptsi_agg_delay) {
 786                         /* We nuke the delay */
 787                         slot -= bbr->r_ctl.rc_hptsi_agg_delay;
 788                         bbr->r_ctl.rc_hptsi_agg_delay = 0;
 789                 } else {
 790                         /* We nuke some of the delay, put in a minimal 100usecs  */
 791                         bbr->r_ctl.rc_hptsi_agg_delay -= slot;
 792                         bbr->r_ctl.rc_last_delay_val = slot = 100;
 793                 }
 794         }
 795         bbr->r_ctl.rc_last_delay_val = slot;
 796         hpts_timeout = bbr_timer_start(tp, bbr, cts);
 797         if (tp->t_flags & TF_DELACK) {
 798                 if (bbr->rc_in_persist == 0) {
 799                         delayed_ack = bbr_delack_time;
 800                 } else {
 801                         /*
 802                          * We are in persists and have
 803                          * gotten a new data element.
 804                          */
 805                         if (hpts_timeout > bbr_delack_time) {
 806                                 /*
 807                                  * Lets make the persists timer (which acks)
 808                                  * be the smaller of hpts_timeout and bbr_delack_time.
 809                                  */
 810                                 hpts_timeout = bbr_delack_time;
 811                         }
 812                 }
 813         }
 814         if (delayed_ack &&
 815             ((hpts_timeout == 0) ||
 816              (delayed_ack < hpts_timeout))) {
 817                 /* We need a Delayed ack timer */
 818                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
 819                 hpts_timeout = delayed_ack;
 820         }
 821         if (slot) {
 822                 /* Mark that we have a pacing timer up */
 823                 BBR_STAT_INC(bbr_paced_segments);
 824                 bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
 825         }
 826         /*
 827          * If no timers are going to run and we will fall off thfe hptsi
 828          * wheel, we resort to a keep-alive timer if its configured.
 829          */
 830         if ((hpts_timeout == 0) &&
 831             (slot == 0)) {
 832                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
 833                     (tp->t_state <= TCPS_CLOSING)) {
 834                         /*
 835                          * Ok we have no timer (persists, rack, tlp, rxt  or
 836                          * del-ack), we don't have segments being paced. So
 837                          * all that is left is the keepalive timer.
 838                          */
 839                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
 840                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
 841                         } else {
 842                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
 843                         }
 844                         bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
 845                 }
 846         }
 847         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
 848             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
 849                 /*
 850                  * RACK, TLP, persists and RXT timers all are restartable
 851                  * based on actions input .. i.e we received a packet (ack
 852                  * or sack) and that changes things (rw, or snd_una etc).
 853                  * Thus we can restart them with a new value. For
 854                  * keep-alive, delayed_ack we keep track of what was left
 855                  * and restart the timer with a smaller value.
 856                  */
 857                 if (left < hpts_timeout)
 858                         hpts_timeout = left;
 859         }
 860         if (bbr->r_ctl.rc_incr_tmrs && slot &&
 861             (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
 862                 /*
 863                  * If configured to do so, and the timer is either
 864                  * the TLP or RXT timer, we need to increase the timeout
 865                  * by the pacing time. Consider the bottleneck at my
 866                  * machine as an example, we are sending something
 867                  * to start a TLP on. The last packet won't be emitted
 868                  * fully until the pacing time (the bottleneck will hold
 869                  * the data in place). Once the packet is emitted that
 870                  * is when we want to start waiting for the TLP. This
 871                  * is most evident with hardware pacing (where the nic
 872                  * is holding the packet(s) before emitting). But it
 873                  * can also show up in the network so we do it for all
 874                  * cases. Technically we would take off one packet from
 875                  * this extra delay but this is easier and being more
 876                  * conservative is probably better.
 877                  */
 878                 hpts_timeout += slot;
 879         }
 880         if (hpts_timeout) {
 881                 /*
 882                  * Hack alert for now we can't time-out over 2147 seconds (a
 883                  * bit more than 35min)
 884                  */
 885                 if (hpts_timeout > 0x7ffffffe)
 886                         hpts_timeout = 0x7ffffffe;
 887                 bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
 888         } else
 889                 bbr->r_ctl.rc_timer_exp = 0;
 890         if ((slot) &&
 891             (bbr->rc_use_google ||
 892              bbr->output_error_seen ||
 893              (slot <= hpts_timeout))  ) {
 894                 /*
 895                  * Tell LRO that it can queue packets while
 896                  * we pace.
 897                  */
 898                 bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
 899                 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
 900                     (bbr->rc_cwnd_limited == 0)) {
 901                         /*
 902                          * If we are not cwnd limited and we
 903                          * are running a rack timer we put on
 904                          * the do not disturbe even for sack.
 905                          */
 906                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
 907                 } else
 908                         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
 909                 bbr->rc_pacer_started = cts;
 910
 911                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
 912                                            __LINE__, &diag);
 913                 bbr->rc_timer_first = 0;
 914                 bbr->bbr_timer_src = frm;
 915                 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1);
 916                 bbr_log_hpts_diag(bbr, cts, &diag);
 917         } else if (hpts_timeout) {
 918                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
 919                                            __LINE__, &diag);
 920                 /*
 921                  * We add the flag here as well if the slot is set,
 922                  * since hpts will call in to clear the queue first before
 923                  * calling the output routine (which does our timers).
 924                  * We don't want to set the flag if its just a timer
 925                  * else the arrival of data might (that causes us
 926                  * to send more) might get delayed. Imagine being
 927                  * on a keep-alive timer and a request comes in for
 928                  * more data.
 929                  */
 930                 if (slot)
 931                         bbr->rc_pacer_started = cts;
 932                 if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
 933                     (bbr->rc_cwnd_limited == 0)) {
 934                         /*
 935                          * For a rack timer, don't wake us even
 936                          * if a sack arrives as long as we are
 937                          * not cwnd limited.
 938                          */
 939                         bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
 940                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
 941                 } else {
 942                         /* All other timers wake us up */
 943                         bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
 944                         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
 945                 }
 946                 bbr->bbr_timer_src = frm;
 947                 bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0);
 948                 bbr_log_hpts_diag(bbr, cts, &diag);
 949                 bbr->rc_timer_first = 1;
 950         }
 951         bbr->rc_tmr_stopped = 0;
 952         bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay);
 953 }
 954
 955 static void
 956 bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb)
 957 {
 958         /*
 959          * We received an ack, and then did not call send or were bounced
 960          * out due to the hpts was running. Now a timer is up as well, is it
 961          * the right timer?
 962          */
 963         struct inpcb *inp;
 964         struct bbr_sendmap *rsm;
 965         uint32_t hpts_timeout;
 966         int tmr_up;
 967
 968         tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
 969         if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
 970                 return;
 971         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
 972         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
 973             (tmr_up == PACE_TMR_RXT)) {
 974                 /* Should be an RXT */
 975                 return;
 976         }
 977         inp = bbr->rc_inp;
 978         if (rsm == NULL) {
 979                 /* Nothing outstanding? */
 980                 if (tp->t_flags & TF_DELACK) {
 981                         if (tmr_up == PACE_TMR_DELACK)
 982                                 /*
 983                                  * We are supposed to have delayed ack up
 984                                  * and we do
 985                                  */
 986                                 return;
 987                 } else if (sbavail(&inp->inp_socket->so_snd) &&
 988                     (tmr_up == PACE_TMR_RXT)) {
 989                         /*
 990                          * if we hit enobufs then we would expect the
 991                          * possiblity of nothing outstanding and the RXT up
 992                          * (and the hptsi timer).
 993                          */
 994                         return;
 995                 } else if (((V_tcp_always_keepalive ||
 996                             inp->inp_socket->so_options & SO_KEEPALIVE) &&
 997                             (tp->t_state <= TCPS_CLOSING)) &&
 998                             (tmr_up == PACE_TMR_KEEP) &&
 999                     (tp->snd_max == tp->snd_una)) {
1000                         /* We should have keep alive up and we do */
1001                         return;
1002                 }
1003         }
1004         if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) {
1005                 if ((tp->t_flags & TF_SENTFIN) &&
1006                     ((tp->snd_max - tp->snd_una) == 1) &&
1007                     (rsm->r_flags & BBR_HAS_FIN)) {
1008                         /* needs to be a RXT */
1009                         if (tmr_up == PACE_TMR_RXT)
1010                                 return;
1011                         else
1012                                 goto wrong_timer;
1013                 } else if (tmr_up == PACE_TMR_RACK)
1014                         return;
1015                 else
1016                         goto wrong_timer;
1017         } else if (rsm && (tmr_up == PACE_TMR_RACK)) {
1018                 /* Rack timer has priority if we have data out */
1019                 return;
1020         } else if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1021                     ((tmr_up == PACE_TMR_TLP) ||
1022             (tmr_up == PACE_TMR_RXT))) {
1023                 /*
1024                  * Either a TLP or RXT is fine if no sack-passed is in place
1025                  * and data is outstanding.
1026                  */
1027                 return;
1028         } else if (tmr_up == PACE_TMR_DELACK) {
1029                 /*
1030                  * If the delayed ack was going to go off before the
1031                  * rtx/tlp/rack timer were going to expire, then that would
1032                  * be the timer in control. Note we don't check the time
1033                  * here trusting the code is correct.
1034                  */
1035                 return;
1036         }
1037         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
1038             ((tmr_up == PACE_TMR_RXT) ||
1039              (tmr_up == PACE_TMR_TLP) ||
1040              (tmr_up == PACE_TMR_RACK))) {
1041                 /*
1042                  * We have outstanding data and
1043                  * we *do* have a RACK, TLP or RXT
1044                  * timer running. We won't restart
1045                  * anything here since thats probably ok we
1046                  * will get called with some timer here shortly.
1047                  */
1048                 return;
1049         }
1050         /*
1051          * Ok the timer originally started is not what we want now. We will
1052          * force the hpts to be stopped if any, and restart with the slot
1053          * set to what was in the saved slot.
1054          */
1055 wrong_timer:
1056         if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) {
1057                 if (inp->inp_in_hpts)
1058                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
1059                 bbr_timer_cancel(bbr, __LINE__, cts);
1060                 bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val,
1061                     0);
1062         } else {
1063                 /*
1064                  * Output is hptsi so we just need to switch the type of
1065                  * timer. We don't bother with keep-alive, since when we
1066                  * jump through the output, it will start the keep-alive if
1067                  * nothing is sent.
1068                  *
1069                  * We only need a delayed-ack added and or the hpts_timeout.
1070                  */
1071                 hpts_timeout = bbr_timer_start(tp, bbr, cts);
1072                 if (tp->t_flags & TF_DELACK) {
1073                         if (hpts_timeout == 0) {
1074                                 hpts_timeout = bbr_delack_time;
1075                                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1076                         }
1077                         else if (hpts_timeout > bbr_delack_time) {
1078                                 hpts_timeout = bbr_delack_time;
1079                                 bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK;
1080                         }
1081                 }
1082                 if (hpts_timeout) {
1083                         if (hpts_timeout > 0x7ffffffe)
1084                                 hpts_timeout = 0x7ffffffe;
1085                         bbr->r_ctl.rc_timer_exp = cts + hpts_timeout;
1086                 }
1087         }
1088 }
1089
1090 int32_t bbr_clear_lost = 0;
1091
1092 /*
1093  * Considers the two time values now (cts) and earlier.
1094  * If cts is smaller than earlier, we could have
1095  * had a sequence wrap (our counter wraps every
1096  * 70 min or so) or it could be just clock skew
1097  * getting us two differnt time values. Clock skew
1098  * will show up within 10ms or so. So in such
1099  * a case (where cts is behind earlier time by
1100  * less than 10ms) we return 0. Otherwise we
1101  * return the true difference between them.
1102  */
1103 static inline uint32_t
1104 bbr_calc_time(uint32_t cts, uint32_t earlier_time) {
1105         /*
1106          * Given two timestamps, the current time stamp cts, and some other
1107          * time-stamp taken in theory earlier return the difference. The
1108          * trick is here sometimes locking will get the other timestamp
1109          * after the cts. If this occurs we need to return 0.
1110          */
1111         if (TSTMP_GEQ(cts, earlier_time))
1112                 return (cts - earlier_time);
1113         /*
1114          * cts is behind earlier_time if its less than 10ms consider it 0.
1115          * If its more than 10ms difference then we had a time wrap. Else
1116          * its just the normal locking foo. I wonder if we should not go to
1117          * 64bit TS and get rid of this issue.
1118          */
1119         if (TSTMP_GEQ((cts + 10000), earlier_time))
1120                 return (0);
1121         /*
1122          * Ok the time must have wrapped. So we need to answer a large
1123          * amount of time, which the normal subtraction should do.
1124          */
1125         return (cts - earlier_time);
1126 }
1127
1128 static int
1129 sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS)
1130 {
1131         uint32_t stat;
1132         int32_t error;
1133
1134         error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t));
1135         if (error || req->newptr == NULL)
1136                 return error;
1137
1138         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
1139         if (error)
1140                 return (error);
1141         if (stat == 1) {
1142 #ifdef BBR_INVARIANTS
1143                 printf("Clearing BBR lost counters\n");
1144 #endif
1145                 COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT);
1146                 COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT);
1147                 COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT);
1148         } else if (stat == 2) {
1149 #ifdef BBR_INVARIANTS
1150                 printf("Clearing BBR option counters\n");
1151 #endif
1152                 COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE);
1153         } else if (stat == 3) {
1154 #ifdef BBR_INVARIANTS
1155                 printf("Clearing BBR stats counters\n");
1156 #endif
1157                 COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE);
1158         } else if (stat == 4) {
1159 #ifdef BBR_INVARIANTS
1160                 printf("Clearing BBR out-size counters\n");
1161 #endif
1162                 COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE);
1163         }
1164         bbr_clear_lost = 0;
1165         return (0);
1166 }
1167
1168 static void
1169 bbr_init_sysctls(void)
1170 {
1171         struct sysctl_oid *bbr_probertt;
1172         struct sysctl_oid *bbr_hptsi;
1173         struct sysctl_oid *bbr_measure;
1174         struct sysctl_oid *bbr_cwnd;
1175         struct sysctl_oid *bbr_timeout;
1176         struct sysctl_oid *bbr_states;
1177         struct sysctl_oid *bbr_startup;
1178         struct sysctl_oid *bbr_policer;
1179
1180         /* Probe rtt controls */
1181         bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1182             SYSCTL_CHILDREN(bbr_sysctl_root),
1183             OID_AUTO,
1184             "probertt",
1185             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1186             "");
1187         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1188             SYSCTL_CHILDREN(bbr_probertt),
1189             OID_AUTO, "gain", CTLFLAG_RW,
1190             &bbr_rttprobe_gain, 192,
1191             "What is the filter gain drop in probe_rtt (0=disable)?");
1192         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1193             SYSCTL_CHILDREN(bbr_probertt),
1194             OID_AUTO, "cwnd", CTLFLAG_RW,
1195             &bbr_rtt_probe_cwndtarg, 4,
1196             "How many mss's are outstanding during probe-rtt");
1197         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1198             SYSCTL_CHILDREN(bbr_probertt),
1199             OID_AUTO, "int", CTLFLAG_RW,
1200             &bbr_rtt_probe_limit, 4000000,
1201             "If RTT has not shrank in this many micro-seconds enter probe-rtt");
1202         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1203             SYSCTL_CHILDREN(bbr_probertt),
1204             OID_AUTO, "mintime", CTLFLAG_RW,
1205             &bbr_rtt_probe_time, 200000,
1206             "How many microseconds in probe-rtt");
1207         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1208             SYSCTL_CHILDREN(bbr_probertt),
1209             OID_AUTO, "filter_len_sec", CTLFLAG_RW,
1210             &bbr_filter_len_sec, 6,
1211             "How long in seconds does the rttProp filter run?");
1212         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1213             SYSCTL_CHILDREN(bbr_probertt),
1214             OID_AUTO, "drain_rtt", CTLFLAG_RW,
1215             &bbr_drain_rtt, BBR_SRTT,
1216             "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?");
1217         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1218             SYSCTL_CHILDREN(bbr_probertt),
1219             OID_AUTO, "can_force", CTLFLAG_RW,
1220             &bbr_can_force_probertt, 0,
1221             "If we keep setting new low rtt's but delay going in probe-rtt can we force in??");
1222         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1223             SYSCTL_CHILDREN(bbr_probertt),
1224             OID_AUTO, "enter_sets_force", CTLFLAG_RW,
1225             &bbr_probertt_sets_rtt, 0,
1226             "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?");
1227         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1228             SYSCTL_CHILDREN(bbr_probertt),
1229             OID_AUTO, "can_adjust", CTLFLAG_RW,
1230             &bbr_can_adjust_probertt, 1,
1231             "Can we dynamically adjust the probe-rtt limits and times?");
1232         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1233             SYSCTL_CHILDREN(bbr_probertt),
1234             OID_AUTO, "is_ratio", CTLFLAG_RW,
1235             &bbr_is_ratio, 0,
1236             "is the limit to filter a ratio?");
1237         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1238             SYSCTL_CHILDREN(bbr_probertt),
1239             OID_AUTO, "use_cwnd", CTLFLAG_RW,
1240             &bbr_prtt_slam_cwnd, 0,
1241             "Should we set/recover cwnd?");
1242         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1243             SYSCTL_CHILDREN(bbr_probertt),
1244             OID_AUTO, "can_use_ts", CTLFLAG_RW,
1245             &bbr_can_use_ts_for_rtt, 1,
1246             "Can we use the ms timestamp if available for retransmistted rtt calculations?");
1247
1248         /* Pacing controls */
1249         bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1250             SYSCTL_CHILDREN(bbr_sysctl_root),
1251             OID_AUTO,
1252             "pacing",
1253             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1254             "");
1255         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1256             SYSCTL_CHILDREN(bbr_hptsi),
1257             OID_AUTO, "hw_pacing", CTLFLAG_RW,
1258             &bbr_allow_hdwr_pacing, 1,
1259             "Do we allow hardware pacing?");
1260         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1261             SYSCTL_CHILDREN(bbr_hptsi),
1262             OID_AUTO, "hw_pacing_limit", CTLFLAG_RW,
1263             &bbr_hardware_pacing_limit, 4000,
1264             "Do we have a limited number of connections for pacing chelsio (0=no limit)?");
1265         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1266             SYSCTL_CHILDREN(bbr_hptsi),
1267             OID_AUTO, "hw_pacing_adj", CTLFLAG_RW,
1268             &bbr_hdwr_pace_adjust, 2,
1269             "Multiplier to calculated tso size?");
1270         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1271             SYSCTL_CHILDREN(bbr_hptsi),
1272             OID_AUTO, "hw_pacing_floor", CTLFLAG_RW,
1273             &bbr_hdwr_pace_floor, 1,
1274             "Do we invoke the hardware pacing floor?");
1275         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1276             SYSCTL_CHILDREN(bbr_hptsi),
1277             OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW,
1278             &bbr_hdwr_pacing_delay_cnt, 10,
1279             "How many packets must be sent after hdwr pacing is enabled");
1280         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1281             SYSCTL_CHILDREN(bbr_hptsi),
1282             OID_AUTO, "bw_cross", CTLFLAG_RW,
1283             &bbr_cross_over, 3000000,
1284             "What is the point where we cross over to linux like TSO size set");
1285         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1286             SYSCTL_CHILDREN(bbr_hptsi),
1287             OID_AUTO, "seg_deltarg", CTLFLAG_RW,
1288             &bbr_hptsi_segments_delay_tar, 7000,
1289             "What is the worse case delay target for hptsi < 48Mbp connections");
1290         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1291             SYSCTL_CHILDREN(bbr_hptsi),
1292             OID_AUTO, "enet_oh", CTLFLAG_RW,
1293             &bbr_include_enet_oh, 0,
1294             "Do we include the ethernet overhead in calculating pacing delay?");
1295         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1296             SYSCTL_CHILDREN(bbr_hptsi),
1297             OID_AUTO, "ip_oh", CTLFLAG_RW,
1298             &bbr_include_ip_oh, 1,
1299             "Do we include the IP overhead in calculating pacing delay?");
1300         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1301             SYSCTL_CHILDREN(bbr_hptsi),
1302             OID_AUTO, "tcp_oh", CTLFLAG_RW,
1303             &bbr_include_tcp_oh, 0,
1304             "Do we include the TCP overhead in calculating pacing delay?");
1305         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1306             SYSCTL_CHILDREN(bbr_hptsi),
1307             OID_AUTO, "google_discount", CTLFLAG_RW,
1308             &bbr_google_discount, 10,
1309             "What is the default google discount percentage wise for pacing (11 = 1.1%%)?");
1310         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1311             SYSCTL_CHILDREN(bbr_hptsi),
1312             OID_AUTO, "all_get_min", CTLFLAG_RW,
1313             &bbr_all_get_min, 0,
1314             "If you are less than a MSS do you just get the min?");
1315         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1316             SYSCTL_CHILDREN(bbr_hptsi),
1317             OID_AUTO, "tso_min", CTLFLAG_RW,
1318             &bbr_hptsi_bytes_min, 1460,
1319             "For 0 -> 24Mbps what is floor number of segments for TSO");
1320         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1321             SYSCTL_CHILDREN(bbr_hptsi),
1322             OID_AUTO, "seg_tso_max", CTLFLAG_RW,
1323             &bbr_hptsi_segments_max, 6,
1324             "For 0 -> 24Mbps what is top number of segments for TSO");
1325         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1326             SYSCTL_CHILDREN(bbr_hptsi),
1327             OID_AUTO, "seg_floor", CTLFLAG_RW,
1328             &bbr_hptsi_segments_floor, 1,
1329             "Minimum TSO size we will fall too in segments");
1330         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1331             SYSCTL_CHILDREN(bbr_hptsi),
1332             OID_AUTO, "utter_max", CTLFLAG_RW,
1333             &bbr_hptsi_utter_max, 0,
1334             "The absolute maximum that any pacing (outside of hardware) can be");
1335         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1336             SYSCTL_CHILDREN(bbr_hptsi),
1337             OID_AUTO, "seg_divisor", CTLFLAG_RW,
1338             &bbr_hptsi_per_second, 100,
1339             "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps ");
1340         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1341             SYSCTL_CHILDREN(bbr_hptsi),
1342             OID_AUTO, "srtt_mul", CTLFLAG_RW,
1343             &bbr_hptsi_max_mul, 1,
1344             "The multiplier for pace len max");
1345         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1346             SYSCTL_CHILDREN(bbr_hptsi),
1347             OID_AUTO, "srtt_div", CTLFLAG_RW,
1348             &bbr_hptsi_max_div, 2,
1349             "The divisor for pace len max");
1350         /* Measurement controls */
1351         bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1352             SYSCTL_CHILDREN(bbr_sysctl_root),
1353             OID_AUTO,
1354             "measure",
1355             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1356             "Measurement controls");
1357         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1358             SYSCTL_CHILDREN(bbr_measure),
1359             OID_AUTO, "min_i_bw", CTLFLAG_RW,
1360             &bbr_initial_bw_bps, 62500,
1361             "Minimum initial b/w in bytes per second");
1362         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1363             SYSCTL_CHILDREN(bbr_measure),
1364             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1365             &bbr_sack_not_required, 0,
1366             "Do we allow bbr to run on connections not supporting SACK?");
1367         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1368             SYSCTL_CHILDREN(bbr_measure),
1369             OID_AUTO, "use_google", CTLFLAG_RW,
1370             &bbr_use_google_algo, 0,
1371             "Use has close to google V1.0 has possible?");
1372         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1373             SYSCTL_CHILDREN(bbr_measure),
1374             OID_AUTO, "ts_limiting", CTLFLAG_RW,
1375             &bbr_ts_limiting, 1,
1376             "Do we attempt to use the peers timestamp to limit b/w caculations?");
1377         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1378             SYSCTL_CHILDREN(bbr_measure),
1379             OID_AUTO, "ts_can_raise", CTLFLAG_RW,
1380             &bbr_ts_can_raise, 0,
1381             "Can we raise the b/w via timestamp b/w calculation?");
1382         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1383             SYSCTL_CHILDREN(bbr_measure),
1384             OID_AUTO, "ts_delta", CTLFLAG_RW,
1385             &bbr_min_usec_delta, 20000,
1386             "How long in usec between ts of our sends in ts validation code?");
1387         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1388             SYSCTL_CHILDREN(bbr_measure),
1389             OID_AUTO, "ts_peer_delta", CTLFLAG_RW,
1390             &bbr_min_peer_delta, 20,
1391             "What min numerical value should be between the peer deltas?");
1392         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1393             SYSCTL_CHILDREN(bbr_measure),
1394             OID_AUTO, "ts_delta_percent", CTLFLAG_RW,
1395             &bbr_delta_percent, 150,
1396             "What percentage (150 = 15.0) do we allow variance for?");
1397         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1398             SYSCTL_CHILDREN(bbr_measure),
1399             OID_AUTO, "min_measure_good_bw", CTLFLAG_RW,
1400             &bbr_min_measurements_req, 1,
1401             "What is the minimum measurment count we need before we switch to our b/w estimate");
1402         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1403             SYSCTL_CHILDREN(bbr_measure),
1404             OID_AUTO, "min_measure_before_pace", CTLFLAG_RW,
1405             &bbr_no_pacing_until, 4,
1406             "How many pkt-epoch's (0 is off) do we need before pacing is on?");
1407         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1408             SYSCTL_CHILDREN(bbr_measure),
1409             OID_AUTO, "quanta", CTLFLAG_RW,
1410             &bbr_quanta, 2,
1411             "Extra quanta to add when calculating the target (ID section 4.2.3.2).");
1412         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1413             SYSCTL_CHILDREN(bbr_measure),
1414             OID_AUTO, "noretran", CTLFLAG_RW,
1415             &bbr_no_retran, 0,
1416             "Should google mode not use retransmission measurements for the b/w estimation?");
1417         /* State controls */
1418         bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1419             SYSCTL_CHILDREN(bbr_sysctl_root),
1420             OID_AUTO,
1421             "states",
1422             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1423             "State controls");
1424         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1425             SYSCTL_CHILDREN(bbr_states),
1426             OID_AUTO, "idle_restart", CTLFLAG_RW,
1427             &bbr_uses_idle_restart, 0,
1428             "Do we use a new special idle_restart state to ramp back up quickly?");
1429         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1430             SYSCTL_CHILDREN(bbr_states),
1431             OID_AUTO, "idle_restart_threshold", CTLFLAG_RW,
1432             &bbr_idle_restart_threshold, 100000,
1433             "How long must we be idle before we restart??");
1434         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1435             SYSCTL_CHILDREN(bbr_states),
1436             OID_AUTO, "use_pkt_epoch", CTLFLAG_RW,
1437             &bbr_state_is_pkt_epoch, 0,
1438             "Do we use a pkt-epoch for substate if 0 rttProp?");
1439         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1440             SYSCTL_CHILDREN(bbr_states),
1441             OID_AUTO, "startup_rtt_gain", CTLFLAG_RW,
1442             &bbr_rtt_gain_thresh, 0,
1443             "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?");
1444         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1445             SYSCTL_CHILDREN(bbr_states),
1446             OID_AUTO, "drain_floor", CTLFLAG_RW,
1447             &bbr_drain_floor, 88,
1448             "What is the lowest we can drain (pg) too?");
1449         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1450             SYSCTL_CHILDREN(bbr_states),
1451             OID_AUTO, "drain_2_target", CTLFLAG_RW,
1452             &bbr_state_drain_2_tar, 1,
1453             "Do we drain to target in drain substate?");
1454         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1455             SYSCTL_CHILDREN(bbr_states),
1456             OID_AUTO, "gain_2_target", CTLFLAG_RW,
1457             &bbr_gain_to_target, 1,
1458             "Does probe bw gain to target??");
1459         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1460             SYSCTL_CHILDREN(bbr_states),
1461             OID_AUTO, "gain_extra_time", CTLFLAG_RW,
1462             &bbr_gain_gets_extra_too, 1,
1463             "Does probe bw gain get the extra time too?");
1464         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1465             SYSCTL_CHILDREN(bbr_states),
1466             OID_AUTO, "ld_div", CTLFLAG_RW,
1467             &bbr_drain_drop_div, 5,
1468             "Long drain drop divider?");
1469         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1470             SYSCTL_CHILDREN(bbr_states),
1471             OID_AUTO, "ld_mul", CTLFLAG_RW,
1472             &bbr_drain_drop_mul, 4,
1473             "Long drain drop multiplier?");
1474         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1475             SYSCTL_CHILDREN(bbr_states),
1476             OID_AUTO, "rand_ot_disc", CTLFLAG_RW,
1477             &bbr_rand_ot, 50,
1478             "Random discount of the ot?");
1479         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1480             SYSCTL_CHILDREN(bbr_states),
1481             OID_AUTO, "dr_filter_life", CTLFLAG_RW,
1482             &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT,
1483             "How many packet-epochs does the b/w delivery rate last?");
1484         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1485             SYSCTL_CHILDREN(bbr_states),
1486             OID_AUTO, "subdrain_applimited", CTLFLAG_RW,
1487             &bbr_sub_drain_app_limit, 0,
1488             "Does our sub-state drain invoke app limited if its long?");
1489         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1490             SYSCTL_CHILDREN(bbr_states),
1491             OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW,
1492             &bbr_sub_drain_slam_cwnd, 0,
1493             "Should we set/recover cwnd for sub-state drain?");
1494         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1495             SYSCTL_CHILDREN(bbr_states),
1496             OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW,
1497             &bbr_slam_cwnd_in_main_drain, 0,
1498             "Should we set/recover cwnd for main-state drain?");
1499         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1500             SYSCTL_CHILDREN(bbr_states),
1501             OID_AUTO, "google_gets_earlyout", CTLFLAG_RW,
1502             &google_allow_early_out, 1,
1503             "Should we allow google probe-bw/drain to exit early at flight target?");
1504         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1505             SYSCTL_CHILDREN(bbr_states),
1506             OID_AUTO, "google_exit_loss", CTLFLAG_RW,
1507             &google_consider_lost, 1,
1508             "Should we have losses exit gain of probebw in google mode??");
1509         /* Startup controls */
1510         bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1511             SYSCTL_CHILDREN(bbr_sysctl_root),
1512             OID_AUTO,
1513             "startup",
1514             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1515             "Startup controls");
1516         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1517             SYSCTL_CHILDREN(bbr_startup),
1518             OID_AUTO, "cheat_iwnd", CTLFLAG_RW,
1519             &bbr_sends_full_iwnd, 1,
1520             "Do we not pace but burst out initial windows has our TSO size?");
1521         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1522             SYSCTL_CHILDREN(bbr_startup),
1523             OID_AUTO, "loss_threshold", CTLFLAG_RW,
1524             &bbr_startup_loss_thresh, 2000,
1525             "In startup what is the loss threshold in a pe that will exit us from startup?");
1526         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1527             SYSCTL_CHILDREN(bbr_startup),
1528             OID_AUTO, "use_lowerpg", CTLFLAG_RW,
1529             &bbr_use_lower_gain_in_startup, 1,
1530             "Should we use a lower hptsi gain if we see loss in startup?");
1531         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1532             SYSCTL_CHILDREN(bbr_startup),
1533             OID_AUTO, "gain", CTLFLAG_RW,
1534             &bbr_start_exit, 25,
1535             "What gain percent do we need to see to stay in startup??");
1536         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1537             SYSCTL_CHILDREN(bbr_startup),
1538             OID_AUTO, "low_gain", CTLFLAG_RW,
1539             &bbr_low_start_exit, 15,
1540             "What gain percent do we need to see to stay in the lower gain startup??");
1541         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1542             SYSCTL_CHILDREN(bbr_startup),
1543             OID_AUTO, "loss_exit", CTLFLAG_RW,
1544             &bbr_exit_startup_at_loss, 1,
1545             "Should we exit startup at loss in an epoch if we are not gaining?");
1546         /* CWND controls */
1547         bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1548             SYSCTL_CHILDREN(bbr_sysctl_root),
1549             OID_AUTO,
1550             "cwnd",
1551             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1552             "Cwnd controls");
1553         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1554             SYSCTL_CHILDREN(bbr_cwnd),
1555             OID_AUTO, "tar_rtt", CTLFLAG_RW,
1556             &bbr_cwndtarget_rtt_touse, 0,
1557             "Target cwnd rtt measurment to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?");
1558         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1559             SYSCTL_CHILDREN(bbr_cwnd),
1560             OID_AUTO, "may_shrink", CTLFLAG_RW,
1561             &bbr_cwnd_may_shrink, 0,
1562             "Can the cwnd shrink if it would grow to more than the target?");
1563         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1564             SYSCTL_CHILDREN(bbr_cwnd),
1565             OID_AUTO, "max_target_limit", CTLFLAG_RW,
1566             &bbr_target_cwnd_mult_limit, 8,
1567             "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?");
1568         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1569             SYSCTL_CHILDREN(bbr_cwnd),
1570             OID_AUTO, "highspeed_min", CTLFLAG_RW,
1571             &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS,
1572             "What is the high-speed min cwnd (rttProp under 1ms)");
1573         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1574             SYSCTL_CHILDREN(bbr_cwnd),
1575             OID_AUTO, "lowspeed_min", CTLFLAG_RW,
1576             &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS,
1577             "What is the min cwnd (rttProp > 1ms)");
1578         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1579             SYSCTL_CHILDREN(bbr_cwnd),
1580             OID_AUTO, "initwin", CTLFLAG_RW,
1581             &bbr_def_init_win, 10,
1582             "What is the BBR initial window, if 0 use tcp version");
1583         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1584             SYSCTL_CHILDREN(bbr_cwnd),
1585             OID_AUTO, "do_loss_red", CTLFLAG_RW,
1586             &bbr_do_red, 600,
1587             "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?");
1588         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1589             SYSCTL_CHILDREN(bbr_cwnd),
1590             OID_AUTO, "red_scale", CTLFLAG_RW,
1591             &bbr_red_scale, 20000,
1592             "What RTT do we scale with?");
1593         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1594             SYSCTL_CHILDREN(bbr_cwnd),
1595             OID_AUTO, "red_growslow", CTLFLAG_RW,
1596             &bbr_red_growth_restrict, 1,
1597             "Do we restrict cwnd growth for whats in flight?");
1598         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1599             SYSCTL_CHILDREN(bbr_cwnd),
1600             OID_AUTO, "red_div", CTLFLAG_RW,
1601             &bbr_red_div, 2,
1602             "If we reduce whats the divisor?");
1603         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1604             SYSCTL_CHILDREN(bbr_cwnd),
1605             OID_AUTO, "red_mul", CTLFLAG_RW,
1606             &bbr_red_mul, 1,
1607             "If we reduce whats the mulitiplier?");
1608         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1609             SYSCTL_CHILDREN(bbr_cwnd),
1610             OID_AUTO, "target_is_unit", CTLFLAG_RW,
1611             &bbr_target_is_bbunit, 0,
1612             "Is the state target the pacing_gain or BBR_UNIT?");
1613         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1614             SYSCTL_CHILDREN(bbr_cwnd),
1615             OID_AUTO, "drop_limit", CTLFLAG_RW,
1616             &bbr_drop_limit, 0,
1617             "Number of segments limit for drop (0=use min_cwnd w/flight)?");
1618
1619         /* Timeout controls */
1620         bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1621             SYSCTL_CHILDREN(bbr_sysctl_root),
1622             OID_AUTO,
1623             "timeout",
1624             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1625             "Time out controls");
1626         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1627             SYSCTL_CHILDREN(bbr_timeout),
1628             OID_AUTO, "delack", CTLFLAG_RW,
1629             &bbr_delack_time, 100000,
1630             "BBR's delayed ack time");
1631         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1632             SYSCTL_CHILDREN(bbr_timeout),
1633             OID_AUTO, "tlp_uses", CTLFLAG_RW,
1634             &bbr_tlp_type_to_use, 3,
1635             "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt");
1636         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1637             SYSCTL_CHILDREN(bbr_timeout),
1638             OID_AUTO, "persmin", CTLFLAG_RW,
1639             &bbr_persist_min, 250000,
1640             "What is the minimum time in microseconds between persists");
1641         SYSCTL_ADD_U32(&bbr_sysctl_ctx,
1642             SYSCTL_CHILDREN(bbr_timeout),
1643             OID_AUTO, "persmax", CTLFLAG_RW,
1644             &bbr_persist_max, 1000000,
1645             "What is the largest delay in microseconds between persists");
1646         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1647             SYSCTL_CHILDREN(bbr_timeout),
1648             OID_AUTO, "tlp_minto", CTLFLAG_RW,
1649             &bbr_tlp_min, 10000,
1650             "TLP Min timeout in usecs");
1651         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1652             SYSCTL_CHILDREN(bbr_timeout),
1653             OID_AUTO, "tlp_dack_time", CTLFLAG_RW,
1654             &bbr_delayed_ack_time, 200000,
1655             "TLP delayed ack compensation value");
1656         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1657             SYSCTL_CHILDREN(bbr_sysctl_root),
1658             OID_AUTO, "minrto", CTLFLAG_RW,
1659             &bbr_rto_min_ms, 30,
1660             "Minimum RTO in ms");
1661         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1662             SYSCTL_CHILDREN(bbr_timeout),
1663             OID_AUTO, "maxrto", CTLFLAG_RW,
1664             &bbr_rto_max_sec, 4,
1665             "Maxiumum RTO in seconds -- should be at least as large as min_rto");
1666         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1667             SYSCTL_CHILDREN(bbr_timeout),
1668             OID_AUTO, "tlp_retry", CTLFLAG_RW,
1669             &bbr_tlp_max_resend, 2,
1670             "How many times does TLP retry a single segment or multiple with no ACK");
1671         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1672             SYSCTL_CHILDREN(bbr_timeout),
1673             OID_AUTO, "minto", CTLFLAG_RW,
1674             &bbr_min_to, 1000,
1675             "Minimum rack timeout in useconds");
1676         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1677             SYSCTL_CHILDREN(bbr_timeout),
1678             OID_AUTO, "pktdelay", CTLFLAG_RW,
1679             &bbr_pkt_delay, 1000,
1680             "Extra RACK time (in useconds) besides reordering thresh");
1681         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1682             SYSCTL_CHILDREN(bbr_timeout),
1683             OID_AUTO, "incr_tmrs", CTLFLAG_RW,
1684             &bbr_incr_timers, 1,
1685             "Increase the RXT/TLP timer by the pacing time used?");
1686         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1687             SYSCTL_CHILDREN(bbr_timeout),
1688             OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW,
1689             &bbr_marks_rxt_sack_passed, 0,
1690             "Mark sack passed on all those not ack'd when a RXT hits?");
1691         /* Policer controls */
1692         bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
1693             SYSCTL_CHILDREN(bbr_sysctl_root),
1694             OID_AUTO,
1695             "policer",
1696             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1697             "Policer controls");
1698         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1699             SYSCTL_CHILDREN(bbr_policer),
1700             OID_AUTO, "detect_enable", CTLFLAG_RW,
1701             &bbr_policer_detection_enabled, 1,
1702             "Is policer detection enabled??");
1703         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1704             SYSCTL_CHILDREN(bbr_policer),
1705             OID_AUTO, "min_pes", CTLFLAG_RW,
1706             &bbr_lt_intvl_min_rtts, 4,
1707             "Minimum number of PE's?");
1708         SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1709             SYSCTL_CHILDREN(bbr_policer),
1710             OID_AUTO, "bwdiff", CTLFLAG_RW,
1711             &bbr_lt_bw_diff, (4000/8),
1712             "Minimal bw diff?");
1713         SYSCTL_ADD_U64(&bbr_sysctl_ctx,
1714             SYSCTL_CHILDREN(bbr_policer),
1715             OID_AUTO, "bwratio", CTLFLAG_RW,
1716             &bbr_lt_bw_ratio, 8,
1717             "Minimal bw diff?");
1718         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1719             SYSCTL_CHILDREN(bbr_policer),
1720             OID_AUTO, "from_rack_rxt", CTLFLAG_RW,
1721             &bbr_policer_call_from_rack_to, 0,
1722             "Do we call the policer detection code from a rack-timeout?");
1723         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1724             SYSCTL_CHILDREN(bbr_policer),
1725             OID_AUTO, "false_postive", CTLFLAG_RW,
1726             &bbr_lt_intvl_fp, 0,
1727             "What packet epoch do we do false-postive detection at (0=no)?");
1728         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1729             SYSCTL_CHILDREN(bbr_policer),
1730             OID_AUTO, "loss_thresh", CTLFLAG_RW,
1731             &bbr_lt_loss_thresh, 196,
1732             "Loss threshold 196 = 19.6%?");
1733         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1734             SYSCTL_CHILDREN(bbr_policer),
1735             OID_AUTO, "false_postive_thresh", CTLFLAG_RW,
1736             &bbr_lt_fd_thresh, 100,
1737             "What percentage is the false detection threshold (150=15.0)?");
1738         /* All the rest */
1739         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1740             SYSCTL_CHILDREN(bbr_sysctl_root),
1741             OID_AUTO, "cheat_rxt", CTLFLAG_RW,
1742             &bbr_use_rack_resend_cheat, 0,
1743             "Do we burst 1ms between sends on retransmissions (like rack)?");
1744         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1745             SYSCTL_CHILDREN(bbr_sysctl_root),
1746             OID_AUTO, "error_paceout", CTLFLAG_RW,
1747             &bbr_error_base_paceout, 10000,
1748             "When we hit an error what is the min to pace out in usec's?");
1749         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1750             SYSCTL_CHILDREN(bbr_sysctl_root),
1751             OID_AUTO, "kill_paceout", CTLFLAG_RW,
1752             &bbr_max_net_error_cnt, 10,
1753             "When we hit this many errors in a row, kill the session?");
1754         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1755             SYSCTL_CHILDREN(bbr_sysctl_root),
1756             OID_AUTO, "data_after_close", CTLFLAG_RW,
1757             &bbr_ignore_data_after_close, 1,
1758             "Do we hold off sending a RST until all pending data is ack'd");
1759         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1760             SYSCTL_CHILDREN(bbr_sysctl_root),
1761             OID_AUTO, "resend_use_tso", CTLFLAG_RW,
1762             &bbr_resends_use_tso, 0,
1763             "Can resends use TSO?");
1764         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1765             SYSCTL_CHILDREN(bbr_sysctl_root),
1766             OID_AUTO, "sblklimit", CTLFLAG_RW,
1767             &bbr_sack_block_limit, 128,
1768             "When do we start ignoring small sack blocks");
1769         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1770             SYSCTL_CHILDREN(bbr_sysctl_root),
1771             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1772             &bbr_verbose_logging, 0,
1773             "Should BBR black box logging be verbose");
1774         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1775             SYSCTL_CHILDREN(bbr_sysctl_root),
1776             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1777             &bbr_reorder_thresh, 2,
1778             "What factor for rack will be added when seeing reordering (shift right)");
1779         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1780             SYSCTL_CHILDREN(bbr_sysctl_root),
1781             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1782             &bbr_reorder_fade, 0,
1783             "Does reorder detection fade, if so how many ms (0 means never)");
1784         SYSCTL_ADD_S32(&bbr_sysctl_ctx,
1785             SYSCTL_CHILDREN(bbr_sysctl_root),
1786             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1787             &bbr_tlp_thresh, 1,
1788             "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1789         /* Stats and counters */
1790         /* The pacing counters for hdwr/software can't be in the array */
1791         bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1792         bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK);
1793         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1794             SYSCTL_CHILDREN(bbr_sysctl_root),
1795             OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD,
1796             &bbr_hdwr_pacing_enobuf,
1797             "Total number of enobufs for hardware paced flows");
1798         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1799             SYSCTL_CHILDREN(bbr_sysctl_root),
1800             OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD,
1801             &bbr_nohdwr_pacing_enobuf,
1802             "Total number of enobufs for non-hardware paced flows");
1803
1804         bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK);
1805         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1806             SYSCTL_CHILDREN(bbr_sysctl_root),
1807             OID_AUTO, "hdwr_pacing", CTLFLAG_RD,
1808             &bbr_flows_whdwr_pacing,
1809             "Total number of hardware paced flows");
1810         bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK);
1811         SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx,
1812             SYSCTL_CHILDREN(bbr_sysctl_root),
1813             OID_AUTO, "software_pacing", CTLFLAG_RD,
1814             &bbr_flows_nohdwr_pacing,
1815             "Total number of software paced flows");
1816         COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK);
1817         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1818             OID_AUTO, "stats", CTLFLAG_RD,
1819             bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats");
1820         COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK);
1821         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1822             OID_AUTO, "opts", CTLFLAG_RD,
1823             bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats");
1824         COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK);
1825         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1826             OID_AUTO, "lost", CTLFLAG_RD,
1827             bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur");
1828         COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK);
1829         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1830             OID_AUTO, "stateresend", CTLFLAG_RD,
1831             bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend");
1832         COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK);
1833         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1834             OID_AUTO, "statetime", CTLFLAG_RD,
1835             bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states");
1836         COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1837         SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root),
1838             OID_AUTO, "outsize", CTLFLAG_RD,
1839             bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls");
1840         SYSCTL_ADD_PROC(&bbr_sysctl_ctx,
1841             SYSCTL_CHILDREN(bbr_sysctl_root),
1842             OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1843             &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters");
1844 }
1845
1846 static void
1847 bbr_counter_destroy(void)
1848 {
1849         COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE);
1850         COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE);
1851         COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE);
1852         COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT);
1853         COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT);
1854         COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT);
1855         counter_u64_free(bbr_nohdwr_pacing_enobuf);
1856         counter_u64_free(bbr_hdwr_pacing_enobuf);
1857         counter_u64_free(bbr_flows_whdwr_pacing);
1858         counter_u64_free(bbr_flows_nohdwr_pacing);
1859
1860 }
1861
1862 static __inline void
1863 bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts)
1864 {
1865         memset(l, 0, sizeof(union tcp_log_stackspecific));
1866         l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate;
1867         l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate);
1868         l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
1869         l->bw_inuse = bbr_get_bw(bbr);
1870         l->inflight = ctf_flight_size(bbr->rc_tp,
1871                           (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
1872         l->applimited = bbr->r_ctl.r_app_limited_until;
1873         l->delivered = bbr->r_ctl.rc_delivered;
1874         l->timeStamp = cts;
1875         l->lost = bbr->r_ctl.rc_lost;
1876         l->bbr_state = bbr->rc_bbr_state;
1877         l->bbr_substate = bbr_state_val(bbr);
1878         l->epoch = bbr->r_ctl.rc_rtt_epoch;
1879         l->lt_epoch = bbr->r_ctl.rc_lt_epoch;
1880         l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
1881         l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain;
1882         l->inhpts = bbr->rc_inp->inp_in_hpts;
1883         l->ininput = bbr->rc_inp->inp_in_input;
1884         l->use_lt_bw = bbr->rc_lt_use_bw;
1885         l->pkts_out = bbr->r_ctl.rc_flight_at_input;
1886         l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch;
1887 }
1888
1889 static void
1890 bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason)
1891 {
1892         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1893                 union tcp_log_stackspecific log;
1894
1895                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1896                 log.u_bbr.flex1 = 0;
1897                 log.u_bbr.flex2 = 0;
1898                 log.u_bbr.flex5 = 0;
1899                 log.u_bbr.flex3 = 0;
1900                 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate;
1901                 log.u_bbr.flex7 = reason;
1902                 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt;
1903                 log.u_bbr.flex8 = 0;
1904                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1905                     &bbr->rc_inp->inp_socket->so_rcv,
1906                     &bbr->rc_inp->inp_socket->so_snd,
1907                     BBR_LOG_BW_RED_EV, 0,
1908                     0, &log, false, &bbr->rc_tv);
1909         }
1910 }
1911
1912 static void
1913 bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count)
1914 {
1915         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1916                 union tcp_log_stackspecific log;
1917
1918                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1919                 log.u_bbr.flex1 = seq;
1920                 log.u_bbr.flex2 = count;
1921                 log.u_bbr.flex8 = mode;
1922                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1923                     &bbr->rc_inp->inp_socket->so_rcv,
1924                     &bbr->rc_inp->inp_socket->so_snd,
1925                     BBR_LOG_LOWGAIN, 0,
1926                     0, &log, false, &bbr->rc_tv);
1927         }
1928 }
1929
1930 static void
1931 bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling,
1932     uint8_t reason, uint32_t p_maxseg, int len)
1933 {
1934         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1935                 union tcp_log_stackspecific log;
1936
1937                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1938                 log.u_bbr.flex1 = p_maxseg;
1939                 log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags;
1940                 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
1941                 log.u_bbr.flex4 = reason;
1942                 log.u_bbr.flex5 = bbr->rc_in_persist;
1943                 log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val;
1944                 log.u_bbr.flex7 = p_maxseg;
1945                 log.u_bbr.flex8 = bbr->rc_in_persist;
1946                 log.u_bbr.pkts_out = 0;
1947                 log.u_bbr.applimited = len;
1948                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1949                     &bbr->rc_inp->inp_socket->so_rcv,
1950                     &bbr->rc_inp->inp_socket->so_snd,
1951                     BBR_LOG_JUSTRET, 0,
1952                     tlen, &log, false, &bbr->rc_tv);
1953         }
1954 }
1955
1956 static void
1957 bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq)
1958 {
1959         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1960                 union tcp_log_stackspecific log;
1961
1962                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
1963                 log.u_bbr.flex1 = seq;
1964                 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
1965                 log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start;
1966                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
1967                     &bbr->rc_inp->inp_socket->so_rcv,
1968                     &bbr->rc_inp->inp_socket->so_snd,
1969                     BBR_LOG_ENTREC, 0,
1970                     0, &log, false, &bbr->rc_tv);
1971         }
1972 }
1973
1974 static void
1975 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)
1976 {
1977         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
1978                 union tcp_log_stackspecific log;
1979
1980                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
1981                 log.u_bbr.flex1 = tso;
1982                 log.u_bbr.flex2 = maxseg;
1983                 log.u_bbr.flex3 = mtu;
1984                 log.u_bbr.flex4 = csum_flags;
1985                 TCP_LOG_EVENTP(tp, NULL,
1986                     &bbr->rc_inp->inp_socket->so_rcv,
1987                     &bbr->rc_inp->inp_socket->so_snd,
1988                     BBR_LOG_MSGSIZE, 0,
1989                     0, &log, false, &bbr->rc_tv);
1990         }
1991 }
1992
1993 static void
1994 bbr_log_flowend(struct tcp_bbr *bbr)
1995 {
1996         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1997                 union tcp_log_stackspecific log;
1998                 struct sockbuf *r, *s;
1999                 struct timeval tv;
2000
2001                 if (bbr->rc_inp->inp_socket) {
2002                         r = &bbr->rc_inp->inp_socket->so_rcv;
2003                         s = &bbr->rc_inp->inp_socket->so_snd;
2004                 } else {
2005                         r = s = NULL;
2006                 }
2007                 bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv));
2008                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2009                     r, s,
2010                     TCP_LOG_FLOWEND, 0,
2011                     0, &log, false, &tv);
2012         }
2013 }
2014
2015 static void
2016 bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line,
2017     uint32_t lost, uint32_t del)
2018 {
2019         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2020                 union tcp_log_stackspecific log;
2021
2022                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2023                 log.u_bbr.flex1 = lost;
2024                 log.u_bbr.flex2 = del;
2025                 log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw;
2026                 log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt;
2027                 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2028                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2029                 log.u_bbr.flex7 = line;
2030                 log.u_bbr.flex8 = 0;
2031                 log.u_bbr.inflight = bbr->r_ctl.r_measurement_count;
2032                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2033                     &bbr->rc_inp->inp_socket->so_rcv,
2034                     &bbr->rc_inp->inp_socket->so_snd,
2035                     BBR_LOG_PKT_EPOCH, 0,
2036                     0, &log, false, &bbr->rc_tv);
2037         }
2038 }
2039
2040 static void
2041 bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time)
2042 {
2043         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2044                 union tcp_log_stackspecific log;
2045
2046                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2047                 log.u_bbr.flex1 = bbr->r_ctl.rc_lost;
2048                 log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat;
2049                 log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat;
2050                 log.u_bbr.flex7 = line;
2051                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2052                     &bbr->rc_inp->inp_socket->so_rcv,
2053                     &bbr->rc_inp->inp_socket->so_snd,
2054                     BBR_LOG_TIME_EPOCH, 0,
2055                     0, &log, false, &bbr->rc_tv);
2056         }
2057 }
2058
2059 static void
2060 bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth)
2061 {
2062         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2063                 union tcp_log_stackspecific log;
2064
2065                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2066                 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2067                 log.u_bbr.flex2 = new_tar;
2068                 log.u_bbr.flex3 = line;
2069                 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2070                 log.u_bbr.flex5 = bbr_quanta;
2071                 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs;
2072                 log.u_bbr.flex7 = bbr->rc_last_options;
2073                 log.u_bbr.flex8 = meth;
2074                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2075                     &bbr->rc_inp->inp_socket->so_rcv,
2076                     &bbr->rc_inp->inp_socket->so_snd,
2077                     BBR_LOG_STATE_TARGET, 0,
2078                     0, &log, false, &bbr->rc_tv);
2079         }
2080
2081 }
2082
2083 static void
2084 bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2085 {
2086         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2087                 union tcp_log_stackspecific log;
2088
2089                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2090                 log.u_bbr.flex1 = line;
2091                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2092                 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2093                 if (bbr_state_is_pkt_epoch)
2094                         log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
2095                 else
2096                         log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP);
2097                 log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch;
2098                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2099                 log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000);
2100                 log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra;
2101                 log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state;
2102                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2103                     &bbr->rc_inp->inp_socket->so_rcv,
2104                     &bbr->rc_inp->inp_socket->so_snd,
2105                     BBR_LOG_STATE, 0,
2106                     0, &log, false, &bbr->rc_tv);
2107         }
2108 }
2109
2110 static void
2111 bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied,
2112                     uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond)
2113 {
2114         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2115                 union tcp_log_stackspecific log;
2116
2117                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2118                 log.u_bbr.flex1 = line;
2119                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2120                 log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt;
2121                 log.u_bbr.flex4 = applied;
2122                 log.u_bbr.flex5 = rtt;
2123                 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2124                 log.u_bbr.flex7 = cond;
2125                 log.u_bbr.flex8 = reas;
2126                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2127                     &bbr->rc_inp->inp_socket->so_rcv,
2128                     &bbr->rc_inp->inp_socket->so_snd,
2129                     BBR_LOG_RTT_SHRINKS, 0,
2130                     0, &log, false, &bbr->rc_tv);
2131         }
2132 }
2133
2134 static void
2135 bbr_log_type_exit_rec(struct tcp_bbr *bbr)
2136 {
2137         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2138                 union tcp_log_stackspecific log;
2139
2140                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2141                 log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start;
2142                 log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent;
2143                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2144                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2145                     &bbr->rc_inp->inp_socket->so_rcv,
2146                     &bbr->rc_inp->inp_socket->so_snd,
2147                     BBR_LOG_EXITREC, 0,
2148                     0, &log, false, &bbr->rc_tv);
2149         }
2150 }
2151
2152 static void
2153 bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg,
2154     uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line)
2155 {
2156         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2157                 union tcp_log_stackspecific log;
2158
2159                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2160                 log.u_bbr.flex1 = line;
2161                 log.u_bbr.flex2 = prev_acked;
2162                 log.u_bbr.flex3 = bytes_this_ack;
2163                 log.u_bbr.flex4 = chg;
2164                 log.u_bbr.flex5 = th_ack;
2165                 log.u_bbr.flex6 = target;
2166                 log.u_bbr.flex8 = meth;
2167                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2168                     &bbr->rc_inp->inp_socket->so_rcv,
2169                     &bbr->rc_inp->inp_socket->so_snd,
2170                     BBR_LOG_CWND, 0,
2171                     0, &log, false, &bbr->rc_tv);
2172         }
2173 }
2174
2175 static void
2176 bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin)
2177 {
2178         /*
2179          * Log the rtt sample we are applying to the srtt algorithm in
2180          * useconds.
2181          */
2182         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2183                 union tcp_log_stackspecific log;
2184
2185                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2186                 log.u_bbr.flex1 = rtt;
2187                 log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time;
2188                 log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay;
2189                 log.u_bbr.flex4 = bbr->rc_tp->ts_offset;
2190                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2191                 log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv);
2192                 log.u_bbr.flex6 = tsin;
2193                 log.u_bbr.flex7 = 0;
2194                 log.u_bbr.flex8 = bbr->rc_ack_was_delayed;
2195                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2196                     &bbr->rc_inp->inp_socket->so_rcv,
2197                     &bbr->rc_inp->inp_socket->so_snd,
2198                     TCP_LOG_RTT, 0,
2199                     0, &log, false, &bbr->rc_tv);
2200         }
2201 }
2202
2203 static void
2204 bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit)
2205 {
2206         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2207                 union tcp_log_stackspecific log;
2208
2209                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2210                 log.u_bbr.flex1 = time_in;
2211                 log.u_bbr.flex2 = line;
2212                 log.u_bbr.flex8 = enter_exit;
2213                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2214                     &bbr->rc_inp->inp_socket->so_rcv,
2215                     &bbr->rc_inp->inp_socket->so_snd,
2216                     BBR_LOG_PERSIST, 0,
2217                     0, &log, false, &bbr->rc_tv);
2218         }
2219 }
2220 static void
2221 bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts)
2222 {
2223         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2224                 union tcp_log_stackspecific log;
2225
2226                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2227                 log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age;
2228                 log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks;
2229                 log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int;
2230                 log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time;
2231                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2232                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2233                     &bbr->rc_inp->inp_socket->so_rcv,
2234                     &bbr->rc_inp->inp_socket->so_snd,
2235                     BBR_LOG_ACKCLEAR, 0,
2236                     0, &log, false, &bbr->rc_tv);
2237         }
2238 }
2239
2240 static void
2241 bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen,
2242                   uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m)
2243 {
2244         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2245                 union tcp_log_stackspecific log;
2246                 struct timeval tv;
2247
2248                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2249                 log.u_bbr.flex1 = nsegs;
2250                 log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes;
2251                 if (m) {
2252                         struct timespec ts;
2253
2254                         log.u_bbr.flex3 = m->m_flags;
2255                         if (m->m_flags & M_TSTMP) {
2256                                 mbuf_tstmp2timespec(m, &ts);
2257                                 tv.tv_sec = ts.tv_sec;
2258                                 tv.tv_usec = ts.tv_nsec / 1000;
2259                                 log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv);
2260                         } else {
2261                                 log.u_bbr.lt_epoch = 0;
2262                         }
2263                         if (m->m_flags & M_TSTMP_LRO) {
2264                                 tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000;
2265                                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000;
2266                                 log.u_bbr.flex5 = tcp_tv_to_usectick(&tv);
2267                         } else {
2268                                 /* No arrival timestamp */
2269                                 log.u_bbr.flex5 = 0;
2270                         }
2271
2272                         log.u_bbr.pkts_out = tcp_get_usecs(&tv);
2273                 } else {
2274                         log.u_bbr.flex3 = 0;
2275                         log.u_bbr.flex5 = 0;
2276                         log.u_bbr.flex6 = 0;
2277                         log.u_bbr.pkts_out = 0;
2278                 }
2279                 log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state;
2280                 log.u_bbr.flex7 = bbr->r_wanted_output;
2281                 log.u_bbr.flex8 = bbr->rc_in_persist;
2282                 TCP_LOG_EVENTP(bbr->rc_tp, th,
2283                     &bbr->rc_inp->inp_socket->so_rcv,
2284                     &bbr->rc_inp->inp_socket->so_snd,
2285                     TCP_LOG_IN, 0,
2286                     tlen, &log, true, &bbr->rc_tv);
2287         }
2288 }
2289
2290 static void
2291 bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out)
2292 {
2293         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2294                 union tcp_log_stackspecific log;
2295
2296                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2297                 log.u_bbr.flex1 = did_out;
2298                 log.u_bbr.flex2 = nxt_pkt;
2299                 log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val;
2300                 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2301                 log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp;
2302                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes;
2303                 log.u_bbr.flex7 = bbr->r_wanted_output;
2304                 log.u_bbr.flex8 = bbr->rc_in_persist;
2305                 log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay;
2306                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2307                     &bbr->rc_inp->inp_socket->so_rcv,
2308                     &bbr->rc_inp->inp_socket->so_snd,
2309                     BBR_LOG_DOSEG_DONE, 0,
2310                     0, &log, true, &bbr->rc_tv);
2311         }
2312 }
2313
2314 static void
2315 bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts,
2316     int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz)
2317 {
2318         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2319                 union tcp_log_stackspecific log;
2320
2321                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2322                 log.u_bbr.flex1 = line;
2323                 log.u_bbr.flex2 = o_len;
2324                 log.u_bbr.flex3 = segcnt;
2325                 log.u_bbr.flex4 = segsiz;
2326                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2327                     &bbr->rc_inp->inp_socket->so_rcv,
2328                     &bbr->rc_inp->inp_socket->so_snd,
2329                     BBR_LOG_ENOBUF_JMP, ENOBUFS,
2330                     len, &log, true, &bbr->rc_tv);
2331         }
2332 }
2333
2334 static void
2335 bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling)
2336 {
2337         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2338                 union tcp_log_stackspecific log;
2339
2340                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2341                 log.u_bbr.flex1 = timers;
2342                 log.u_bbr.flex2 = ret;
2343                 log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp;
2344                 log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags;
2345                 log.u_bbr.flex5 = cts;
2346                 log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state;
2347                 log.u_bbr.flex8 = hpts_calling;
2348                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2349                     &bbr->rc_inp->inp_socket->so_rcv,
2350                     &bbr->rc_inp->inp_socket->so_snd,
2351                     BBR_LOG_TO_PROCESS, 0,
2352                     0, &log, false, &bbr->rc_tv);
2353         }
2354 }
2355
2356 static void
2357 bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num)
2358 {
2359         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2360                 union tcp_log_stackspecific log;
2361                 uint64_t ar;
2362
2363                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2364                 log.u_bbr.flex1 = bbr->bbr_timer_src;
2365                 log.u_bbr.flex2 = 0;
2366                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2367                 ar = (uint64_t)(bbr->r_ctl.rc_resend);
2368                 ar >>= 32;
2369                 ar &= 0x00000000ffffffff;
2370                 log.u_bbr.flex4 = (uint32_t)ar;
2371                 ar = (uint64_t)bbr->r_ctl.rc_resend;
2372                 ar &= 0x00000000ffffffff;
2373                 log.u_bbr.flex5 = (uint32_t)ar;
2374                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2375                 log.u_bbr.flex8 = to_num;
2376                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2377                     &bbr->rc_inp->inp_socket->so_rcv,
2378                     &bbr->rc_inp->inp_socket->so_snd,
2379                     BBR_LOG_RTO, 0,
2380                     0, &log, false, &bbr->rc_tv);
2381         }
2382 }
2383
2384 static void
2385 bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason)
2386 {
2387         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2388                 union tcp_log_stackspecific log;
2389
2390                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2391                 log.u_bbr.flex1 = flex1;
2392                 log.u_bbr.flex2 = flex2;
2393                 log.u_bbr.flex3 = flex3;
2394                 log.u_bbr.flex4 = 0;
2395                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2396                 log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup;
2397                 log.u_bbr.flex8 = reason;
2398                 log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw;
2399                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2400                     &bbr->rc_inp->inp_socket->so_rcv,
2401                     &bbr->rc_inp->inp_socket->so_snd,
2402                     BBR_LOG_REDUCE, 0,
2403                     0, &log, false, &bbr->rc_tv);
2404         }
2405 }
2406
2407 static void
2408 bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag)
2409 {
2410         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2411                 union tcp_log_stackspecific log;
2412
2413                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2414                 log.u_bbr.flex1 = diag->p_nxt_slot;
2415                 log.u_bbr.flex2 = diag->p_cur_slot;
2416                 log.u_bbr.flex3 = diag->slot_req;
2417                 log.u_bbr.flex4 = diag->inp_hptsslot;
2418                 log.u_bbr.flex5 = diag->slot_remaining;
2419                 log.u_bbr.flex6 = diag->need_new_to;
2420                 log.u_bbr.flex7 = diag->p_hpts_active;
2421                 log.u_bbr.flex8 = diag->p_on_min_sleep;
2422                 /* Hijack other fields as needed  */
2423                 log.u_bbr.epoch = diag->have_slept;
2424                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
2425                 log.u_bbr.pkts_out = diag->co_ret;
2426                 log.u_bbr.applimited = diag->hpts_sleep_time;
2427                 log.u_bbr.delivered = diag->p_prev_slot;
2428                 log.u_bbr.inflight = diag->p_runningtick;
2429                 log.u_bbr.bw_inuse = diag->wheel_tick;
2430                 log.u_bbr.rttProp = diag->wheel_cts;
2431                 log.u_bbr.delRate = diag->maxticks;
2432                 log.u_bbr.cur_del_rate = diag->p_curtick;
2433                 log.u_bbr.cur_del_rate <<= 32;
2434                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
2435                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2436                     &bbr->rc_inp->inp_socket->so_rcv,
2437                     &bbr->rc_inp->inp_socket->so_snd,
2438                     BBR_LOG_HPTSDIAG, 0,
2439                     0, &log, false, &bbr->rc_tv);
2440         }
2441 }
2442
2443 static void
2444 bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt,
2445     uint32_t thresh, uint32_t to)
2446 {
2447         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2448                 union tcp_log_stackspecific log;
2449
2450                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2451                 log.u_bbr.flex1 = bbr->rc_tp->t_rttvar;
2452                 log.u_bbr.flex2 = time_since_sent;
2453                 log.u_bbr.flex3 = srtt;
2454                 log.u_bbr.flex4 = thresh;
2455                 log.u_bbr.flex5 = to;
2456                 log.u_bbr.flex6 = bbr->rc_tp->t_srtt;
2457                 log.u_bbr.flex8 = mode;
2458                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2459                     &bbr->rc_inp->inp_socket->so_rcv,
2460                     &bbr->rc_inp->inp_socket->so_snd,
2461                     BBR_LOG_TIMERPREP, 0,
2462                     0, &log, false, &bbr->rc_tv);
2463         }
2464 }
2465
2466 static void
2467 bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len,
2468     uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod)
2469 {
2470         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2471                 union tcp_log_stackspecific log;
2472
2473                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2474                 log.u_bbr.flex1 = usecs;
2475                 log.u_bbr.flex2 = len;
2476                 log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff);
2477                 log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff);
2478                 if (override)
2479                         log.u_bbr.flex5 = (1 << 2);
2480                 else
2481                         log.u_bbr.flex5 = 0;
2482                 log.u_bbr.flex6 = override;
2483                 log.u_bbr.flex7 = gain;
2484                 log.u_bbr.flex8 = mod;
2485                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2486                     &bbr->rc_inp->inp_socket->so_rcv,
2487                     &bbr->rc_inp->inp_socket->so_snd,
2488                     BBR_LOG_HPTSI_CALC, 0,
2489                     len, &log, false, &bbr->rc_tv);
2490         }
2491 }
2492
2493 static void
2494 bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2495 {
2496         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2497                 union tcp_log_stackspecific log;
2498
2499                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2500
2501                 log.u_bbr.flex1 = bbr->bbr_timer_src;
2502                 log.u_bbr.flex2 = to;
2503                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2504                 log.u_bbr.flex4 = slot;
2505                 log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot;
2506                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2507                 log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2;
2508                 log.u_bbr.flex8 = which;
2509                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2510                     &bbr->rc_inp->inp_socket->so_rcv,
2511                     &bbr->rc_inp->inp_socket->so_snd,
2512                     BBR_LOG_TIMERSTAR, 0,
2513                     0, &log, false, &bbr->rc_tv);
2514         }
2515 }
2516
2517 static void
2518 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)
2519 {
2520         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2521                 union tcp_log_stackspecific log;
2522
2523                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2524                 log.u_bbr.flex1 = thresh;
2525                 log.u_bbr.flex2 = lro;
2526                 log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts;
2527                 log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
2528                 log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2529                 log.u_bbr.flex6 = srtt;
2530                 log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift;
2531                 log.u_bbr.flex8 = frm;
2532                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2533                     &bbr->rc_inp->inp_socket->so_rcv,
2534                     &bbr->rc_inp->inp_socket->so_snd,
2535                     BBR_LOG_THRESH_CALC, 0,
2536                     0, &log, false, &bbr->rc_tv);
2537         }
2538 }
2539
2540 static void
2541 bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed)
2542 {
2543         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2544                 union tcp_log_stackspecific log;
2545
2546                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2547                 log.u_bbr.flex1 = line;
2548                 log.u_bbr.flex2 = bbr->bbr_timer_src;
2549                 log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags;
2550                 log.u_bbr.flex4 = bbr->rc_in_persist;
2551                 log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state;
2552                 log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
2553                 log.u_bbr.flex8 = hpts_removed;
2554                 log.u_bbr.pkts_out = bbr->rc_pacer_started;
2555                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2556                     &bbr->rc_inp->inp_socket->so_rcv,
2557                     &bbr->rc_inp->inp_socket->so_snd,
2558                     BBR_LOG_TIMERCANC, 0,
2559                     0, &log, false, &bbr->rc_tv);
2560         }
2561 }
2562
2563 static void
2564 bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta)
2565 {
2566         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2567                 union tcp_log_stackspecific log;
2568
2569                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2570                 log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio;
2571                 log.u_bbr.flex2 = (peer_delta >> 32);
2572                 log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff);
2573                 log.u_bbr.flex4 = (delta >> 32);
2574                 log.u_bbr.flex5 = (delta & 0x00000000ffffffff);
2575                 log.u_bbr.flex7 = bbr->rc_ts_clock_set;
2576                 log.u_bbr.flex8 = bbr->rc_ts_cant_be_used;
2577                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2578                     &bbr->rc_inp->inp_socket->so_rcv,
2579                     &bbr->rc_inp->inp_socket->so_snd,
2580                     BBR_LOG_TSTMP_VAL, 0,
2581                     0, &log, false, &bbr->rc_tv);
2582         }
2583 }
2584
2585 static void
2586 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)
2587 {
2588         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2589                 union tcp_log_stackspecific log;
2590
2591                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2592                 log.u_bbr.flex1 = tsosz;
2593                 log.u_bbr.flex2 = tls;
2594                 log.u_bbr.flex3 = tcp_min_hptsi_time;
2595                 log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min;
2596                 log.u_bbr.flex5 = old_val;
2597                 log.u_bbr.flex6 = maxseg;
2598                 log.u_bbr.flex7 = bbr->rc_no_pacing;
2599                 log.u_bbr.flex7 <<= 1;
2600                 log.u_bbr.flex7 |= bbr->rc_past_init_win;
2601                 if (hdwr)
2602                         log.u_bbr.flex8 = 0x80 | bbr->rc_use_google;
2603                 else
2604                         log.u_bbr.flex8 = bbr->rc_use_google;
2605                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2606                     &bbr->rc_inp->inp_socket->so_rcv,
2607                     &bbr->rc_inp->inp_socket->so_snd,
2608                     BBR_LOG_BBRTSO, 0,
2609                     0, &log, false, &bbr->rc_tv);
2610         }
2611 }
2612
2613 static void
2614 bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm,
2615                       uint32_t flags, uint32_t line)
2616 {
2617         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2618                 union tcp_log_stackspecific log;
2619
2620                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2621                 log.u_bbr.flex1 = line;
2622                 log.u_bbr.flex2 = rsm->r_start;
2623                 log.u_bbr.flex3 = rsm->r_end;
2624                 log.u_bbr.flex4 = rsm->r_delivered;
2625                 log.u_bbr.flex5 = rsm->r_rtr_cnt;
2626                 log.u_bbr.flex6 = rsm->r_dupack;
2627                 log.u_bbr.flex7 = rsm->r_tim_lastsent[0];
2628                 log.u_bbr.flex8 = rsm->r_flags;
2629                 /* Hijack the pkts_out fids */
2630                 log.u_bbr.applimited = flags;
2631                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2632                     &bbr->rc_inp->inp_socket->so_rcv,
2633                     &bbr->rc_inp->inp_socket->so_snd,
2634                     BBR_RSM_CLEARED, 0,
2635                     0, &log, false, &bbr->rc_tv);
2636         }
2637 }
2638
2639 static void
2640 bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts,
2641     uint32_t flex3, uint32_t flex2, uint32_t flex5,
2642     uint32_t flex6, uint32_t pkts_out, int flex7,
2643     uint32_t flex4, uint32_t flex1)
2644 {
2645
2646         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2647                 union tcp_log_stackspecific log;
2648
2649                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2650                 log.u_bbr.flex1 = flex1;
2651                 log.u_bbr.flex2 = flex2;
2652                 log.u_bbr.flex3 = flex3;
2653                 log.u_bbr.flex4 = flex4;
2654                 log.u_bbr.flex5 = flex5;
2655                 log.u_bbr.flex6 = flex6;
2656                 log.u_bbr.flex7 = flex7;
2657                 /* Hijack the pkts_out fids */
2658                 log.u_bbr.pkts_out = pkts_out;
2659                 log.u_bbr.flex8 = flex8;
2660                 if (bbr->rc_ack_was_delayed)
2661                         log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay;
2662                 else
2663                         log.u_bbr.epoch = 0;
2664                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2665                     &bbr->rc_inp->inp_socket->so_rcv,
2666                     &bbr->rc_inp->inp_socket->so_snd,
2667                     BBR_LOG_BBRUPD, 0,
2668                     flex2, &log, false, &bbr->rc_tv);
2669         }
2670 }
2671
2672 static void
2673 bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason,
2674         uint32_t newbw, uint32_t obw, uint32_t diff,
2675         uint32_t tim)
2676 {
2677         if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2678                 union tcp_log_stackspecific log;
2679
2680                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2681                 log.u_bbr.flex1 = reason;
2682                 log.u_bbr.flex2 = newbw;
2683                 log.u_bbr.flex3 = obw;
2684                 log.u_bbr.flex4 = diff;
2685                 log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost;
2686                 log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del;
2687                 log.u_bbr.flex7 = bbr->rc_lt_is_sampling;
2688                 log.u_bbr.pkts_out = tim;
2689                 log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw;
2690                 if (bbr->rc_lt_use_bw == 0)
2691                         log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
2692                 else
2693                         log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
2694                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2695                     &bbr->rc_inp->inp_socket->so_rcv,
2696                     &bbr->rc_inp->inp_socket->so_snd,
2697                     BBR_LOG_BWSAMP, 0,
2698                     0, &log, false, &bbr->rc_tv);
2699         }
2700 }
2701
2702 static inline void
2703 bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line)
2704 {
2705         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2706                 union tcp_log_stackspecific log;
2707
2708                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2709                 log.u_bbr.flex1 = line;
2710                 log.u_bbr.flex2 = tick;
2711                 log.u_bbr.flex3 = tp->t_maxunacktime;
2712                 log.u_bbr.flex4 = tp->t_acktime;
2713                 log.u_bbr.flex8 = event;
2714                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2715                     &bbr->rc_inp->inp_socket->so_rcv,
2716                     &bbr->rc_inp->inp_socket->so_snd,
2717                     BBR_LOG_PROGRESS, 0,
2718                     0, &log, false, &bbr->rc_tv);
2719         }
2720 }
2721
2722 static void
2723 bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp,
2724                          uint64_t rate, uint64_t hw_rate, int line, uint32_t cts,
2725                          int error)
2726 {
2727         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2728                 union tcp_log_stackspecific log;
2729
2730                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2731                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2732                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2733                 log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2734                 log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2735                 log.u_bbr.bw_inuse = rate;
2736                 log.u_bbr.flex5 = line;
2737                 log.u_bbr.flex6 = error;
2738                 log.u_bbr.flex8 = bbr->skip_gain;
2739                 log.u_bbr.flex8 <<= 1;
2740                 log.u_bbr.flex8 |= bbr->gain_is_limited;
2741                 log.u_bbr.flex8 <<= 1;
2742                 log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing;
2743                 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
2744                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2745                     &bbr->rc_inp->inp_socket->so_rcv,
2746                     &bbr->rc_inp->inp_socket->so_snd,
2747                     BBR_LOG_HDWR_PACE, 0,
2748                     0, &log, false, &bbr->rc_tv);
2749         }
2750 }
2751
2752 static void
2753 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)
2754 {
2755         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2756                 union tcp_log_stackspecific log;
2757
2758                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2759                 log.u_bbr.flex1 = slot;
2760                 log.u_bbr.flex2 = del_by;
2761                 log.u_bbr.flex3 = prev_delay;
2762                 log.u_bbr.flex4 = line;
2763                 log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val;
2764                 log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay;
2765                 log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags);
2766                 log.u_bbr.flex8 = bbr->rc_in_persist;
2767                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2768                     &bbr->rc_inp->inp_socket->so_rcv,
2769                     &bbr->rc_inp->inp_socket->so_snd,
2770                     BBR_LOG_BBRSND, 0,
2771                     len, &log, false, &bbr->rc_tv);
2772         }
2773 }
2774
2775 static void
2776 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)
2777 {
2778         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2779                 union tcp_log_stackspecific log;
2780
2781                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2782                 log.u_bbr.flex1 = bbr->r_ctl.rc_delivered;
2783                 log.u_bbr.flex2 = 0;
2784                 log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt;
2785                 log.u_bbr.flex4 = end;
2786                 log.u_bbr.flex5 = seq;
2787                 log.u_bbr.flex6 = t;
2788                 log.u_bbr.flex7 = match;
2789                 log.u_bbr.flex8 = flags;
2790                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2791                     &bbr->rc_inp->inp_socket->so_rcv,
2792                     &bbr->rc_inp->inp_socket->so_snd,
2793                     BBR_LOG_BBRRTT, 0,
2794                     0, &log, false, &bbr->rc_tv);
2795         }
2796 }
2797
2798 static void
2799 bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method)
2800 {
2801         if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2802                 union tcp_log_stackspecific log;
2803
2804                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
2805                 log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state;
2806                 log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
2807                 log.u_bbr.flex3 = bbr->r_ctl.gain_epoch;
2808                 log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs;
2809                 log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs;
2810                 log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight;
2811                 log.u_bbr.flex7 = 0;
2812                 log.u_bbr.flex8 = entry_method;
2813                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2814                     &bbr->rc_inp->inp_socket->so_rcv,
2815                     &bbr->rc_inp->inp_socket->so_snd,
2816                     BBR_LOG_EXIT_GAIN, 0,
2817                     0, &log, false, &bbr->rc_tv);
2818         }
2819 }
2820
2821 static void
2822 bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired)
2823 {
2824         if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2825                 union tcp_log_stackspecific log;
2826
2827                 bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime);
2828                 /* R-HU */
2829                 log.u_bbr.flex1 = 0;
2830                 log.u_bbr.flex2 = 0;
2831                 log.u_bbr.flex3 = 0;
2832                 log.u_bbr.flex4 = 0;
2833                 log.u_bbr.flex7 = 0;
2834                 log.u_bbr.flex8 = settings_desired;
2835
2836                 TCP_LOG_EVENTP(bbr->rc_tp, NULL,
2837                     &bbr->rc_inp->inp_socket->so_rcv,
2838                     &bbr->rc_inp->inp_socket->so_snd,
2839                     BBR_LOG_SETTINGS_CHG, 0,
2840                     0, &log, false, &bbr->rc_tv);
2841         }
2842 }
2843
2844 /*
2845  * Returns the bw from the our filter.
2846  */
2847 static inline uint64_t
2848 bbr_get_full_bw(struct tcp_bbr *bbr)
2849 {
2850         uint64_t bw;
2851
2852         bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2853
2854         return (bw);
2855 }
2856
2857 static inline void
2858 bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2859 {
2860         uint64_t calclr;
2861         uint32_t lost, del;
2862
2863         if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch)
2864                 lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch;
2865         else
2866                 lost = 0;
2867         del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del;
2868         if (lost == 0)  {
2869                 calclr = 0;
2870         } else if (del) {
2871                 calclr = lost;
2872                 calclr *= (uint64_t)1000;
2873                 calclr /= (uint64_t)del;
2874         } else {
2875                 /* Nothing delivered? 100.0% loss */
2876                 calclr = 1000;
2877         }
2878         bbr->r_ctl.rc_pkt_epoch_loss_rate =  (uint32_t)calclr;
2879         if (IN_RECOVERY(bbr->rc_tp->t_flags))
2880                 bbr->r_ctl.recovery_lr += (uint32_t)calclr;
2881         bbr->r_ctl.rc_pkt_epoch++;
2882         if (bbr->rc_no_pacing &&
2883             (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) {
2884                 bbr->rc_no_pacing = 0;
2885                 tcp_bbr_tso_size_check(bbr, cts);
2886         }
2887         bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time);
2888         bbr->r_ctl.rc_pkt_epoch_time = cts;
2889         /* What was our loss rate */
2890         bbr_log_pkt_epoch(bbr, cts, line, lost, del);
2891         bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered;
2892         bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost;
2893 }
2894
2895 static inline void
2896 bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
2897 {
2898         uint32_t epoch_time;
2899
2900         /* Tick the RTT clock */
2901         bbr->r_ctl.rc_rtt_epoch++;
2902         epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start;
2903         bbr_log_time_epoch(bbr, cts, line, epoch_time);
2904         bbr->r_ctl.rc_rcv_epoch_start = cts;
2905 }
2906
2907 static inline void
2908 bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked)
2909 {
2910         if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) {
2911                 bbr->rc_is_pkt_epoch_now = 1;
2912         }
2913 }
2914
2915 /*
2916  * Returns the bw from either the b/w filter
2917  * or from the lt_bw (if the connection is being
2918  * policed).
2919  */
2920 static inline uint64_t
2921 __bbr_get_bw(struct tcp_bbr *bbr)
2922 {
2923         uint64_t bw, min_bw;
2924         uint64_t rtt;
2925         int gm_measure_cnt = 1;
2926
2927         /*
2928          * For startup we make, like google, a
2929          * minimum b/w. This is generated from the
2930          * IW and the rttProp. We do fall back to srtt
2931          * if for some reason (initial handshake) we don't
2932          * have a rttProp. We, in the worst case, fall back
2933          * to the configured min_bw (rc_initial_hptsi_bw).
2934          */
2935         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
2936                 /* Attempt first to use rttProp */
2937                 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2938                 if (rtt && (rtt < 0xffffffff)) {
2939 measure:
2940                         min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2941                                 ((uint64_t)1000000);
2942                         min_bw /= rtt;
2943                         if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2944                                 min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2945                         }
2946
2947                 } else if (bbr->rc_tp->t_srtt != 0) {
2948                         /* No rttProp, use srtt? */
2949                         rtt = bbr_get_rtt(bbr, BBR_SRTT);
2950                         goto measure;
2951                 } else {
2952                         min_bw = bbr->r_ctl.rc_initial_hptsi_bw;
2953                 }
2954         } else
2955                 min_bw = 0;
2956
2957         if ((bbr->rc_past_init_win == 0) &&
2958             (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp)))
2959                 bbr->rc_past_init_win = 1;
2960         if ((bbr->rc_use_google)  && (bbr->r_ctl.r_measurement_count >= 1))
2961                 gm_measure_cnt = 0;
2962         if (gm_measure_cnt &&
2963             ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) ||
2964              (bbr->rc_past_init_win == 0))) {
2965                 /* For google we use our guess rate until we get 1 measurement */
2966
2967 use_initial_window:
2968                 rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop);
2969                 if (rtt && (rtt < 0xffffffff)) {
2970                         /*
2971                          * We have an RTT measurment. Use that in
2972                          * combination with our initial window to calculate
2973                          * a b/w.
2974                          */
2975                         bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) *
2976                                 ((uint64_t)1000000);
2977                         bw /= rtt;
2978                         if (bw < bbr->r_ctl.rc_initial_hptsi_bw) {
2979                                 bw = bbr->r_ctl.rc_initial_hptsi_bw;
2980                         }
2981                 } else {
2982                         /* Drop back to the 40 and punt to a default */
2983                         bw = bbr->r_ctl.rc_initial_hptsi_bw;
2984                 }
2985                 if (bw < 1)
2986                         /* Probably should panic */
2987                         bw = 1;
2988                 if (bw > min_bw)
2989                         return (bw);
2990                 else
2991                         return (min_bw);
2992         }
2993         if (bbr->rc_lt_use_bw)
2994                 bw = bbr->r_ctl.rc_lt_bw;
2995         else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0))
2996                 bw = bbr->r_ctl.red_bw;
2997         else
2998                 bw = get_filter_value(&bbr->r_ctl.rc_delrate);
2999         if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) {
3000                 /*
3001                  * Enforce user set rate limit, keep in mind that
3002                  * t_peakrate_thr is in B/s already
3003                  */
3004                 bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw);
3005         }
3006         if (bw == 0) {
3007                 /* We should not be at 0, go to the initial window then  */
3008                 goto use_initial_window;
3009         }
3010         if (bw < 1)
3011                 /* Probably should panic */
3012                 bw = 1;
3013         if (bw < min_bw)
3014                 bw = min_bw;
3015         return (bw);
3016 }
3017
3018 static inline uint64_t
3019 bbr_get_bw(struct tcp_bbr *bbr)
3020 {
3021         uint64_t bw;
3022
3023         bw = __bbr_get_bw(bbr);
3024         return (bw);
3025 }
3026
3027 static inline void
3028 bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts)
3029 {
3030         bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch;
3031         bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time;
3032         bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3033         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3034 }
3035
3036 static inline void
3037 bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts)
3038 {
3039         bbr->rc_lt_is_sampling = 0;
3040         bbr->rc_lt_use_bw = 0;
3041         bbr->r_ctl.rc_lt_bw = 0;
3042         bbr_reset_lt_bw_interval(bbr, cts);
3043 }
3044
3045 static inline void
3046 bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin)
3047 {
3048         uint64_t diff;
3049
3050         /* Do we have a previous sample? */
3051         if (bbr->r_ctl.rc_lt_bw) {
3052                 /* Get the diff in bytes per second */
3053                 if (bbr->r_ctl.rc_lt_bw > bw)
3054                         diff = bbr->r_ctl.rc_lt_bw - bw;
3055                 else
3056                         diff = bw - bbr->r_ctl.rc_lt_bw;
3057                 if ((diff <= bbr_lt_bw_diff) ||
3058                     (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) {
3059                         /* Consider us policed */
3060                         uint32_t saved_bw;
3061
3062                         saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw;
3063                         bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2;   /* average of two */
3064                         bbr->rc_lt_use_bw = 1;
3065                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
3066                         /*
3067                          * Use pkt based epoch for measuring length of
3068                          * policer up
3069                          */
3070                         bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch;
3071                         /*
3072                          * reason 4 is we need to start consider being
3073                          * policed
3074                          */
3075                         bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin);
3076                         return;
3077                 }
3078         }
3079         bbr->r_ctl.rc_lt_bw = bw;
3080         bbr_reset_lt_bw_interval(bbr, cts);
3081         bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin);
3082 }
3083
3084 static void
3085 bbr_randomize_extra_state_time(struct tcp_bbr *bbr)
3086 {
3087         uint32_t ran, deduct;
3088
3089         ran = arc4random_uniform(bbr_rand_ot);
3090         if (ran) {
3091                 deduct = bbr->r_ctl.rc_level_state_extra / ran;
3092                 bbr->r_ctl.rc_level_state_extra -= deduct;
3093         }
3094 }
3095 /*
3096  * Return randomly the starting state
3097  * to use in probebw.
3098  */
3099 static uint8_t
3100 bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts)
3101 {
3102         uint32_t ran;
3103         uint8_t ret_val;
3104
3105         /* Initialize the offset to 0 */
3106         bbr->r_ctl.rc_exta_time_gd = 0;
3107         bbr->rc_hit_state_1 = 0;
3108         bbr->r_ctl.rc_level_state_extra = 0;
3109         ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1));
3110         /*
3111          * The math works funny here :) the return value is used to set the
3112          * substate and then the state change is called which increments by
3113          * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when
3114          * we fully enter the state. Note that the (8 - 1 - ran) assures that
3115          * we return 1 - 7, so we dont return 0 and end up starting in
3116          * state 1 (DRAIN).
3117          */
3118         ret_val = BBR_SUBSTATE_COUNT - 1 - ran;
3119         /* Set an epoch */
3120         if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP))
3121                 bbr_set_epoch(bbr, cts, __LINE__);
3122
3123         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
3124         return (ret_val);
3125 }
3126
3127 static void
3128 bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected)
3129 {
3130         uint32_t diff, d_time;
3131         uint64_t del_time, bw, lost, delivered;
3132
3133         if (bbr->r_use_policer == 0)
3134                 return;
3135         if (bbr->rc_lt_use_bw) {
3136                 /* We are using lt bw do we stop yet? */
3137                 diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use;
3138                 if (diff > bbr_lt_bw_max_rtts) {
3139                         /* Reset it all */
3140 reset_all:
3141                         bbr_reset_lt_bw_sampling(bbr, cts);
3142                         if (bbr->rc_filled_pipe) {
3143                                 bbr_set_epoch(bbr, cts, __LINE__);
3144                                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
3145                                 bbr_substate_change(bbr, cts, __LINE__, 0);
3146                                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
3147                                 bbr_log_type_statechange(bbr, cts, __LINE__);
3148                         } else {
3149                                 /*
3150                                  * This should not happen really
3151                                  * unless we remove the startup/drain
3152                                  * restrictions above.
3153                                  */
3154                                 bbr->rc_bbr_state = BBR_STATE_STARTUP;
3155                                 bbr_set_epoch(bbr, cts, __LINE__);
3156                                 bbr->r_ctl.rc_bbr_state_time = cts;
3157                                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
3158                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
3159                                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
3160                                 bbr_set_state_target(bbr, __LINE__);
3161                                 bbr_log_type_statechange(bbr, cts, __LINE__);
3162                         }
3163                         /* reason 0 is to stop using lt-bw */
3164                         bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0);
3165                         return;
3166                 }
3167                 if (bbr_lt_intvl_fp == 0) {
3168                         /* Not doing false-postive detection */
3169                         return;
3170                 }
3171                 /* False positive detection */
3172                 if (diff == bbr_lt_intvl_fp) {
3173                         /* At bbr_lt_intvl_fp we record the lost */
3174                         bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered;
3175                         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
3176                 } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) {
3177                         /* Now is our loss rate still high? */
3178                         lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3179                         delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3180                         if ((delivered == 0) ||
3181                             (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) {
3182                                 /* No still below our threshold */
3183                                 bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0);
3184                         } else {
3185                                 /* Yikes its still high, it must be a false positive */
3186                                 bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0);
3187                                 goto reset_all;
3188                         }
3189                 }
3190                 return;
3191         }
3192         /*
3193          * Wait for the first loss before sampling, to let the policer
3194          * exhaust its tokens and estimate the steady-state rate allowed by
3195          * the policer. Starting samples earlier includes bursts that
3196          * over-estimate the bw.
3197          */
3198         if (bbr->rc_lt_is_sampling == 0) {
3199                 /* reason 1 is to begin doing the sampling  */
3200                 if (loss_detected == 0)
3201                         return;
3202                 bbr_reset_lt_bw_interval(bbr, cts);
3203                 bbr->rc_lt_is_sampling = 1;
3204                 bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0);
3205                 return;
3206         }
3207         /* Now how long were we delivering long term last> */
3208         if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time))
3209                 d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time;
3210         else
3211                 d_time = 0;
3212
3213         /* To avoid underestimates, reset sampling if we run out of data. */
3214         if (bbr->r_ctl.r_app_limited_until) {
3215                 /* Can not measure in app-limited state */
3216                 bbr_reset_lt_bw_sampling(bbr, cts);
3217                 /* reason 2 is to reset sampling due to app limits  */
3218                 bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time);
3219                 return;
3220         }
3221         diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch;
3222         if (diff < bbr_lt_intvl_min_rtts) {
3223                 /*
3224                  * need more samples (we don't
3225                  * start on a round like linux so
3226                  * we need 1 more).
3227                  */
3228                 /* 6 is not_enough time or no-loss */
3229                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3230                 return;
3231         }
3232         if (diff > (4 * bbr_lt_intvl_min_rtts)) {
3233                 /*
3234                  * For now if we wait too long, reset all sampling. We need
3235                  * to do some research here, its possible that we should
3236                  * base this on how much loss as occurred.. something like
3237                  * if its under 10% (or some thresh) reset all otherwise
3238                  * don't.  Thats for phase II I guess.
3239                  */
3240                 bbr_reset_lt_bw_sampling(bbr, cts);
3241                 /* reason 3 is to reset sampling due too long of sampling */
3242                 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3243                 return;
3244         }
3245         /*
3246          * End sampling interval when a packet is lost, so we estimate the
3247          * policer tokens were exhausted. Stopping the sampling before the
3248          * tokens are exhausted under-estimates the policed rate.
3249          */
3250         if (loss_detected == 0) {
3251                 /* 6 is not_enough time or no-loss */
3252                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3253                 return;
3254         }
3255         /* Calculate packets lost and delivered in sampling interval. */
3256         lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost;
3257         delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del;
3258         if ((delivered == 0) ||
3259             (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) {
3260                 bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time);
3261                 return;
3262         }
3263         if (d_time < 1000) {
3264                 /* Not enough time. wait */
3265                 /* 6 is not_enough time or no-loss */
3266                 bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time);
3267                 return;
3268         }
3269         if (d_time >= (0xffffffff / USECS_IN_MSEC)) {
3270                 /* Too long */
3271                 bbr_reset_lt_bw_sampling(bbr, cts);
3272                 /* reason 3 is to reset sampling due too long of sampling */
3273                 bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time);
3274                 return;
3275         }
3276         del_time = d_time;
3277         bw = delivered;
3278         bw *= (uint64_t)USECS_IN_SECOND;
3279         bw /= del_time;
3280         bbr_lt_bw_samp_done(bbr, bw, cts, d_time);
3281 }
3282
3283 /*
3284  * Allocate a sendmap from our zone.
3285  */
3286 static struct bbr_sendmap *
3287 bbr_alloc(struct tcp_bbr *bbr)
3288 {
3289         struct bbr_sendmap *rsm;
3290
3291         BBR_STAT_INC(bbr_to_alloc);
3292         rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO));
3293         if (rsm) {
3294                 bbr->r_ctl.rc_num_maps_alloced++;
3295                 return (rsm);
3296         }
3297         if (bbr->r_ctl.rc_free_cnt) {
3298                 BBR_STAT_INC(bbr_to_alloc_emerg);
3299                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
3300                 TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
3301                 bbr->r_ctl.rc_free_cnt--;
3302                 return (rsm);
3303         }
3304         BBR_STAT_INC(bbr_to_alloc_failed);
3305         return (NULL);
3306 }
3307
3308 static struct bbr_sendmap *
3309 bbr_alloc_full_limit(struct tcp_bbr *bbr)
3310 {
3311         if ((V_tcp_map_entries_limit > 0) &&
3312             (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3313                 BBR_STAT_INC(bbr_alloc_limited);
3314                 if (!bbr->alloc_limit_reported) {
3315                         bbr->alloc_limit_reported = 1;
3316                         BBR_STAT_INC(bbr_alloc_limited_conns);
3317                 }
3318                 return (NULL);
3319         }
3320         return (bbr_alloc(bbr));
3321 }
3322
3323 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3324 static struct bbr_sendmap *
3325 bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type)
3326 {
3327         struct bbr_sendmap *rsm;
3328
3329         if (limit_type) {
3330                 /* currently there is only one limit type */
3331                 if (V_tcp_map_split_limit > 0 &&
3332                     bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
3333                         BBR_STAT_INC(bbr_split_limited);
3334                         if (!bbr->alloc_limit_reported) {
3335                                 bbr->alloc_limit_reported = 1;
3336                                 BBR_STAT_INC(bbr_alloc_limited_conns);
3337                         }
3338                         return (NULL);
3339                 }
3340         }
3341
3342         /* allocate and mark in the limit type, if set */
3343         rsm = bbr_alloc(bbr);
3344         if (rsm != NULL && limit_type) {
3345                 rsm->r_limit_type = limit_type;
3346                 bbr->r_ctl.rc_num_split_allocs++;
3347         }
3348         return (rsm);
3349 }
3350
3351 static void
3352 bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
3353 {
3354         if (rsm->r_limit_type) {
3355                 /* currently there is only one limit type */
3356                 bbr->r_ctl.rc_num_split_allocs--;
3357         }
3358         if (rsm->r_is_smallmap)
3359                 bbr->r_ctl.rc_num_small_maps_alloced--;
3360         if (bbr->r_ctl.rc_tlp_send == rsm)
3361                 bbr->r_ctl.rc_tlp_send = NULL;
3362         if (bbr->r_ctl.rc_resend == rsm) {
3363                 bbr->r_ctl.rc_resend = NULL;
3364         }
3365         if (bbr->r_ctl.rc_next == rsm)
3366                 bbr->r_ctl.rc_next = NULL;
3367         if (bbr->r_ctl.rc_sacklast == rsm)
3368                 bbr->r_ctl.rc_sacklast = NULL;
3369         if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
3370                 memset(rsm, 0, sizeof(struct bbr_sendmap));
3371                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
3372                 rsm->r_limit_type = 0;
3373                 bbr->r_ctl.rc_free_cnt++;
3374                 return;
3375         }
3376         bbr->r_ctl.rc_num_maps_alloced--;
3377         uma_zfree(bbr_zone, rsm);
3378 }
3379
3380 /*
3381  * Returns the BDP.
3382  */
3383 static uint64_t
3384 bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) {
3385         /*
3386          * Calculate the bytes in flight needed given the bw (in bytes per
3387          * second) and the specifyed rtt in useconds. We need to put out the
3388          * returned value per RTT to match that rate. Gain will normaly
3389          * raise it up from there.
3390          *
3391          * This should not overflow as long as the bandwidth is below 1
3392          * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller
3393          * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30).
3394          */
3395         uint64_t usec_per_sec;
3396
3397         usec_per_sec = USECS_IN_SECOND;
3398         return ((rtt * bw) / usec_per_sec);
3399 }
3400
3401 /*
3402  * Return the initial cwnd.
3403  */
3404 static uint32_t
3405 bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp)
3406 {
3407         uint32_t i_cwnd;
3408
3409         if (bbr->rc_init_win) {
3410                 i_cwnd = bbr->rc_init_win * tp->t_maxseg;
3411         } else if (V_tcp_initcwnd_segments)
3412                 i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg),
3413                     max(2 * tp->t_maxseg, 14600));
3414         else if (V_tcp_do_rfc3390)
3415                 i_cwnd = min(4 * tp->t_maxseg,
3416                     max(2 * tp->t_maxseg, 4380));
3417         else {
3418                 /* Per RFC5681 Section 3.1 */
3419                 if (tp->t_maxseg > 2190)
3420                         i_cwnd = 2 * tp->t_maxseg;
3421                 else if (tp->t_maxseg > 1095)
3422                         i_cwnd = 3 * tp->t_maxseg;
3423                 else
3424                         i_cwnd = 4 * tp->t_maxseg;
3425         }
3426         return (i_cwnd);
3427 }
3428
3429 /*
3430  * Given a specified gain, return the target
3431  * cwnd based on that gain.
3432  */
3433 static uint32_t
3434 bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw)
3435 {
3436         uint64_t bdp, rtt;
3437         uint32_t cwnd;
3438
3439         if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) ||
3440             (bbr_get_full_bw(bbr) == 0)) {
3441                 /* No measurements yet */
3442                 return (bbr_initial_cwnd(bbr, bbr->rc_tp));
3443         }
3444         /*
3445          * Get bytes per RTT needed (rttProp is normally in
3446          * bbr_cwndtarget_rtt_touse)
3447          */
3448         rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse);
3449         /* Get the bdp from the two values */
3450         bdp = bbr_get_bw_delay_prod(rtt, bw);
3451         /* Now apply the gain */
3452         cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT));
3453
3454         return (cwnd);
3455 }
3456
3457 static uint32_t
3458 bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain)
3459 {
3460         uint32_t cwnd, mss;
3461
3462         mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
3463         /* Get the base cwnd with gain rounded to a mss */
3464         cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss);
3465         /*
3466          * Add in N (2 default since we do not have a
3467          * fq layer to trap packets in) quanta's per the I-D
3468          * section 4.2.3.2 quanta adjust.
3469          */
3470         cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs);
3471         if (bbr->rc_use_google) {
3472                 if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3473                    (bbr_state_val(bbr) == BBR_SUB_GAIN)) {
3474                         /*
3475                          * The linux implementation adds
3476                          * an extra 2 x mss in gain cycle which
3477                          * is documented no-where except in the code.
3478                          * so we add more for Neal undocumented feature
3479                          */
3480                         cwnd += 2 * mss;
3481                 }
3482                 if ((cwnd / mss) & 0x1) {
3483                         /* Round up for odd num mss */
3484                         cwnd += mss;
3485                 }
3486         }
3487         /* Are we below the min cwnd? */
3488         if (cwnd < get_min_cwnd(bbr))
3489                 return (get_min_cwnd(bbr));
3490         return (cwnd);
3491 }
3492
3493 static uint16_t
3494 bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain)
3495 {
3496         if (gain < 1)
3497                 gain = 1;
3498         return (gain);
3499 }
3500
3501 static uint32_t
3502 bbr_get_header_oh(struct tcp_bbr *bbr)
3503 {
3504         int seg_oh;
3505
3506         seg_oh = 0;
3507         if (bbr->r_ctl.rc_inc_tcp_oh) {
3508                 /* Do we include TCP overhead? */
3509                 seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr));
3510         }
3511         if (bbr->r_ctl.rc_inc_ip_oh) {
3512                 /* Do we include IP overhead? */
3513 #ifdef INET6
3514                 if (bbr->r_is_v6)
3515                         seg_oh += sizeof(struct ip6_hdr);
3516                 else
3517 #endif
3518 #ifdef INET
3519                         seg_oh += sizeof(struct ip);
3520 #endif
3521         }
3522         if (bbr->r_ctl.rc_inc_enet_oh) {
3523                 /* Do we include the ethernet overhead?  */
3524                 seg_oh += sizeof(struct ether_header);
3525         }
3526         return(seg_oh);
3527 }
3528
3529 static uint32_t
3530 bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw)
3531 {
3532         uint64_t divor, res, tim;
3533
3534         if (useconds_time == 0)
3535                 return (0);
3536         gain = bbr_gain_adjust(bbr, gain);
3537         divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT;
3538         tim = useconds_time;
3539         res = (tim * bw * gain) / divor;
3540         if (res == 0)
3541                 res = 1;
3542         return ((uint32_t)res);
3543 }
3544
3545 /*
3546  * Given a gain and a length return the delay in useconds that
3547  * should be used to evenly space out packets
3548  * on the connection (based on the gain factor).
3549  */
3550 static uint32_t
3551 bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog)
3552 {
3553         uint64_t bw, lentim, res;
3554         uint32_t usecs, srtt, over = 0;
3555         uint32_t seg_oh, num_segs, maxseg;
3556
3557         if (len == 0)
3558                 return (0);
3559
3560         maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
3561         num_segs = (len + maxseg - 1) / maxseg;
3562         if (bbr->rc_use_google == 0) {
3563                 seg_oh = bbr_get_header_oh(bbr);
3564                 len += (num_segs * seg_oh);
3565         }
3566         gain = bbr_gain_adjust(bbr, gain);
3567         bw = bbr_get_bw(bbr);
3568         if (bbr->rc_use_google) {
3569                 uint64_t cbw;
3570
3571                 /*
3572                  * Reduce the b/w by the google discount
3573                  * factor 10 = 1%.
3574                  */
3575                 cbw = bw *  (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount);
3576                 cbw /= (uint64_t)1000;
3577                 /* We don't apply a discount if it results in 0 */
3578                 if (cbw > 0)
3579                         bw = cbw;
3580         }
3581         lentim = ((uint64_t)len *
3582                   (uint64_t)USECS_IN_SECOND *
3583                   (uint64_t)BBR_UNIT);
3584         res = lentim / ((uint64_t)gain * bw);
3585         if (res == 0)
3586                 res = 1;
3587         usecs = (uint32_t)res;
3588         srtt = bbr_get_rtt(bbr, BBR_SRTT);
3589         if (bbr_hptsi_max_mul && bbr_hptsi_max_div &&
3590             (bbr->rc_use_google == 0) &&
3591             (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) {
3592                 /*
3593                  * We cannot let the delay be more than 1/2 the srtt time.
3594                  * Otherwise we cannot pace out or send properly.
3595                  */
3596                 over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div;
3597                 BBR_STAT_INC(bbr_hpts_min_time);
3598         }
3599         if (!nolog)
3600                 bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1);
3601         return (usecs);
3602 }
3603
3604 static void
3605 bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack,
3606                  uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses)
3607 {
3608         INP_WLOCK_ASSERT(tp->t_inpcb);
3609         uint64_t bw;
3610         uint32_t cwnd, target_cwnd, saved_bytes, maxseg;
3611         int32_t meth;
3612
3613 #ifdef STATS
3614         if ((tp->t_flags & TF_GPUTINPROG) &&
3615             SEQ_GEQ(th->th_ack, tp->gput_ack)) {
3616                 /*
3617                  * Strech acks and compressed acks will cause this to
3618                  * oscillate but we are doing it the same way as the main
3619                  * stack so it will be compariable (though possibly not
3620                  * ideal).
3621                  */
3622                 int32_t cgput;
3623                 int64_t gput, time_stamp;
3624
3625                 gput = (int64_t) (th->th_ack - tp->gput_seq) * 8;
3626                 time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000));
3627                 cgput = gput / time_stamp;
3628                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
3629                                          cgput);
3630                 if (tp->t_stats_gput_prev > 0)
3631                         stats_voi_update_abs_s32(tp->t_stats,
3632                                                  VOI_TCP_GPUT_ND,
3633                                                  ((gput - tp->t_stats_gput_prev) * 100) /
3634                                                  tp->t_stats_gput_prev);
3635                 tp->t_flags &= ~TF_GPUTINPROG;
3636                 tp->t_stats_gput_prev = cgput;
3637         }
3638 #endif
3639         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3640             ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
3641                 /* We don't change anything in probe-rtt */
3642                 return;
3643         }
3644         maxseg = tp->t_maxseg - bbr->rc_last_options;
3645         saved_bytes = bytes_this_ack;
3646         bytes_this_ack += sack_changed;
3647         if (bytes_this_ack > prev_acked) {
3648                 bytes_this_ack -= prev_acked;
3649                 /*
3650                  * A byte ack'd gives us a full mss
3651                  * to be like linux i.e. they count packets.
3652                  */
3653                 if ((bytes_this_ack < maxseg) && bbr->rc_use_google)
3654                         bytes_this_ack = maxseg;
3655         } else {
3656                 /* Unlikely */
3657                 bytes_this_ack = 0;
3658         }
3659         cwnd = tp->snd_cwnd;
3660         bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3661         if (bw)
3662                 target_cwnd = bbr_get_target_cwnd(bbr,
3663                                                   bw,
3664                                                   (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain);
3665         else
3666                 target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp);
3667         if (IN_RECOVERY(tp->t_flags) &&
3668             (bbr->bbr_prev_in_rec == 0)) {
3669                 /*
3670                  * We are entering recovery and
3671                  * thus packet conservation.
3672                  */
3673                 bbr->pkt_conservation = 1;
3674                 bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime;
3675                 cwnd = ctf_flight_size(tp,
3676                                        (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3677                         bytes_this_ack;
3678         }
3679         if (IN_RECOVERY(tp->t_flags)) {
3680                 uint32_t flight;
3681
3682                 bbr->bbr_prev_in_rec = 1;
3683                 if (cwnd > losses) {
3684                         cwnd -= losses;
3685                         if (cwnd < maxseg)
3686                                 cwnd = maxseg;
3687                 } else
3688                         cwnd = maxseg;
3689                 flight = ctf_flight_size(tp,
3690                                          (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3691                 bbr_log_type_cwndupd(bbr, flight, 0,
3692                                      losses, 10, 0, 0, line);
3693                 if (bbr->pkt_conservation) {
3694                         uint32_t time_in;
3695
3696                         if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start))
3697                                 time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start;
3698                         else
3699                                 time_in = 0;
3700
3701                         if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
3702                                 /* Clear packet conservation after an rttProp */
3703                                 bbr->pkt_conservation = 0;
3704                         } else {
3705                                 if ((flight + bytes_this_ack) > cwnd)
3706                                         cwnd = flight + bytes_this_ack;
3707                                 if (cwnd < get_min_cwnd(bbr))
3708                                         cwnd = get_min_cwnd(bbr);
3709                                 tp->snd_cwnd = cwnd;
3710                                 bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed,
3711                                                      prev_acked, 1, target_cwnd, th->th_ack, line);
3712                                 return;
3713                         }
3714                 }
3715         } else
3716                 bbr->bbr_prev_in_rec = 0;
3717         if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) {
3718                 bbr->r_ctl.restrict_growth--;
3719                 if (bytes_this_ack > maxseg)
3720                         bytes_this_ack = maxseg;
3721         }
3722         if (bbr->rc_filled_pipe) {
3723                 /*
3724                  * Here we have exited startup and filled the pipe. We will
3725                  * thus allow the cwnd to shrink to the target. We hit here
3726                  * mostly.
3727                  */
3728                 uint32_t s_cwnd;
3729
3730                 meth = 2;
3731                 s_cwnd = min((cwnd + bytes_this_ack), target_cwnd);
3732                 if (s_cwnd > cwnd)
3733                         cwnd = s_cwnd;
3734                 else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing)
3735                         cwnd = s_cwnd;
3736         } else {
3737                 /*
3738                  * Here we are still in startup, we increase cwnd by what
3739                  * has been acked.
3740                  */
3741                 if ((cwnd < target_cwnd) ||
3742                     (bbr->rc_past_init_win == 0)) {
3743                         meth = 3;
3744                         cwnd += bytes_this_ack;
3745                 } else {
3746                         /*
3747                          * Method 4 means we are at target so no gain in
3748                          * startup and past the initial window.
3749                          */
3750                         meth = 4;
3751                 }
3752         }
3753         tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr));
3754         bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line);
3755 }
3756
3757 static void
3758 tcp_bbr_partialack(struct tcpcb *tp)
3759 {
3760         struct tcp_bbr *bbr;
3761
3762         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3763         INP_WLOCK_ASSERT(tp->t_inpcb);
3764         if (ctf_flight_size(tp,
3765                 (bbr->r_ctl.rc_sacked  + bbr->r_ctl.rc_lost_bytes)) <=
3766             tp->snd_cwnd) {
3767                 bbr->r_wanted_output = 1;
3768         }
3769 }
3770
3771 static void
3772 bbr_post_recovery(struct tcpcb *tp)
3773 {
3774         struct tcp_bbr *bbr;
3775         uint32_t  flight;
3776
3777         INP_WLOCK_ASSERT(tp->t_inpcb);
3778         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3779         /*
3780          * Here we just exit recovery.
3781          */
3782         EXIT_RECOVERY(tp->t_flags);
3783         /* Lock in our b/w reduction for the specified number of pkt-epochs */
3784         bbr->r_recovery_bw = 0;
3785         tp->snd_recover = tp->snd_una;
3786         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3787         bbr->pkt_conservation = 0;
3788         if (bbr->rc_use_google == 0) {
3789                 /*
3790                  * For non-google mode lets
3791                  * go ahead and make sure we clear
3792                  * the recovery state so if we
3793                  * bounce back in to recovery we
3794                  * will do PC.
3795                  */
3796                 bbr->bbr_prev_in_rec = 0;
3797         }
3798         bbr_log_type_exit_rec(bbr);
3799         if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3800                 tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
3801                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__);
3802         } else {
3803                 /* For probe-rtt case lets fix up its saved_cwnd */
3804                 if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) {
3805                         bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent;
3806                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__);
3807                 }
3808         }
3809         flight = ctf_flight_size(tp,
3810                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
3811         if ((bbr->rc_use_google == 0) &&
3812             bbr_do_red) {
3813                 uint64_t val, lr2use;
3814                 uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd;
3815                 uint32_t *cwnd_p;
3816
3817                 if (bbr_get_rtt(bbr, BBR_SRTT)) {
3818                         val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000);
3819                         val /= bbr_get_rtt(bbr, BBR_SRTT);
3820                         ratio = (uint32_t)val;
3821                 } else
3822                         ratio = 1000;
3823
3824                 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div,
3825                                      bbr->r_ctl.recovery_lr, 21,
3826                                      ratio,
3827                                      bbr->r_ctl.rc_red_cwnd_pe,
3828                                      __LINE__);
3829                 if ((ratio < bbr_do_red) || (bbr_do_red == 0))
3830                         goto done;
3831                 if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
3832                      bbr_prtt_slam_cwnd) ||
3833                     (bbr_sub_drain_slam_cwnd &&
3834                      (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
3835                      bbr->rc_hit_state_1 &&
3836                      (bbr_state_val(bbr) == BBR_SUB_DRAIN)) ||
3837                     ((bbr->rc_bbr_state == BBR_STATE_DRAIN) &&
3838                      bbr_slam_cwnd_in_main_drain)) {
3839                         /*
3840                          * Here we must poke at the saved cwnd
3841                          * as well as the cwnd.
3842                          */
3843                         cwnd = bbr->r_ctl.rc_saved_cwnd;
3844                         cwnd_p = &bbr->r_ctl.rc_saved_cwnd;
3845                 } else {
3846                         cwnd = tp->snd_cwnd;
3847                         cwnd_p = &tp->snd_cwnd;
3848                 }
3849                 maxseg = tp->t_maxseg - bbr->rc_last_options;
3850                 /* Add the overall lr with the recovery lr */
3851                 if (bbr->r_ctl.rc_lost == 0)
3852                         lr2use = 0;
3853                 else if (bbr->r_ctl.rc_delivered == 0)
3854                         lr2use = 1000;
3855                 else {
3856                         lr2use = bbr->r_ctl.rc_lost * 1000;
3857                         lr2use /= bbr->r_ctl.rc_delivered;
3858                 }
3859                 lr2use += bbr->r_ctl.recovery_lr;
3860                 acks_inflight = (flight / (maxseg * 2));
3861                 if (bbr_red_scale) {
3862                         lr2use *= bbr_get_rtt(bbr, BBR_SRTT);
3863                         lr2use /= bbr_red_scale;
3864                         if ((bbr_red_growth_restrict) &&
3865                             ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1))
3866                             bbr->r_ctl.restrict_growth += acks_inflight;
3867                 }
3868                 if (lr2use) {
3869                         val = (uint64_t)cwnd * lr2use;
3870                         val /= 1000;
3871                         if (cwnd > val)
3872                                 newcwnd = roundup((cwnd - val), maxseg);
3873                         else
3874                                 newcwnd = maxseg;
3875                 } else {
3876                         val = (uint64_t)cwnd * (uint64_t)bbr_red_mul;
3877                         val /= (uint64_t)bbr_red_div;
3878                         newcwnd = roundup((uint32_t)val, maxseg);
3879                 }
3880                 /* with standard delayed acks how many acks can I expect? */
3881                 if (bbr_drop_limit == 0) {
3882                         /*
3883                          * Anticpate how much we will
3884                          * raise the cwnd based on the acks.
3885                          */
3886                         if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) {
3887                                 /* We do enforce the min (with the acks) */
3888                                 newcwnd = (get_min_cwnd(bbr) - acks_inflight);
3889                         }
3890                 } else {
3891                         /*
3892                          * A strict drop limit of N is is inplace
3893                          */
3894                         if (newcwnd < (bbr_drop_limit * maxseg)) {
3895                                 newcwnd = bbr_drop_limit * maxseg;
3896                         }
3897                 }
3898                 /* For the next N acks do we restrict the growth */
3899                 *cwnd_p = newcwnd;
3900                 if (tp->snd_cwnd > newcwnd)
3901                         tp->snd_cwnd = newcwnd;
3902                 bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22,
3903                                      (uint32_t)lr2use,
3904                                      bbr_get_rtt(bbr, BBR_SRTT), __LINE__);
3905                 bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch;
3906         }
3907 done:
3908         bbr->r_ctl.recovery_lr = 0;
3909         if (flight <= tp->snd_cwnd) {
3910                 bbr->r_wanted_output = 1;
3911         }
3912         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3913 }
3914
3915 static void
3916 bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts)
3917 {
3918         bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
3919         /* Limit the drop in b/w to 1/2 our current filter. */
3920         if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate)
3921                 bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate;
3922         if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2))
3923                 bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2;
3924         tcp_bbr_tso_size_check(bbr, cts);
3925 }
3926
3927 static void
3928 bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm)
3929 {
3930         struct tcp_bbr *bbr;
3931
3932         INP_WLOCK_ASSERT(tp->t_inpcb);
3933         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
3934         switch (type) {
3935         case CC_NDUPACK:
3936                 if (!IN_RECOVERY(tp->t_flags)) {
3937                         tp->snd_recover = tp->snd_max;
3938                         /* Start a new epoch */
3939                         bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
3940                         if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) {
3941                                 /*
3942                                  * Move forward the lt epoch
3943                                  * so it won't count the truncated
3944                                  * epoch.
3945                                  */
3946                                 bbr->r_ctl.rc_lt_epoch++;
3947                         }
3948                         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
3949                                 /*
3950                                  * Just like the policer detection code
3951                                  * if we are in startup we must push
3952                                  * forward the last startup epoch
3953                                  * to hide the truncated PE.
3954                                  */
3955                                 bbr->r_ctl.rc_bbr_last_startup_epoch++;
3956                         }
3957                         bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd;
3958                         ENTER_RECOVERY(tp->t_flags);
3959                         bbr->rc_tlp_rtx_out = 0;
3960                         bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate;
3961                         tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime);
3962                         if (bbr->rc_inp->inp_in_hpts &&
3963                             ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) {
3964                                 /*
3965                                  * When we enter recovery, we need to restart
3966                                  * any timers. This may mean we gain an agg
3967                                  * early, which will be made up for at the last
3968                                  * rxt out.
3969                                  */
3970                                 bbr->rc_timer_first = 1;
3971                                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
3972                         }
3973                         /*
3974                          * Calculate a new cwnd based on to the current
3975                          * delivery rate with no gain. We get the bdp
3976                          * without gaining it up like we normally would and
3977                          * we use the last cur_del_rate.
3978                          */
3979                         if ((bbr->rc_use_google == 0) &&
3980                             (bbr->r_ctl.bbr_rttprobe_gain_val ||
3981                              (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) {
3982                                 tp->snd_cwnd = ctf_flight_size(tp,
3983                                                    (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
3984                                         (tp->t_maxseg - bbr->rc_last_options);
3985                                 if (tp->snd_cwnd < get_min_cwnd(bbr)) {
3986                                         /* We always gate to min cwnd */
3987                                         tp->snd_cwnd = get_min_cwnd(bbr);
3988                                 }
3989                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__);
3990                         }
3991                         bbr_log_type_enter_rec(bbr, rsm->r_start);
3992                 }
3993                 break;
3994         case CC_RTO_ERR:
3995                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
3996                 /* RTO was unnecessary, so reset everything. */
3997                 bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime);
3998                 if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
3999                         tp->snd_cwnd = tp->snd_cwnd_prev;
4000                         tp->snd_ssthresh = tp->snd_ssthresh_prev;
4001                         tp->snd_recover = tp->snd_recover_prev;
4002                         tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent);
4003                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__);
4004                 }
4005                 tp->t_badrxtwin = 0;
4006                 break;
4007         }
4008 }
4009
4010 /*
4011  * Indicate whether this ack should be delayed.  We can delay the ack if
4012  * following conditions are met:
4013  *      - There is no delayed ack timer in progress.
4014  *      - Our last ack wasn't a 0-sized window. We never want to delay
4015  *        the ack that opens up a 0-sized window.
4016  *      - LRO wasn't used for this segment. We make sure by checking that the
4017  *        segment size is not larger than the MSS.
4018  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4019  *        connection.
4020  *      - The data being acked is less than a full segment (a stretch ack
4021  *        of more than a segment we should ack.
4022  *      - nsegs is 1 (if its more than that we received more than 1 ack).
4023  */
4024 #define DELAY_ACK(tp, bbr, nsegs)                               \
4025         (((tp->t_flags & TF_RXWIN0SENT) == 0) &&                \
4026          ((tp->t_flags & TF_DELACK) == 0) &&                    \
4027          ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) &&  \
4028          (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4029
4030 /*
4031  * Return the lowest RSM in the map of
4032  * packets still in flight that is not acked.
4033  * This should normally find on the first one
4034  * since we remove packets from the send
4035  * map after they are marked ACKED.
4036  */
4037 static struct bbr_sendmap *
4038 bbr_find_lowest_rsm(struct tcp_bbr *bbr)
4039 {
4040         struct bbr_sendmap *rsm;
4041
4042         /*
4043          * Walk the time-order transmitted list looking for an rsm that is
4044          * not acked. This will be the one that was sent the longest time
4045          * ago that is still outstanding.
4046          */
4047         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) {
4048                 if (rsm->r_flags & BBR_ACKED) {
4049                         continue;
4050                 }
4051                 goto finish;
4052         }
4053 finish:
4054         return (rsm);
4055 }
4056
4057 static struct bbr_sendmap *
4058 bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
4059 {
4060         struct bbr_sendmap *prsm;
4061
4062         /*
4063          * Walk the sequence order list backward until we hit and arrive at
4064          * the highest seq not acked. In theory when this is called it
4065          * should be the last segment (which it was not).
4066          */
4067         prsm = rsm;
4068         TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4069                 if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) {
4070                         continue;
4071                 }
4072                 return (prsm);
4073         }
4074         return (NULL);
4075 }
4076
4077 /*
4078  * Returns to the caller the number of microseconds that
4079  * the packet can be outstanding before we think we
4080  * should have had an ack returned.
4081  */
4082 static uint32_t
4083 bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm)
4084 {
4085         /*
4086          * lro is the flag we use to determine if we have seen reordering.
4087          * If it gets set we have seen reordering. The reorder logic either
4088          * works in one of two ways:
4089          *
4090          * If reorder-fade is configured, then we track the last time we saw
4091          * re-ordering occur. If we reach the point where enough time as
4092          * passed we no longer consider reordering has occuring.
4093          *
4094          * Or if reorder-face is 0, then once we see reordering we consider
4095          * the connection to alway be subject to reordering and just set lro
4096          * to 1.
4097          *
4098          * In the end if lro is non-zero we add the extra time for
4099          * reordering in.
4100          */
4101         int32_t lro;
4102         uint32_t thresh, t_rxtcur;
4103
4104         if (srtt == 0)
4105                 srtt = 1;
4106         if (bbr->r_ctl.rc_reorder_ts) {
4107                 if (bbr->r_ctl.rc_reorder_fade) {
4108                         if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) {
4109                                 lro = cts - bbr->r_ctl.rc_reorder_ts;
4110                                 if (lro == 0) {
4111                                         /*
4112                                          * No time as passed since the last
4113                                          * reorder, mark it as reordering.
4114                                          */
4115                                         lro = 1;
4116                                 }
4117                         } else {
4118                                 /* Negative time? */
4119                                 lro = 0;
4120                         }
4121                         if (lro > bbr->r_ctl.rc_reorder_fade) {
4122                                 /* Turn off reordering seen too */
4123                                 bbr->r_ctl.rc_reorder_ts = 0;
4124                                 lro = 0;
4125                         }
4126                 } else {
4127                         /* Reodering does not fade */
4128                         lro = 1;
4129                 }
4130         } else {
4131                 lro = 0;
4132         }
4133         thresh = srtt + bbr->r_ctl.rc_pkt_delay;
4134         if (lro) {
4135                 /* It must be set, if not you get 1/4 rtt */
4136                 if (bbr->r_ctl.rc_reorder_shift)
4137                         thresh += (srtt >> bbr->r_ctl.rc_reorder_shift);
4138                 else
4139                         thresh += (srtt >> 2);
4140         } else {
4141                 thresh += 1000;
4142         }
4143         /* We don't let the rack timeout be above a RTO */
4144         if ((bbr->rc_tp)->t_srtt == 0)
4145                 t_rxtcur = BBR_INITIAL_RTO;
4146         else
4147                 t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur);
4148         if (thresh > t_rxtcur) {
4149                 thresh = t_rxtcur;
4150         }
4151         /* And we don't want it above the RTO max either */
4152         if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4153                 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4154         }
4155         bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK);
4156         return (thresh);
4157 }
4158
4159 /*
4160  * Return to the caller the amount of time in mico-seconds
4161  * that should be used for the TLP timer from the last
4162  * send time of this packet.
4163  */
4164 static uint32_t
4165 bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr,
4166     struct bbr_sendmap *rsm, uint32_t srtt,
4167     uint32_t cts)
4168 {
4169         uint32_t thresh, len, maxseg, t_rxtcur;
4170         struct bbr_sendmap *prsm;
4171
4172         if (srtt == 0)
4173                 srtt = 1;
4174         if (bbr->rc_tlp_threshold)
4175                 thresh = srtt + (srtt / bbr->rc_tlp_threshold);
4176         else
4177                 thresh = (srtt * 2);
4178         maxseg = tp->t_maxseg - bbr->rc_last_options;
4179         /* Get the previous sent packet, if any  */
4180         len = rsm->r_end - rsm->r_start;
4181
4182         /* 2.1 behavior */
4183         prsm = TAILQ_PREV(rsm, bbr_head, r_tnext);
4184         if (prsm && (len <= maxseg)) {
4185                 /*
4186                  * Two packets outstanding, thresh should be (2*srtt) +
4187                  * possible inter-packet delay (if any).
4188                  */
4189                 uint32_t inter_gap = 0;
4190                 int idx, nidx;
4191
4192                 idx = rsm->r_rtr_cnt - 1;
4193                 nidx = prsm->r_rtr_cnt - 1;
4194                 if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
4195                         /* Yes it was sent later (or at the same time) */
4196                         inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
4197                 }
4198                 thresh += inter_gap;
4199         } else if (len <= maxseg) {
4200                 /*
4201                  * Possibly compensate for delayed-ack.
4202                  */
4203                 uint32_t alt_thresh;
4204
4205                 alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time;
4206                 if (alt_thresh > thresh)
4207                         thresh = alt_thresh;
4208         }
4209         /* Not above the current  RTO */
4210         if (tp->t_srtt == 0)
4211                 t_rxtcur = BBR_INITIAL_RTO;
4212         else
4213                 t_rxtcur = TICKS_2_USEC(tp->t_rxtcur);
4214
4215         bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP);
4216         /* Not above an RTO */
4217         if (thresh > t_rxtcur) {
4218                 thresh = t_rxtcur;
4219         }
4220         /* Not above a RTO max */
4221         if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) {
4222                 thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND);
4223         }
4224         /* And now apply the user TLP min */
4225         if (thresh < bbr_tlp_min) {
4226                 thresh = bbr_tlp_min;
4227         }
4228         return (thresh);
4229 }
4230
4231 /*
4232  * Return one of three RTTs to use (in microseconds).
4233  */
4234 static __inline uint32_t
4235 bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type)
4236 {
4237         uint32_t f_rtt;
4238         uint32_t srtt;
4239
4240         f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
4241         if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) {
4242                 /* We have no rtt at all */
4243                 if (bbr->rc_tp->t_srtt == 0)
4244                         f_rtt = BBR_INITIAL_RTO;
4245                 else
4246                         f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4247                 /*
4248                  * Since we don't know how good the rtt is apply a
4249                  * delayed-ack min
4250                  */
4251                 if (f_rtt < bbr_delayed_ack_time) {
4252                         f_rtt = bbr_delayed_ack_time;
4253                 }
4254         }
4255         /* Take the filter version or last measured pkt-rtt */
4256         if (rtt_type == BBR_RTT_PROP) {
4257                 srtt = f_rtt;
4258         } else if (rtt_type == BBR_RTT_PKTRTT) {
4259                 if (bbr->r_ctl.rc_pkt_epoch_rtt) {
4260                         srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
4261                 } else {
4262                         /* No pkt rtt yet */
4263                         srtt = f_rtt;
4264                 }
4265         } else if (rtt_type == BBR_RTT_RACK) {
4266                 srtt = bbr->r_ctl.rc_last_rtt;
4267                 /* We need to add in any internal delay for our timer */
4268                 if (bbr->rc_ack_was_delayed)
4269                         srtt += bbr->r_ctl.rc_ack_hdwr_delay;
4270         } else if (rtt_type == BBR_SRTT) {
4271                 srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
4272         } else {
4273                 /* TSNH */
4274                 srtt = f_rtt;
4275 #ifdef BBR_INVARIANTS
4276                 panic("Unknown rtt request type %d", rtt_type);
4277 #endif
4278         }
4279         return (srtt);
4280 }
4281
4282 static int
4283 bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts)
4284 {
4285         uint32_t thresh;
4286
4287         thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK),
4288                                       cts, rsm);
4289         if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) {
4290                 /* It is lost (past time) */
4291                 return (1);
4292         }
4293         return (0);
4294 }
4295
4296 /*
4297  * Return a sendmap if we need to retransmit something.
4298  */
4299 static struct bbr_sendmap *
4300 bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4301 {
4302         /*
4303          * Check to see that we don't need to fall into recovery. We will
4304          * need to do so if our oldest transmit is past the time we should
4305          * have had an ack.
4306          */
4307
4308         struct bbr_sendmap *rsm;
4309         int32_t idx;
4310
4311         if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) {
4312                 /* Nothing outstanding that we know of */
4313                 return (NULL);
4314         }
4315         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
4316         if (rsm == NULL) {
4317                 /* Nothing in the transmit map */
4318                 return (NULL);
4319         }
4320         if (tp->t_flags & TF_SENTFIN) {
4321                 /* Fin restricted, don't find anything once a fin is sent */
4322                 return (NULL);
4323         }
4324         if (rsm->r_flags & BBR_ACKED) {
4325                 /*
4326                  * Ok the first one is acked (this really should not happen
4327                  * since we remove the from the tmap once they are acked)
4328                  */
4329                 rsm = bbr_find_lowest_rsm(bbr);
4330                 if (rsm == NULL)
4331                         return (NULL);
4332         }
4333         idx = rsm->r_rtr_cnt - 1;
4334         if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) {
4335                 /* Send timestamp is the same or less? can't be ready */
4336                 return (NULL);
4337         }
4338         /* Get our RTT time */
4339         if (bbr_is_lost(bbr, rsm, cts) &&
4340             ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
4341              (rsm->r_flags & BBR_SACK_PASSED))) {
4342                 if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4343                         rsm->r_flags |= BBR_MARKED_LOST;
4344                         bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4345                         bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4346                 }
4347                 bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm);
4348 #ifdef BBR_INVARIANTS
4349                 if ((rsm->r_end - rsm->r_start) == 0)
4350                         panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm);
4351 #endif
4352                 return (rsm);
4353         }
4354         return (NULL);
4355 }
4356
4357 /*
4358  * RACK Timer, here we simply do logging and house keeping.
4359  * the normal bbr_output_wtime() function will call the
4360  * appropriate thing to check if we need to do a RACK retransmit.
4361  * We return 1, saying don't proceed with bbr_output_wtime only
4362  * when all timers have been stopped (destroyed PCB?).
4363  */
4364 static int
4365 bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4366 {
4367         /*
4368          * This timer simply provides an internal trigger to send out data.
4369          * The check_recovery_mode call will see if there are needed
4370          * retransmissions, if so we will enter fast-recovery. The output
4371          * call may or may not do the same thing depending on sysctl
4372          * settings.
4373          */
4374         uint32_t lost;
4375
4376         if (bbr->rc_all_timers_stopped) {
4377                 return (1);
4378         }
4379         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4380                 /* Its not time yet */
4381                 return (0);
4382         }
4383         BBR_STAT_INC(bbr_to_tot);
4384         lost = bbr->r_ctl.rc_lost;
4385         if (bbr->r_state && (bbr->r_state != tp->t_state))
4386                 bbr_set_state(tp, bbr, 0);
4387         bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK);
4388         if (bbr->r_ctl.rc_resend == NULL) {
4389                 /* Lets do the check here */
4390                 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
4391         }
4392         if (bbr_policer_call_from_rack_to)
4393                 bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4394         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
4395         return (0);
4396 }
4397
4398 static __inline void
4399 bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start)
4400 {
4401         int idx;
4402
4403         nrsm->r_start = start;
4404         nrsm->r_end = rsm->r_end;
4405         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
4406         nrsm->r_flags = rsm->r_flags;
4407         /* We don't transfer forward the SYN flag */
4408         nrsm->r_flags &= ~BBR_HAS_SYN;
4409         /* We move forward the FIN flag, not that this should happen */
4410         rsm->r_flags &= ~BBR_HAS_FIN;
4411         nrsm->r_dupack = rsm->r_dupack;
4412         nrsm->r_rtr_bytes = 0;
4413         nrsm->r_is_gain = rsm->r_is_gain;
4414         nrsm->r_is_drain = rsm->r_is_drain;
4415         nrsm->r_delivered = rsm->r_delivered;
4416         nrsm->r_ts_valid = rsm->r_ts_valid;
4417         nrsm->r_del_ack_ts = rsm->r_del_ack_ts;
4418         nrsm->r_del_time = rsm->r_del_time;
4419         nrsm->r_app_limited = rsm->r_app_limited;
4420         nrsm->r_first_sent_time = rsm->r_first_sent_time;
4421         nrsm->r_flight_at_send = rsm->r_flight_at_send;
4422         /* We split a piece the lower section looses any just_ret flag. */
4423         nrsm->r_bbr_state = rsm->r_bbr_state;
4424         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
4425                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
4426         }
4427         rsm->r_end = nrsm->r_start;
4428         idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs);
4429         idx /= 8;
4430         /* Check if we got too small */
4431         if ((rsm->r_is_smallmap == 0) &&
4432             ((rsm->r_end - rsm->r_start) <= idx)) {
4433                 bbr->r_ctl.rc_num_small_maps_alloced++;
4434                 rsm->r_is_smallmap = 1;
4435         }
4436         /* Check the new one as well */
4437         if ((nrsm->r_end - nrsm->r_start) <= idx) {
4438                 bbr->r_ctl.rc_num_small_maps_alloced++;
4439                 nrsm->r_is_smallmap = 1;
4440         }
4441 }
4442
4443 static int
4444 bbr_sack_mergable(struct bbr_sendmap *at,
4445                   uint32_t start, uint32_t end)
4446 {
4447         /*
4448          * Given a sack block defined by
4449          * start and end, and a current postion
4450          * at. Return 1 if either side of at
4451          * would show that the block is mergable
4452          * to that side. A block to be mergable
4453          * must have overlap with the start/end
4454          * and be in the SACK'd state.
4455          */
4456         struct bbr_sendmap *l_rsm;
4457         struct bbr_sendmap *r_rsm;
4458
4459         /* first get the either side blocks */
4460         l_rsm = TAILQ_PREV(at, bbr_head, r_next);
4461         r_rsm = TAILQ_NEXT(at, r_next);
4462         if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) {
4463                 /* Potentially mergeable */
4464                 if ((l_rsm->r_end == start) ||
4465                     (SEQ_LT(start, l_rsm->r_end) &&
4466                      SEQ_GT(end, l_rsm->r_end))) {
4467                             /*
4468                              * map blk   |------|
4469                              * sack blk         |------|
4470                              * <or>
4471                              * map blk   |------|
4472                              * sack blk      |------|
4473                              */
4474                             return (1);
4475                     }
4476         }
4477         if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) {
4478                 /* Potentially mergeable */
4479                 if ((r_rsm->r_start == end) ||
4480                     (SEQ_LT(start, r_rsm->r_start) &&
4481                      SEQ_GT(end, r_rsm->r_start))) {
4482                         /*
4483                          * map blk          |---------|
4484                          * sack blk    |----|
4485                          * <or>
4486                          * map blk          |---------|
4487                          * sack blk    |-------|
4488                          */
4489                         return (1);
4490                 }
4491         }
4492         return (0);
4493 }
4494
4495 static struct bbr_sendmap *
4496 bbr_merge_rsm(struct tcp_bbr *bbr,
4497               struct bbr_sendmap *l_rsm,
4498               struct bbr_sendmap *r_rsm)
4499 {
4500         /*
4501          * We are merging two ack'd RSM's,
4502          * the l_rsm is on the left (lower seq
4503          * values) and the r_rsm is on the right
4504          * (higher seq value). The simplest way
4505          * to merge these is to move the right
4506          * one into the left. I don't think there
4507          * is any reason we need to try to find
4508          * the oldest (or last oldest retransmitted).
4509          */
4510         l_rsm->r_end = r_rsm->r_end;
4511         if (l_rsm->r_dupack < r_rsm->r_dupack)
4512                 l_rsm->r_dupack = r_rsm->r_dupack;
4513         if (r_rsm->r_rtr_bytes)
4514                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
4515         if (r_rsm->r_in_tmap) {
4516                 /* This really should not happen */
4517                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext);
4518         }
4519         if (r_rsm->r_app_limited)
4520                 l_rsm->r_app_limited = r_rsm->r_app_limited;
4521         /* Now the flags */
4522         if (r_rsm->r_flags & BBR_HAS_FIN)
4523                 l_rsm->r_flags |= BBR_HAS_FIN;
4524         if (r_rsm->r_flags & BBR_TLP)
4525                 l_rsm->r_flags |= BBR_TLP;
4526         if (r_rsm->r_flags & BBR_RWND_COLLAPSED)
4527                 l_rsm->r_flags |= BBR_RWND_COLLAPSED;
4528         if (r_rsm->r_flags & BBR_MARKED_LOST) {
4529                 /* This really should not happen */
4530                 bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start;
4531         }
4532         TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next);
4533         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
4534                 /* Transfer the split limit to the map we free */
4535                 r_rsm->r_limit_type = l_rsm->r_limit_type;
4536                 l_rsm->r_limit_type = 0;
4537         }
4538         bbr_free(bbr, r_rsm);
4539         return(l_rsm);
4540 }
4541
4542 /*
4543  * TLP Timer, here we simply setup what segment we want to
4544  * have the TLP expire on, the normal bbr_output_wtime() will then
4545  * send it out.
4546  *
4547  * We return 1, saying don't proceed with bbr_output_wtime only
4548  * when all timers have been stopped (destroyed PCB?).
4549  */
4550 static int
4551 bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4552 {
4553         /*
4554          * Tail Loss Probe.
4555          */
4556         struct bbr_sendmap *rsm = NULL;
4557         struct socket *so;
4558         uint32_t amm;
4559         uint32_t out, avail;
4560         uint32_t maxseg;
4561         int collapsed_win = 0;
4562
4563         if (bbr->rc_all_timers_stopped) {
4564                 return (1);
4565         }
4566         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
4567                 /* Its not time yet */
4568                 return (0);
4569         }
4570         if (ctf_progress_timeout_check(tp, true)) {
4571                 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4572                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4573                 return (1);
4574         }
4575         /* Did we somehow get into persists? */
4576         if (bbr->rc_in_persist) {
4577                 return (0);
4578         }
4579         if (bbr->r_state && (bbr->r_state != tp->t_state))
4580                 bbr_set_state(tp, bbr, 0);
4581         BBR_STAT_INC(bbr_tlp_tot);
4582         maxseg = tp->t_maxseg - bbr->rc_last_options;
4583         /*
4584          * A TLP timer has expired. We have been idle for 2 rtts. So we now
4585          * need to figure out how to force a full MSS segment out.
4586          */
4587         so = tp->t_inpcb->inp_socket;
4588         avail = sbavail(&so->so_snd);
4589         out = ctf_outstanding(tp);
4590         if (out > tp->snd_wnd) {
4591                 /* special case, we need a retransmission */
4592                 collapsed_win = 1;
4593                 goto need_retran;
4594         }
4595         if (avail > out) {
4596                 /* New data is available */
4597                 amm = avail - out;
4598                 if (amm > maxseg) {
4599                         amm = maxseg;
4600                 } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) {
4601                         /* not enough to fill a MTU and no-delay is off */
4602                         goto need_retran;
4603                 }
4604                 /* Set the send-new override */
4605                 if ((out + amm) <= tp->snd_wnd) {
4606                         bbr->rc_tlp_new_data = 1;
4607                 } else {
4608                         goto need_retran;
4609                 }
4610                 bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4611                 bbr->r_ctl.rc_last_tlp_seq = tp->snd_max;
4612                 bbr->r_ctl.rc_tlp_send = NULL;
4613                 /* cap any slots */
4614                 BBR_STAT_INC(bbr_tlp_newdata);
4615                 goto send;
4616         }
4617 need_retran:
4618         /*
4619          * Ok we need to arrange the last un-acked segment to be re-sent, or
4620          * optionally the first un-acked segment.
4621          */
4622         if (collapsed_win == 0) {
4623                 rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
4624                 if (rsm && (BBR_ACKED | BBR_HAS_FIN)) {
4625                         rsm = bbr_find_high_nonack(bbr, rsm);
4626                 }
4627                 if (rsm == NULL) {
4628                         goto restore;
4629                 }
4630         } else {
4631                 /*
4632                  * We must find the last segment
4633                  * that was acceptable by the client.
4634                  */
4635                 TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
4636                         if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) {
4637                                 /* Found one */
4638                                 break;
4639                         }
4640                 }
4641                 if (rsm == NULL) {
4642                         /* None? if so send the first */
4643                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4644                         if (rsm == NULL)
4645                                 goto restore;
4646                 }
4647         }
4648         if ((rsm->r_end - rsm->r_start) > maxseg) {
4649                 /*
4650                  * We need to split this the last segment in two.
4651                  */
4652                 struct bbr_sendmap *nrsm;
4653
4654                 nrsm = bbr_alloc_full_limit(bbr);
4655                 if (nrsm == NULL) {
4656                         /*
4657                          * We can't get memory to split, we can either just
4658                          * not split it. Or retransmit the whole piece, lets
4659                          * do the large send (BTLP :-) ).
4660                          */
4661                         goto go_for_it;
4662                 }
4663                 bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg));
4664                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
4665                 if (rsm->r_in_tmap) {
4666                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
4667                         nrsm->r_in_tmap = 1;
4668                 }
4669                 rsm->r_flags &= (~BBR_HAS_FIN);
4670                 rsm = nrsm;
4671         }
4672 go_for_it:
4673         bbr->r_ctl.rc_tlp_send = rsm;
4674         bbr->rc_tlp_rtx_out = 1;
4675         if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) {
4676                 bbr->r_ctl.rc_tlp_seg_send_cnt++;
4677                 tp->t_rxtshift++;
4678         } else {
4679                 bbr->r_ctl.rc_last_tlp_seq = rsm->r_start;
4680                 bbr->r_ctl.rc_tlp_seg_send_cnt = 1;
4681         }
4682 send:
4683         if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) {
4684                 /*
4685                  * Can't [re]/transmit a segment we have retranmitted the
4686                  * max times. We need the retransmit timer to take over.
4687                  */
4688 restore:
4689                 bbr->rc_tlp_new_data = 0;
4690                 bbr->r_ctl.rc_tlp_send = NULL;
4691                 if (rsm)
4692                         rsm->r_flags &= ~BBR_TLP;
4693                 BBR_STAT_INC(bbr_tlp_retran_fail);
4694                 return (0);
4695         } else if (rsm) {
4696                 rsm->r_flags |= BBR_TLP;
4697         }
4698         if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) &&
4699             (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) {
4700                 /*
4701                  * We have retransmitted to many times for TLP. Switch to
4702                  * the regular RTO timer
4703                  */
4704                 goto restore;
4705         }
4706         bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP);
4707         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
4708         return (0);
4709 }
4710
4711 /*
4712  * Delayed ack Timer, here we simply need to setup the
4713  * ACK_NOW flag and remove the DELACK flag. From there
4714  * the output routine will send the ack out.
4715  *
4716  * We only return 1, saying don't proceed, if all timers
4717  * are stopped (destroyed PCB?).
4718  */
4719 static int
4720 bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4721 {
4722         if (bbr->rc_all_timers_stopped) {
4723                 return (1);
4724         }
4725         bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK);
4726         tp->t_flags &= ~TF_DELACK;
4727         tp->t_flags |= TF_ACKNOW;
4728         KMOD_TCPSTAT_INC(tcps_delack);
4729         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
4730         return (0);
4731 }
4732
4733 /*
4734  * Here we send a KEEP-ALIVE like probe to the
4735  * peer, we do not send data.
4736  *
4737  * We only return 1, saying don't proceed, if all timers
4738  * are stopped (destroyed PCB?).
4739  */
4740 static int
4741 bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4742 {
4743         struct tcptemp *t_template;
4744         int32_t retval = 1;
4745
4746         if (bbr->rc_all_timers_stopped) {
4747                 return (1);
4748         }
4749         if (bbr->rc_in_persist == 0)
4750                 return (0);
4751         KASSERT(tp->t_inpcb != NULL,
4752             ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
4753         /*
4754          * Persistence timer into zero window. Force a byte to be output, if
4755          * possible.
4756          */
4757         bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST);
4758         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
4759         KMOD_TCPSTAT_INC(tcps_persisttimeo);
4760         /*
4761          * Have we exceeded the user specified progress time?
4762          */
4763         if (ctf_progress_timeout_check(tp, true)) {
4764                 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4765                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4766                 goto out;
4767         }
4768         /*
4769          * Hack: if the peer is dead/unreachable, we do not time out if the
4770          * window is closed.  After a full backoff, drop the connection if
4771          * the idle time (no responses to probes) reaches the maximum
4772          * backoff that we would use if retransmitting.
4773          */
4774         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
4775             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
4776             ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
4777                 KMOD_TCPSTAT_INC(tcps_persistdrop);
4778                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4779                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4780                 goto out;
4781         }
4782         if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) &&
4783             tp->snd_una == tp->snd_max) {
4784                 bbr_exit_persist(tp, bbr, cts, __LINE__);
4785                 retval = 0;
4786                 goto out;
4787         }
4788         /*
4789          * If the user has closed the socket then drop a persisting
4790          * connection after a much reduced timeout.
4791          */
4792         if (tp->t_state > TCPS_CLOSE_WAIT &&
4793             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
4794                 KMOD_TCPSTAT_INC(tcps_persistdrop);
4795                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
4796                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4797                 goto out;
4798         }
4799         t_template = tcpip_maketemplate(bbr->rc_inp);
4800         if (t_template) {
4801                 tcp_respond(tp, t_template->tt_ipgen,
4802                             &t_template->tt_t, (struct mbuf *)NULL,
4803                             tp->rcv_nxt, tp->snd_una - 1, 0);
4804                 /* This sends an ack */
4805                 if (tp->t_flags & TF_DELACK)
4806                         tp->t_flags &= ~TF_DELACK;
4807                 free(t_template, M_TEMP);
4808         }
4809         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
4810                 tp->t_rxtshift++;
4811         bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0);
4812 out:
4813         return (retval);
4814 }
4815
4816 /*
4817  * If a keepalive goes off, we had no other timers
4818  * happening. We always return 1 here since this
4819  * routine either drops the connection or sends
4820  * out a segment with respond.
4821  */
4822 static int
4823 bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4824 {
4825         struct tcptemp *t_template;
4826         struct inpcb *inp;
4827
4828         if (bbr->rc_all_timers_stopped) {
4829                 return (1);
4830         }
4831         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
4832         inp = tp->t_inpcb;
4833         bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP);
4834         /*
4835          * Keep-alive timer went off; send something or drop connection if
4836          * idle for too long.
4837          */
4838         KMOD_TCPSTAT_INC(tcps_keeptimeo);
4839         if (tp->t_state < TCPS_ESTABLISHED)
4840                 goto dropit;
4841         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
4842             tp->t_state <= TCPS_CLOSING) {
4843                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
4844                         goto dropit;
4845                 /*
4846                  * Send a packet designed to force a response if the peer is
4847                  * up and reachable: either an ACK if the connection is
4848                  * still alive, or an RST if the peer has closed the
4849                  * connection due to timeout or reboot. Using sequence
4850                  * number tp->snd_una-1 causes the transmitted zero-length
4851                  * segment to lie outside the receive window; by the
4852                  * protocol spec, this requires the correspondent TCP to
4853                  * respond.
4854                  */
4855                 KMOD_TCPSTAT_INC(tcps_keepprobe);
4856                 t_template = tcpip_maketemplate(inp);
4857                 if (t_template) {
4858                         tcp_respond(tp, t_template->tt_ipgen,
4859                             &t_template->tt_t, (struct mbuf *)NULL,
4860                             tp->rcv_nxt, tp->snd_una - 1, 0);
4861                         free(t_template, M_TEMP);
4862                 }
4863         }
4864         bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0);
4865         return (1);
4866 dropit:
4867         KMOD_TCPSTAT_INC(tcps_keepdrops);
4868         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
4869         tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4870         return (1);
4871 }
4872
4873 /*
4874  * Retransmit helper function, clear up all the ack
4875  * flags and take care of important book keeping.
4876  */
4877 static void
4878 bbr_remxt_tmr(struct tcpcb *tp)
4879 {
4880         /*
4881          * The retransmit timer went off, all sack'd blocks must be
4882          * un-acked.
4883          */
4884         struct bbr_sendmap *rsm, *trsm = NULL;
4885         struct tcp_bbr *bbr;
4886         uint32_t cts, lost;
4887
4888         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
4889         cts = tcp_get_usecs(&bbr->rc_tv);
4890         lost = bbr->r_ctl.rc_lost;
4891         if (bbr->r_state && (bbr->r_state != tp->t_state))
4892                 bbr_set_state(tp, bbr, 0);
4893
4894         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
4895                 if (rsm->r_flags & BBR_ACKED) {
4896                         uint32_t old_flags;
4897
4898                         rsm->r_dupack = 0;
4899                         if (rsm->r_in_tmap == 0) {
4900                                 /* We must re-add it back to the tlist */
4901                                 if (trsm == NULL) {
4902                                         TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
4903                                 } else {
4904                                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext);
4905                                 }
4906                                 rsm->r_in_tmap = 1;
4907                         }
4908                         old_flags = rsm->r_flags;
4909                         rsm->r_flags |= BBR_RXT_CLEARED;
4910                         rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS);
4911                         bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
4912                 } else {
4913                         if ((tp->t_state < TCPS_ESTABLISHED) &&
4914                             (rsm->r_start == tp->snd_una)) {
4915                                 /*
4916                                  * Special case for TCP FO. Where
4917                                  * we sent more data beyond the snd_max.
4918                                  * We don't mark that as lost and stop here.
4919                                  */
4920                                 break;
4921                         }
4922                         if ((rsm->r_flags & BBR_MARKED_LOST) == 0) {
4923                                 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
4924                                 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
4925                         }
4926                         if (bbr_marks_rxt_sack_passed) {
4927                                 /*
4928                                  * With this option, we will rack out
4929                                  * in 1ms increments the rest of the packets.
4930                                  */
4931                                 rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST;
4932                                 rsm->r_flags &= ~BBR_WAS_SACKPASS;
4933                         } else {
4934                                 /*
4935                                  * With this option we only mark them lost
4936                                  * and remove all sack'd markings. We will run
4937                                  * another RXT or a TLP. This will cause
4938                                  * us to eventually send more based on what
4939                                  * ack's come in.
4940                                  */
4941                                 rsm->r_flags |= BBR_MARKED_LOST;
4942                                 rsm->r_flags &= ~BBR_WAS_SACKPASS;
4943                                 rsm->r_flags &= ~BBR_SACK_PASSED;
4944                         }
4945                 }
4946                 trsm = rsm;
4947         }
4948         bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map);
4949         /* Clear the count (we just un-acked them) */
4950         bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR);
4951         bbr->rc_tlp_new_data = 0;
4952         bbr->r_ctl.rc_tlp_seg_send_cnt = 0;
4953         /* zap the behindness on a rxt */
4954         bbr->r_ctl.rc_hptsi_agg_delay = 0;
4955         bbr->r_agg_early_set = 0;
4956         bbr->r_ctl.rc_agg_early = 0;
4957         bbr->rc_tlp_rtx_out = 0;
4958         bbr->r_ctl.rc_sacked = 0;
4959         bbr->r_ctl.rc_sacklast = NULL;
4960         bbr->r_timer_override = 1;
4961         bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost));
4962 }
4963
4964 /*
4965  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
4966  * we will setup to retransmit the lowest seq number outstanding.
4967  */
4968 static int
4969 bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
4970 {
4971         int32_t rexmt;
4972         int32_t retval = 0;
4973         bool isipv6;
4974
4975         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
4976         if (bbr->rc_all_timers_stopped) {
4977                 return (1);
4978         }
4979         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
4980             (tp->snd_una == tp->snd_max)) {
4981                 /* Nothing outstanding .. nothing to do */
4982                 return (0);
4983         }
4984         /*
4985          * Retransmission timer went off.  Message has not been acked within
4986          * retransmit interval.  Back off to a longer retransmit interval
4987          * and retransmit one segment.
4988          */
4989         if (ctf_progress_timeout_check(tp, true)) {
4990                 retval = 1;
4991                 bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
4992                 tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT);
4993                 goto out;
4994         }
4995         bbr_remxt_tmr(tp);
4996         if ((bbr->r_ctl.rc_resend == NULL) ||
4997             ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) {
4998                 /*
4999                  * If the rwnd collapsed on
5000                  * the one we are retransmitting
5001                  * it does not count against the
5002                  * rxt count.
5003                  */
5004                 tp->t_rxtshift++;
5005         }
5006         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
5007                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
5008                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
5009                 retval = 1;
5010                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
5011                 tcp_set_inp_to_drop(bbr->rc_inp,
5012                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
5013                 goto out;
5014         }
5015         if (tp->t_state == TCPS_SYN_SENT) {
5016                 /*
5017                  * If the SYN was retransmitted, indicate CWND to be limited
5018                  * to 1 segment in cc_conn_init().
5019                  */
5020                 tp->snd_cwnd = 1;
5021         } else if (tp->t_rxtshift == 1) {
5022                 /*
5023                  * first retransmit; record ssthresh and cwnd so they can be
5024                  * recovered if this turns out to be a "bad" retransmit. A
5025                  * retransmit is considered "bad" if an ACK for this segment
5026                  * is received within RTT/2 interval; the assumption here is
5027                  * that the ACK was already in flight.  See "On Estimating
5028                  * End-to-End Network Path Properties" by Allman and Paxson
5029                  * for more details.
5030                  */
5031                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5032                 if (!IN_RECOVERY(tp->t_flags)) {
5033                         tp->snd_cwnd_prev = tp->snd_cwnd;
5034                         tp->snd_ssthresh_prev = tp->snd_ssthresh;
5035                         tp->snd_recover_prev = tp->snd_recover;
5036                         tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
5037                         tp->t_flags |= TF_PREVVALID;
5038                 } else {
5039                         tp->t_flags &= ~TF_PREVVALID;
5040                 }
5041                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5042         } else {
5043                 tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options;
5044                 tp->t_flags &= ~TF_PREVVALID;
5045         }
5046         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
5047         if ((tp->t_state == TCPS_SYN_SENT) ||
5048             (tp->t_state == TCPS_SYN_RECEIVED))
5049                 rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift];
5050         else
5051                 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
5052         TCPT_RANGESET(tp->t_rxtcur, rexmt,
5053             MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms),
5054             MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
5055         /*
5056          * We enter the path for PLMTUD if connection is established or, if
5057          * connection is FIN_WAIT_1 status, reason for the last is that if
5058          * amount of data we send is very small, we could send it in couple
5059          * of packets and process straight to FIN. In that case we won't
5060          * catch ESTABLISHED state.
5061          */
5062 #ifdef INET6
5063         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
5064 #else
5065         isipv6 = false;
5066 #endif
5067         if (((V_tcp_pmtud_blackhole_detect == 1) ||
5068             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
5069             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
5070             ((tp->t_state == TCPS_ESTABLISHED) ||
5071             (tp->t_state == TCPS_FIN_WAIT_1))) {
5072                 /*
5073                  * Idea here is that at each stage of mtu probe (usually,
5074                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
5075                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
5076                  * should take care of that.
5077                  */
5078                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
5079                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
5080                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
5081                     tp->t_rxtshift % 2 == 0)) {
5082                         /*
5083                          * Enter Path MTU Black-hole Detection mechanism: -
5084                          * Disable Path MTU Discovery (IP "DF" bit). -
5085                          * Reduce MTU to lower value than what we negotiated
5086                          * with peer.
5087                          */
5088                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
5089                                 /*
5090                                  * Record that we may have found a black
5091                                  * hole.
5092                                  */
5093                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
5094                                 /* Keep track of previous MSS. */
5095                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
5096                         }
5097                         /*
5098                          * Reduce the MSS to blackhole value or to the
5099                          * default in an attempt to retransmit.
5100                          */
5101 #ifdef INET6
5102                         isipv6 = bbr->r_is_v6;
5103                         if (isipv6 &&
5104                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
5105                                 /* Use the sysctl tuneable blackhole MSS. */
5106                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
5107                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5108                         } else if (isipv6) {
5109                                 /* Use the default MSS. */
5110                                 tp->t_maxseg = V_tcp_v6mssdflt;
5111                                 /*
5112                                  * Disable Path MTU Discovery when we switch
5113                                  * to minmss.
5114                                  */
5115                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5116                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5117                         }
5118 #endif
5119 #if defined(INET6) && defined(INET)
5120                         else
5121 #endif
5122 #ifdef INET
5123                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
5124                                 /* Use the sysctl tuneable blackhole MSS. */
5125                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
5126                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
5127                         } else {
5128                                 /* Use the default MSS. */
5129                                 tp->t_maxseg = V_tcp_mssdflt;
5130                                 /*
5131                                  * Disable Path MTU Discovery when we switch
5132                                  * to minmss.
5133                                  */
5134                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
5135                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
5136                         }
5137 #endif
5138                 } else {
5139                         /*
5140                          * If further retransmissions are still unsuccessful
5141                          * with a lowered MTU, maybe this isn't a blackhole
5142                          * and we restore the previous MSS and blackhole
5143                          * detection flags. The limit '6' is determined by
5144                          * giving each probe stage (1448, 1188, 524) 2
5145                          * chances to recover.
5146                          */
5147                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
5148                             (tp->t_rxtshift >= 6)) {
5149                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
5150                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
5151                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
5152                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
5153                         }
5154                 }
5155         }
5156         /*
5157          * Disable RFC1323 and SACK if we haven't got any response to our
5158          * third SYN to work-around some broken terminal servers (most of
5159          * which have hopefully been retired) that have bad VJ header
5160          * compression code which trashes TCP segments containing
5161          * unknown-to-them TCP options.
5162          */
5163         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
5164             (tp->t_rxtshift == 3))
5165                 tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT);
5166         /*
5167          * If we backed off this far, our srtt estimate is probably bogus.
5168          * Clobber it so we'll take the next rtt measurement as our srtt;
5169          * move the current srtt into rttvar to keep the current retransmit
5170          * times until then.
5171          */
5172         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
5173 #ifdef INET6
5174                 if (bbr->r_is_v6)
5175                         in6_losing(tp->t_inpcb);
5176                 else
5177 #endif
5178                         in_losing(tp->t_inpcb);
5179                 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
5180                 tp->t_srtt = 0;
5181         }
5182         sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
5183         tp->snd_recover = tp->snd_max;
5184         tp->t_flags |= TF_ACKNOW;
5185         tp->t_rtttime = 0;
5186 out:
5187         return (retval);
5188 }
5189
5190 static int
5191 bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling)
5192 {
5193         int32_t ret = 0;
5194         int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
5195
5196         if (timers == 0) {
5197                 return (0);
5198         }
5199         if (tp->t_state == TCPS_LISTEN) {
5200                 /* no timers on listen sockets */
5201                 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
5202                         return (0);
5203                 return (1);
5204         }
5205         if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) {
5206                 uint32_t left;
5207
5208                 if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
5209                         ret = -1;
5210                         bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5211                         return (0);
5212                 }
5213                 if (hpts_calling == 0) {
5214                         ret = -2;
5215                         bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling);
5216                         return (0);
5217                 }
5218                 /*
5219                  * Ok our timer went off early and we are not paced false
5220                  * alarm, go back to sleep.
5221                  */
5222                 left = bbr->r_ctl.rc_timer_exp - cts;
5223                 ret = -3;
5224                 bbr_log_to_processing(bbr, cts, ret, left, hpts_calling);
5225                 tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left));
5226                 return (1);
5227         }
5228         bbr->rc_tmr_stopped = 0;
5229         bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
5230         if (timers & PACE_TMR_DELACK) {
5231                 ret = bbr_timeout_delack(tp, bbr, cts);
5232         } else if (timers & PACE_TMR_PERSIT) {
5233                 ret = bbr_timeout_persist(tp, bbr, cts);
5234         } else if (timers & PACE_TMR_RACK) {
5235                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5236                 ret = bbr_timeout_rack(tp, bbr, cts);
5237         } else if (timers & PACE_TMR_TLP) {
5238                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5239                 ret = bbr_timeout_tlp(tp, bbr, cts);
5240         } else if (timers & PACE_TMR_RXT) {
5241                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
5242                 ret = bbr_timeout_rxt(tp, bbr, cts);
5243         } else if (timers & PACE_TMR_KEEP) {
5244                 ret = bbr_timeout_keepalive(tp, bbr, cts);
5245         }
5246         bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling);
5247         return (ret);
5248 }
5249
5250 static void
5251 bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts)
5252 {
5253         if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
5254                 uint8_t hpts_removed = 0;
5255
5256                 if (bbr->rc_inp->inp_in_hpts &&
5257                     (bbr->rc_timer_first == 1)) {
5258                         /*
5259                          * If we are canceling timer's when we have the
5260                          * timer ahead of the output being paced. We also
5261                          * must remove ourselves from the hpts.
5262                          */
5263                         hpts_removed = 1;
5264                         tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
5265                         if (bbr->r_ctl.rc_last_delay_val) {
5266                                 /* Update the last hptsi delay too */
5267                                 uint32_t time_since_send;
5268
5269                                 if (TSTMP_GT(cts, bbr->rc_pacer_started))
5270                                         time_since_send = cts - bbr->rc_pacer_started;
5271                                 else
5272                                         time_since_send = 0;
5273                                 if (bbr->r_ctl.rc_last_delay_val > time_since_send) {
5274                                         /* Cut down our slot time */
5275                                         bbr->r_ctl.rc_last_delay_val -= time_since_send;
5276                                 } else {
5277                                         bbr->r_ctl.rc_last_delay_val = 0;
5278                                 }
5279                                 bbr->rc_pacer_started = cts;
5280                         }
5281                 }
5282                 bbr->rc_timer_first = 0;
5283                 bbr_log_to_cancel(bbr, line, cts, hpts_removed);
5284                 bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
5285                 bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
5286         }
5287 }
5288
5289 static void
5290 bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type)
5291 {
5292         struct tcp_bbr *bbr;
5293
5294         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
5295         bbr->rc_all_timers_stopped = 1;
5296         return;
5297 }
5298
5299 /*
5300  * stop all timers always returning 0.
5301  */
5302 static int
5303 bbr_stopall(struct tcpcb *tp)
5304 {
5305         return (0);
5306 }
5307
5308 static void
5309 bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
5310 {
5311         return;
5312 }
5313
5314 /*
5315  * return true if a bbr timer (rack or tlp) is active.
5316  */
5317 static int
5318 bbr_timer_active(struct tcpcb *tp, uint32_t timer_type)
5319 {
5320         return (0);
5321 }
5322
5323 static uint32_t
5324 bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts)
5325 {
5326         struct bbr_sendmap *rsm;
5327
5328         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
5329         if ((rsm == NULL) || (u_rsm == rsm))
5330                 return (cts);
5331         return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
5332 }
5333
5334 static void
5335 bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr,
5336      struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time)
5337 {
5338         int32_t idx;
5339
5340         rsm->r_rtr_cnt++;
5341         rsm->r_dupack = 0;
5342         if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) {
5343                 rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS;
5344                 rsm->r_flags |= BBR_OVERMAX;
5345         }
5346         if (rsm->r_flags & BBR_RWND_COLLAPSED) {
5347                 /* Take off the collapsed flag at rxt */
5348                 rsm->r_flags &= ~BBR_RWND_COLLAPSED;
5349         }
5350         if (rsm->r_flags & BBR_MARKED_LOST) {
5351                 /* We have retransmitted, its no longer lost */
5352                 rsm->r_flags &= ~BBR_MARKED_LOST;
5353                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
5354         }
5355         if (rsm->r_flags & BBR_RXT_CLEARED) {
5356                 /*
5357                  * We hit a RXT timer on it and
5358                  * we cleared the "acked" flag.
5359                  * We now have it going back into
5360                  * flight, we can remove the cleared
5361                  * flag and possibly do accounting on
5362                  * this piece.
5363                  */
5364                 rsm->r_flags &= ~BBR_RXT_CLEARED;
5365         }
5366         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) {
5367                 bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
5368                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
5369         }
5370         idx = rsm->r_rtr_cnt - 1;
5371         rsm->r_tim_lastsent[idx] = cts;
5372         rsm->r_pacing_delay = pacing_time;
5373         rsm->r_delivered = bbr->r_ctl.rc_delivered;
5374         rsm->r_ts_valid = bbr->rc_ts_valid;
5375         if (bbr->rc_ts_valid)
5376                 rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
5377         if (bbr->r_ctl.r_app_limited_until)
5378                 rsm->r_app_limited = 1;
5379         else
5380                 rsm->r_app_limited = 0;
5381         if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
5382                 rsm->r_bbr_state = bbr_state_val(bbr);
5383         else
5384                 rsm->r_bbr_state = 8;
5385         if (rsm->r_flags & BBR_ACKED) {
5386                 /* Problably MTU discovery messing with us */
5387                 uint32_t old_flags;
5388
5389                 old_flags = rsm->r_flags;
5390                 rsm->r_flags &= ~BBR_ACKED;
5391                 bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__);
5392                 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
5393                 if (bbr->r_ctl.rc_sacked == 0)
5394                         bbr->r_ctl.rc_sacklast = NULL;
5395         }
5396         if (rsm->r_in_tmap) {
5397                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5398         }
5399         TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
5400         rsm->r_in_tmap = 1;
5401         if (rsm->r_flags & BBR_SACK_PASSED) {
5402                 /* We have retransmitted due to the SACK pass */
5403                 rsm->r_flags &= ~BBR_SACK_PASSED;
5404                 rsm->r_flags |= BBR_WAS_SACKPASS;
5405         }
5406         rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
5407         rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
5408                                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
5409         bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
5410         if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
5411                 rsm->r_is_gain = 1;
5412                 rsm->r_is_drain = 0;
5413         } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
5414                 rsm->r_is_drain = 1;
5415                 rsm->r_is_gain = 0;
5416         } else {
5417                 rsm->r_is_drain = 0;
5418                 rsm->r_is_gain = 0;
5419         }
5420         rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */
5421 }
5422
5423 /*
5424  * Returns 0, or the sequence where we stopped
5425  * updating. We also update the lenp to be the amount
5426  * of data left.
5427  */
5428
5429 static uint32_t
5430 bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr,
5431     struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time)
5432 {
5433         /*
5434          * We (re-)transmitted starting at rsm->r_start for some length
5435          * (possibly less than r_end.
5436          */
5437         struct bbr_sendmap *nrsm;
5438         uint32_t c_end;
5439         int32_t len;
5440
5441         len = *lenp;
5442         c_end = rsm->r_start + len;
5443         if (SEQ_GEQ(c_end, rsm->r_end)) {
5444                 /*
5445                  * We retransmitted the whole piece or more than the whole
5446                  * slopping into the next rsm.
5447                  */
5448                 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5449                 if (c_end == rsm->r_end) {
5450                         *lenp = 0;
5451                         return (0);
5452                 } else {
5453                         int32_t act_len;
5454
5455                         /* Hangs over the end return whats left */
5456                         act_len = rsm->r_end - rsm->r_start;
5457                         *lenp = (len - act_len);
5458                         return (rsm->r_end);
5459                 }
5460                 /* We don't get out of this block. */
5461         }
5462         /*
5463          * Here we retransmitted less than the whole thing which means we
5464          * have to split this into what was transmitted and what was not.
5465          */
5466         nrsm = bbr_alloc_full_limit(bbr);
5467         if (nrsm == NULL) {
5468                 *lenp = 0;
5469                 return (0);
5470         }
5471         /*
5472          * So here we are going to take the original rsm and make it what we
5473          * retransmitted. nrsm will be the tail portion we did not
5474          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
5475          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
5476          * 1, 6 and the new piece will be 6, 11.
5477          */
5478         bbr_clone_rsm(bbr, nrsm, rsm, c_end);
5479         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
5480         nrsm->r_dupack = 0;
5481         if (rsm->r_in_tmap) {
5482                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
5483                 nrsm->r_in_tmap = 1;
5484         }
5485         rsm->r_flags &= (~BBR_HAS_FIN);
5486         bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
5487         *lenp = 0;
5488         return (0);
5489 }
5490
5491 static uint64_t
5492 bbr_get_hardware_rate(struct tcp_bbr *bbr)
5493 {
5494         uint64_t bw;
5495
5496         bw = bbr_get_bw(bbr);
5497         bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN];
5498         bw /= (uint64_t)BBR_UNIT;
5499         return(bw);
5500 }
5501
5502 static void
5503 bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts,
5504                        uint64_t act_rate, uint64_t rate_wanted)
5505 {
5506         /*
5507          * We could not get a full gains worth
5508          * of rate.
5509          */
5510         if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) {
5511                 /* we can't even get the real rate */
5512                 uint64_t red;
5513
5514                 bbr->skip_gain = 1;
5515                 bbr->gain_is_limited = 0;
5516                 red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate;
5517                 if (red)
5518                         filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts);
5519         } else {
5520                 /* We can use a lower gain */
5521                 bbr->skip_gain = 0;
5522                 bbr->gain_is_limited = 1;
5523         }
5524 }
5525
5526 static void
5527 bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts)
5528 {
5529         const struct tcp_hwrate_limit_table *nrte;
5530         int error, rate = -1;
5531
5532         if (bbr->r_ctl.crte == NULL)
5533                 return;
5534         if ((bbr->rc_inp->inp_route.ro_nh == NULL) ||
5535             (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) {
5536                 /* Lost our routes? */
5537                 /* Clear the way for a re-attempt */
5538                 bbr->bbr_attempt_hdwr_pace = 0;
5539 lost_rate:
5540                 bbr->gain_is_limited = 0;
5541                 bbr->skip_gain = 0;
5542                 bbr->bbr_hdrw_pacing = 0;
5543                 counter_u64_add(bbr_flows_whdwr_pacing, -1);
5544                 counter_u64_add(bbr_flows_nohdwr_pacing, 1);
5545                 tcp_bbr_tso_size_check(bbr, cts);
5546                 return;
5547         }
5548         rate = bbr_get_hardware_rate(bbr);
5549         nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte,
5550                                    bbr->rc_tp,
5551                                    bbr->rc_inp->inp_route.ro_nh->nh_ifp,
5552                                    rate,
5553                                    (RS_PACING_GEQ|RS_PACING_SUB_OK),
5554                                    &error);
5555         if (nrte == NULL) {
5556                 goto lost_rate;
5557         }
5558         if (nrte != bbr->r_ctl.crte) {
5559                 bbr->r_ctl.crte = nrte;
5560                 if (error == 0)  {
5561                         BBR_STAT_INC(bbr_hdwr_rl_mod_ok);
5562                         if (bbr->r_ctl.crte->rate < rate) {
5563                                 /* We have a problem */
5564                                 bbr_setup_less_of_rate(bbr, cts,
5565                                                        bbr->r_ctl.crte->rate, rate);
5566                         } else {
5567                                 /* We are good */
5568                                 bbr->gain_is_limited = 0;
5569                                 bbr->skip_gain = 0;
5570                         }
5571                 } else {
5572                         /* A failure should release the tag */
5573                         BBR_STAT_INC(bbr_hdwr_rl_mod_fail);
5574                         bbr->gain_is_limited = 0;
5575                         bbr->skip_gain = 0;
5576                         bbr->bbr_hdrw_pacing = 0;
5577                 }
5578                 bbr_type_log_hdwr_pacing(bbr,
5579                                          bbr->r_ctl.crte->ptbl->rs_ifp,
5580                                          rate,
5581                                          ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate),
5582                                          __LINE__,
5583                                          cts,
5584                                          error);
5585         }
5586 }
5587
5588 static void
5589 bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts)
5590 {
5591         /*
5592          * If we have hardware pacing support
5593          * we need to factor that in for our
5594          * TSO size.
5595          */
5596         const struct tcp_hwrate_limit_table *rlp;
5597         uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay;
5598
5599         if ((bbr->bbr_hdrw_pacing == 0) ||
5600             (IN_RECOVERY(bbr->rc_tp->t_flags)) ||
5601             (bbr->r_ctl.crte == NULL))
5602                 return;
5603         if (bbr->hw_pacing_set == 0) {
5604                 /* Not yet by the hdwr pacing count delay */
5605                 return;
5606         }
5607         if (bbr_hdwr_pace_adjust == 0) {
5608                 /* No adjustment */
5609                 return;
5610         }
5611         rlp = bbr->r_ctl.crte;
5612         if (bbr->rc_tp->t_maxseg > bbr->rc_last_options)
5613                 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5614         else
5615                 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5616         /*
5617          * So lets first get the
5618          * time we will take between
5619          * TSO sized sends currently without
5620          * hardware help.
5621          */
5622         cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT,
5623                         bbr->r_ctl.rc_pace_max_segs, cts, 1);
5624         hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg;
5625         hdwr_delay *= rlp->time_between;
5626         if (cur_delay > hdwr_delay)
5627                 delta = cur_delay - hdwr_delay;
5628         else
5629                 delta = 0;
5630         bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay,
5631                              (bbr->r_ctl.rc_pace_max_segs / maxseg),
5632                              1);
5633         if (delta &&
5634             (delta < (max(rlp->time_between,
5635                           bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) {
5636                 /*
5637                  * Now lets divide by the pacing
5638                  * time between each segment the
5639                  * hardware sends rounding up and
5640                  * derive a bytes from that. We multiply
5641                  * that by bbr_hdwr_pace_adjust to get
5642                  * more bang for our buck.
5643                  *
5644                  * The goal is to have the software pacer
5645                  * waiting no more than an additional
5646                  * pacing delay if we can (without the
5647                  * compensation i.e. x bbr_hdwr_pace_adjust).
5648                  */
5649                 seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between),
5650                              (bbr->r_ctl.rc_pace_max_segs/maxseg));
5651                 seg_sz *= bbr_hdwr_pace_adjust;
5652                 if (bbr_hdwr_pace_floor &&
5653                     (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5654                         /* Currently hardware paces
5655                          * out rs_min_seg segments at a time.
5656                          * We need to make sure we always send at least
5657                          * a full burst of bbr_hdwr_pace_floor down.
5658                          */
5659                         seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5660                 }
5661                 seg_sz *= maxseg;
5662         } else if (delta == 0) {
5663                 /*
5664                  * The highest pacing rate is
5665                  * above our b/w gained. This means
5666                  * we probably are going quite fast at
5667                  * the hardware highest rate. Lets just multiply
5668                  * the calculated TSO size by the
5669                  * multiplier factor (its probably
5670                  * 4 segments in the default config for
5671                  * mlx).
5672                  */
5673                 seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust;
5674                 if (bbr_hdwr_pace_floor &&
5675                     (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) {
5676                         /* Currently hardware paces
5677                          * out rs_min_seg segments at a time.
5678                          * We need to make sure we always send at least
5679                          * a full burst of bbr_hdwr_pace_floor down.
5680                          */
5681                         seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg;
5682                 }
5683         } else {
5684                 /*
5685                  * The pacing time difference is so
5686                  * big that the hardware will
5687                  * pace out more rapidly then we
5688                  * really want and then we
5689                  * will have a long delay. Lets just keep
5690                  * the same TSO size so its as if
5691                  * we were not using hdwr pacing (we
5692                  * just gain a bit of spacing from the
5693                  * hardware if seg_sz > 1).
5694                  */
5695                 seg_sz = bbr->r_ctl.rc_pace_max_segs;
5696         }
5697         if (seg_sz > bbr->r_ctl.rc_pace_max_segs)
5698                 new_tso = seg_sz;
5699         else
5700                 new_tso = bbr->r_ctl.rc_pace_max_segs;
5701         if (new_tso >= (PACE_MAX_IP_BYTES-maxseg))
5702                 new_tso = PACE_MAX_IP_BYTES - maxseg;
5703
5704         if (new_tso != bbr->r_ctl.rc_pace_max_segs) {
5705                 bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0);
5706                 bbr->r_ctl.rc_pace_max_segs = new_tso;
5707         }
5708 }
5709
5710 static void
5711 tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts)
5712 {
5713         uint64_t bw;
5714         uint32_t old_tso = 0, new_tso;
5715         uint32_t maxseg, bytes;
5716         uint32_t tls_seg=0;
5717         /*
5718          * Google/linux uses the following algorithm to determine
5719          * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18):
5720          *
5721          *  bytes = bw_in_bytes_per_second / 1000
5722          *  bytes = min(bytes, 64k)
5723          *  tso_segs = bytes / MSS
5724          *  if (bw < 1.2Mbs)
5725          *      min_tso_segs = 1
5726          *  else
5727          *      min_tso_segs = 2
5728          * tso_segs = max(tso_segs, min_tso_segs)
5729          *
5730          * * Note apply a device specific limit (we apply this in the
5731          *   tcp_m_copym).
5732          * Note that before the initial measurement is made google bursts out
5733          * a full iwnd just like new-reno/cubic.
5734          *
5735          * We do not use this algorithm. Instead we
5736          * use a two phased approach:
5737          *
5738          *  if ( bw <= per-tcb-cross-over)
5739          *     goal_tso =  calculate how much with this bw we
5740          *                 can send in goal-time seconds.
5741          *     if (goal_tso > mss)
5742          *         seg = goal_tso / mss
5743          *         tso = seg * mss
5744          *     else
5745          *         tso = mss
5746          *     if (tso > per-tcb-max)
5747          *         tso = per-tcb-max
5748          *  else if ( bw > 512Mbps)
5749          *     tso = max-tso (64k/mss)
5750          *  else
5751          *     goal_tso = bw / per-tcb-divsor
5752          *     seg = (goal_tso + mss-1)/mss
5753          *     tso = seg * mss
5754          *
5755          * if (tso < per-tcb-floor)
5756          *    tso = per-tcb-floor
5757          * if (tso > per-tcb-utter_max)
5758          *    tso = per-tcb-utter_max
5759          *
5760          * Note the default per-tcb-divisor is 1000 (same as google).
5761          * the goal cross over is 30Mbps however. To recreate googles
5762          * algorithm you need to set:
5763          *
5764          * cross-over = 23,168,000 bps
5765          * goal-time = 18000
5766          * per-tcb-max = 2
5767          * per-tcb-divisor = 1000
5768          * per-tcb-floor = 1
5769          *
5770          * This will get you "google bbr" behavior with respect to tso size.
5771          *
5772          * Note we do set anything TSO size until we are past the initial
5773          * window. Before that we gnerally use either a single MSS
5774          * or we use the full IW size (so we burst a IW at a time)
5775          */
5776
5777         if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) {
5778                 maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5779         } else {
5780                 maxseg = BBR_MIN_SEG - bbr->rc_last_options;
5781         }
5782         old_tso = bbr->r_ctl.rc_pace_max_segs;
5783         if (bbr->rc_past_init_win == 0) {
5784                 /*
5785                  * Not enough data has been acknowledged to make a
5786                  * judgement. Set up the initial TSO based on if we
5787                  * are sending a full IW at once or not.
5788                  */
5789                 if (bbr->rc_use_google)
5790                         bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2);
5791                 else if (bbr->bbr_init_win_cheat)
5792                         bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp);
5793                 else
5794                         bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
5795                 if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg)
5796                         bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg;
5797                 if (bbr->r_ctl.rc_pace_max_segs == 0) {
5798                         bbr->r_ctl.rc_pace_max_segs = maxseg;
5799                 }
5800                 bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0);
5801                         bbr_adjust_for_hw_pacing(bbr, cts);
5802                 return;
5803         }
5804         /**
5805          * Now lets set the TSO goal based on our delivery rate in
5806          * bytes per second. Note we only do this if
5807          * we have acked at least the initial cwnd worth of data.
5808          */
5809         bw = bbr_get_bw(bbr);
5810         if (IN_RECOVERY(bbr->rc_tp->t_flags) &&
5811              (bbr->rc_use_google == 0)) {
5812                 /* We clamp to one MSS in recovery */
5813                 new_tso = maxseg;
5814         } else if (bbr->rc_use_google) {
5815                 int min_tso_segs;
5816
5817                 /* Google considers the gain too */
5818                 if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) {
5819                         bw *= bbr->r_ctl.rc_bbr_hptsi_gain;
5820                         bw /= BBR_UNIT;
5821                 }
5822                 bytes = bw / 1024;
5823                 if (bytes > (64 * 1024))
5824                         bytes = 64 * 1024;
5825                 new_tso = bytes / maxseg;
5826                 if (bw < ONE_POINT_TWO_MEG)
5827                         min_tso_segs = 1;
5828                 else
5829                         min_tso_segs = 2;
5830                 if (new_tso < min_tso_segs)
5831                         new_tso = min_tso_segs;
5832                 new_tso *= maxseg;
5833         } else if (bbr->rc_no_pacing) {
5834                 new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg;
5835         } else if (bw <= bbr->r_ctl.bbr_cross_over) {
5836                 /*
5837                  * Calculate the worse case b/w TSO if we are inserting no
5838                  * more than a delay_target number of TSO's.
5839                  */
5840                 uint32_t tso_len, min_tso;
5841
5842                 tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw);
5843                 if (tso_len > maxseg) {
5844                         new_tso = tso_len / maxseg;
5845                         if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max)
5846                                 new_tso = bbr->r_ctl.bbr_hptsi_segments_max;
5847                         new_tso *= maxseg;
5848                 } else {
5849                         /*
5850                          * less than a full sized frame yikes.. long rtt or
5851                          * low bw?
5852                          */
5853                         min_tso = bbr_minseg(bbr);
5854                         if ((tso_len > min_tso) && (bbr_all_get_min == 0))
5855                                 new_tso = rounddown(tso_len, min_tso);
5856                         else
5857                                 new_tso = min_tso;
5858                 }
5859         } else if (bw > FIVETWELVE_MBPS) {
5860                 /*
5861                  * This guy is so fast b/w wise that we can TSO as large as
5862                  * possible of segments that the NIC will allow.
5863                  */
5864                 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5865         } else {
5866                 /*
5867                  * This formula is based on attempting to send a segment or
5868                  * more every bbr_hptsi_per_second. The default is 1000
5869                  * which means you are targeting what you can send every 1ms
5870                  * based on the peers bw.
5871                  *
5872                  * If the number drops to say 500, then you are looking more
5873                  * at 2ms and you will raise how much we send in a single
5874                  * TSO thus saving CPU (less bbr_output_wtime() calls). The
5875                  * trade off of course is you will send more at once and
5876                  * thus tend to clump up the sends into larger "bursts"
5877                  * building a queue.
5878                  */
5879                 bw /= bbr->r_ctl.bbr_hptsi_per_second;
5880                 new_tso = roundup(bw, (uint64_t)maxseg);
5881                 /*
5882                  * Gate the floor to match what our lower than 48Mbps
5883                  * algorithm does. The ceiling (bbr_hptsi_segments_max) thus
5884                  * becomes the floor for this calculation.
5885                  */
5886                 if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg))
5887                         new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg);
5888         }
5889         if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor)))
5890                 new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor;
5891         if (new_tso > PACE_MAX_IP_BYTES)
5892                 new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg);
5893         /* Enforce an utter maximum. */
5894         if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) {
5895                 new_tso = bbr->r_ctl.bbr_utter_max * maxseg;
5896         }
5897         if (old_tso != new_tso) {
5898                 /* Only log changes */
5899                 bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0);
5900                 bbr->r_ctl.rc_pace_max_segs = new_tso;
5901         }
5902         /* We have hardware pacing! */
5903         bbr_adjust_for_hw_pacing(bbr, cts);
5904 }
5905
5906 static void
5907 bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len,
5908     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t cts,
5909     struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc,
5910     struct sockbuf *sb)
5911 {
5912
5913         struct bbr_sendmap *rsm, *nrsm;
5914         register uint32_t snd_max, snd_una;
5915         uint32_t pacing_time;
5916         /*
5917          * Add to the RACK log of packets in flight or retransmitted. If
5918          * there is a TS option we will use the TS echoed, if not we will
5919          * grab a TS.
5920          *
5921          * Retransmissions will increment the count and move the ts to its
5922          * proper place. Note that if options do not include TS's then we
5923          * won't be able to effectively use the ACK for an RTT on a retran.
5924          *
5925          * Notes about r_start and r_end. Lets consider a send starting at
5926          * sequence 1 for 10 bytes. In such an example the r_start would be
5927          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
5928          * This means that r_end is actually the first sequence for the next
5929          * slot (11).
5930          *
5931          */
5932         INP_WLOCK_ASSERT(tp->t_inpcb);
5933         if (err) {
5934                 /*
5935                  * We don't log errors -- we could but snd_max does not
5936                  * advance in this case either.
5937                  */
5938                 return;
5939         }
5940         if (th_flags & TH_RST) {
5941                 /*
5942                  * We don't log resets and we return immediately from
5943                  * sending
5944                  */
5945                 *abandon = 1;
5946                 return;
5947         }
5948         snd_una = tp->snd_una;
5949         if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) {
5950                 /*
5951                  * The call to bbr_log_output is made before bumping
5952                  * snd_max. This means we can record one extra byte on a SYN
5953                  * or FIN if seq_out is adding more on and a FIN is present
5954                  * (and we are not resending).
5955                  */
5956                 if ((th_flags & TH_SYN) && (tp->iss == seq_out))
5957                         len++;
5958                 if (th_flags & TH_FIN)
5959                         len++;
5960         }
5961         if (SEQ_LEQ((seq_out + len), snd_una)) {
5962                 /* Are sending an old segment to induce an ack (keep-alive)? */
5963                 return;
5964         }
5965         if (SEQ_LT(seq_out, snd_una)) {
5966                 /* huh? should we panic? */
5967                 uint32_t end;
5968
5969                 end = seq_out + len;
5970                 seq_out = snd_una;
5971                 len = end - seq_out;
5972         }
5973         snd_max = tp->snd_max;
5974         if (len == 0) {
5975                 /* We don't log zero window probes */
5976                 return;
5977         }
5978         pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1);
5979         /* First question is it a retransmission? */
5980         if (seq_out == snd_max) {
5981 again:
5982                 rsm = bbr_alloc(bbr);
5983                 if (rsm == NULL) {
5984                         return;
5985                 }
5986                 rsm->r_flags = 0;
5987                 if (th_flags & TH_SYN)
5988                         rsm->r_flags |= BBR_HAS_SYN;
5989                 if (th_flags & TH_FIN)
5990                         rsm->r_flags |= BBR_HAS_FIN;
5991                 rsm->r_tim_lastsent[0] = cts;
5992                 rsm->r_rtr_cnt = 1;
5993                 rsm->r_rtr_bytes = 0;
5994                 rsm->r_start = seq_out;
5995                 rsm->r_end = rsm->r_start + len;
5996                 rsm->r_dupack = 0;
5997                 rsm->r_delivered = bbr->r_ctl.rc_delivered;
5998                 rsm->r_pacing_delay = pacing_time;
5999                 rsm->r_ts_valid = bbr->rc_ts_valid;
6000                 if (bbr->rc_ts_valid)
6001                         rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts;
6002                 rsm->r_del_time = bbr->r_ctl.rc_del_time;
6003                 if (bbr->r_ctl.r_app_limited_until)
6004                         rsm->r_app_limited = 1;
6005                 else
6006                         rsm->r_app_limited = 0;
6007                 rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts);
6008                 rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp,
6009                                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
6010                 /*
6011                  * Here we must also add in this rsm since snd_max
6012                  * is updated after we return from a new send.
6013                  */
6014                 rsm->r_flight_at_send += len;
6015                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
6016                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
6017                 rsm->r_in_tmap = 1;
6018                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
6019                         rsm->r_bbr_state = bbr_state_val(bbr);
6020                 else
6021                         rsm->r_bbr_state = 8;
6022                 if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) {
6023                         rsm->r_is_gain = 1;
6024                         rsm->r_is_drain = 0;
6025                 } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) {
6026                         rsm->r_is_drain = 1;
6027                         rsm->r_is_gain = 0;
6028                 } else {
6029                         rsm->r_is_drain = 0;
6030                         rsm->r_is_gain = 0;
6031                 }
6032                 return;
6033         }
6034         /*
6035          * If we reach here its a retransmission and we need to find it.
6036          */
6037 more:
6038         if (hintrsm && (hintrsm->r_start == seq_out)) {
6039                 rsm = hintrsm;
6040                 hintrsm = NULL;
6041         } else if (bbr->r_ctl.rc_next) {
6042                 /* We have a hint from a previous run */
6043                 rsm = bbr->r_ctl.rc_next;
6044         } else {
6045                 /* No hints sorry */
6046                 rsm = NULL;
6047         }
6048         if ((rsm) && (rsm->r_start == seq_out)) {
6049                 /*
6050                  * We used rc_next or hintrsm  to retransmit, hopefully the
6051                  * likely case.
6052                  */
6053                 seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6054                 if (len == 0) {
6055                         return;
6056                 } else {
6057                         goto more;
6058                 }
6059         }
6060         /* Ok it was not the last pointer go through it the hard way. */
6061         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6062                 if (rsm->r_start == seq_out) {
6063                         seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time);
6064                         bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
6065                         if (len == 0) {
6066                                 return;
6067                         } else {
6068                                 continue;
6069                         }
6070                 }
6071                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
6072                         /* Transmitted within this piece */
6073                         /*
6074                          * Ok we must split off the front and then let the
6075                          * update do the rest
6076                          */
6077                         nrsm = bbr_alloc_full_limit(bbr);
6078                         if (nrsm == NULL) {
6079                                 bbr_update_rsm(tp, bbr, rsm, cts, pacing_time);
6080                                 return;
6081                         }
6082                         /*
6083                          * copy rsm to nrsm and then trim the front of rsm
6084                          * to not include this part.
6085                          */
6086                         bbr_clone_rsm(bbr, nrsm, rsm, seq_out);
6087                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
6088                         if (rsm->r_in_tmap) {
6089                                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6090                                 nrsm->r_in_tmap = 1;
6091                         }
6092                         rsm->r_flags &= (~BBR_HAS_FIN);
6093                         seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time);
6094                         if (len == 0) {
6095                                 return;
6096                         }
6097                 }
6098         }
6099         /*
6100          * Hmm not found in map did they retransmit both old and on into the
6101          * new?
6102          */
6103         if (seq_out == tp->snd_max) {
6104                 goto again;
6105         } else if (SEQ_LT(seq_out, tp->snd_max)) {
6106 #ifdef BBR_INVARIANTS
6107                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
6108                     seq_out, len, tp->snd_una, tp->snd_max);
6109                 printf("Starting Dump of all rack entries\n");
6110                 TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
6111                         printf("rsm:%p start:%u end:%u\n",
6112                             rsm, rsm->r_start, rsm->r_end);
6113                 }
6114                 printf("Dump complete\n");
6115                 panic("seq_out not found rack:%p tp:%p",
6116                     bbr, tp);
6117 #endif
6118         } else {
6119 #ifdef BBR_INVARIANTS
6120                 /*
6121                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
6122                  * flag)
6123                  */
6124                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
6125                     seq_out, len, tp->snd_max, tp);
6126 #endif
6127         }
6128 }
6129
6130 static void
6131 bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt)
6132 {
6133         /*
6134          * Collapse timeout back the cum-ack moved.
6135          */
6136         tp->t_rxtshift = 0;
6137         tp->t_softerror = 0;
6138 }
6139
6140 static void
6141 tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin)
6142 {
6143         bbr->rtt_valid = 1;
6144         bbr->r_ctl.cur_rtt = rtt_usecs;
6145         bbr->r_ctl.ts_in = tsin;
6146         if (rsm_send_time)
6147                 bbr->r_ctl.cur_rtt_send_time = rsm_send_time;
6148 }
6149
6150 static void
6151 bbr_make_timestamp_determination(struct tcp_bbr *bbr)
6152 {
6153         /**
6154          * We have in our bbr control:
6155          * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp).
6156          * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts).
6157          * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts)
6158          * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time)
6159          *
6160          * Now we can calculate the time between the sends by doing:
6161          *
6162          * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts
6163          *
6164          * And the peer's time between receiving them by doing:
6165          *
6166          * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp
6167          *
6168          * We want to figure out if the timestamp values are in msec, 10msec or usec.
6169          * We also may find that we can't use the timestamps if say we see
6170          * that the peer_delta indicates that though we may have taken 10ms to
6171          * pace out the data, it only saw 1ms between the two packets. This would
6172          * indicate that somewhere on the path is a batching entity that is giving
6173          * out time-slices of the actual b/w. This would mean we could not use
6174          * reliably the peers timestamps.
6175          *
6176          * We expect delta > peer_delta initially. Until we figure out the
6177          * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio.
6178          * If we place 1000 there then its a ms vs our usec. If we place 10000 there
6179          * then its 10ms vs our usec. If the peer is running a usec clock we would
6180          * put a 1 there. If the value is faster then ours, we will disable the
6181          * use of timestamps (though we could revist this later if we find it to be not
6182          * just an isolated one or two flows)).
6183          *
6184          * To detect the batching middle boxes we will come up with our compensation and
6185          * if with it in place, we find the peer is drastically off (by some margin) in
6186          * the smaller direction, then we will assume the worst case and disable use of timestamps.
6187          *
6188          */
6189         uint64_t delta, peer_delta, delta_up;
6190
6191         delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts;
6192         if (delta < bbr_min_usec_delta) {
6193                 /*
6194                  * Have not seen a min amount of time
6195                  * between our send times so we can
6196                  * make a determination of the timestamp
6197                  * yet.
6198                  */
6199                 return;
6200         }
6201         peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp;
6202         if (peer_delta < bbr_min_peer_delta) {
6203                 /*
6204                  * We may have enough in the form of
6205                  * our delta but the peers number
6206                  * has not changed that much. It could
6207                  * be its clock ratio is such that
6208                  * we need more data (10ms tick) or
6209                  * there may be other compression scenarios
6210                  * going on. In any event we need the
6211                  * spread to be larger.
6212                  */
6213                 return;
6214         }
6215         /* Ok lets first see which way our delta is going */
6216         if (peer_delta > delta) {
6217                 /* Very unlikely, the peer without
6218                  * compensation shows that it saw
6219                  * the two sends arrive further apart
6220                  * then we saw then in micro-seconds.
6221                  */
6222                 if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) {
6223                         /* well it looks like the peer is a micro-second clock. */
6224                         bbr->rc_ts_clock_set = 1;
6225                         bbr->r_ctl.bbr_peer_tsratio = 1;
6226                 } else {
6227                         bbr->rc_ts_cant_be_used = 1;
6228                         bbr->rc_ts_clock_set = 1;
6229                 }
6230                 return;
6231         }
6232         /* Ok we know that the peer_delta is smaller than our send distance */
6233         bbr->rc_ts_clock_set = 1;
6234         /* First question is it within the percentage that they are using usec time? */
6235         delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent;
6236         if ((peer_delta + delta_up) >= delta) {
6237                 /* Its a usec clock */
6238                 bbr->r_ctl.bbr_peer_tsratio = 1;
6239                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6240                 return;
6241         }
6242         /* Ok if not usec, what about 10usec (though unlikely)? */
6243         delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent;
6244         if (((peer_delta * 10) + delta_up) >= delta) {
6245                 bbr->r_ctl.bbr_peer_tsratio = 10;
6246                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6247                 return;
6248         }
6249         /* And what about 100usec (though again unlikely)? */
6250         delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent;
6251         if (((peer_delta * 100) + delta_up) >= delta) {
6252                 bbr->r_ctl.bbr_peer_tsratio = 100;
6253                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6254                 return;
6255         }
6256         /* And how about 1 msec (the most likely one)? */
6257         delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent;
6258         if (((peer_delta * 1000) + delta_up) >= delta) {
6259                 bbr->r_ctl.bbr_peer_tsratio = 1000;
6260                 bbr_log_tstmp_validation(bbr, peer_delta, delta);
6261                 return;
6262         }
6263         /* Ok if not msec could it be 10 msec? */
6264         delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent;
6265         if (((peer_delta * 10000) + delta_up) >= delta) {
6266                 bbr->r_ctl.bbr_peer_tsratio = 10000;
6267                 return;
6268         }
6269         /* If we fall down here the clock tick so slowly we can't use it */
6270         bbr->rc_ts_cant_be_used = 1;
6271         bbr->r_ctl.bbr_peer_tsratio = 0;
6272         bbr_log_tstmp_validation(bbr, peer_delta, delta);
6273 }
6274
6275 /*
6276  * Collect new round-trip time estimate
6277  * and update averages and current timeout.
6278  */
6279 static void
6280 tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts)
6281 {
6282         int32_t delta;
6283         uint32_t rtt, tsin;
6284         int32_t rtt_ticks;
6285
6286         if (bbr->rtt_valid == 0)
6287                 /* No valid sample */
6288                 return;
6289
6290         rtt = bbr->r_ctl.cur_rtt;
6291         tsin = bbr->r_ctl.ts_in;
6292         if (bbr->rc_prtt_set_ts) {
6293                 /*
6294                  * We are to force feed the rttProp filter due
6295                  * to an entry into PROBE_RTT. This assures
6296                  * that the times are sync'd between when we
6297                  * go into PROBE_RTT and the filter expiration.
6298                  *
6299                  * Google does not use a true filter, so they do
6300                  * this implicitly since they only keep one value
6301                  * and when they enter probe-rtt they update the
6302                  * value to the newest rtt.
6303                  */
6304                 uint32_t rtt_prop;
6305
6306                 bbr->rc_prtt_set_ts = 0;
6307                 rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop);
6308                 if (rtt > rtt_prop)
6309                         filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts);
6310                 else
6311                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6312         }
6313         if (bbr->rc_ack_was_delayed)
6314                 rtt += bbr->r_ctl.rc_ack_hdwr_delay;
6315
6316         if (rtt < bbr->r_ctl.rc_lowest_rtt)
6317                 bbr->r_ctl.rc_lowest_rtt = rtt;
6318         bbr_log_rtt_sample(bbr, rtt, tsin);
6319         if (bbr->r_init_rtt) {
6320                 /*
6321                  * The initial rtt is not-trusted, nuke it and lets get
6322                  * our first valid measurement in.
6323                  */
6324                 bbr->r_init_rtt = 0;
6325                 tp->t_srtt = 0;
6326         }
6327         if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) {
6328                 /*
6329                  * So we have not yet figured out
6330                  * what the peers TSTMP value is
6331                  * in (most likely ms). We need a
6332                  * series of cum-ack's to determine
6333                  * this reliably.
6334                  */
6335                 if (bbr->rc_ack_is_cumack) {
6336                         if (bbr->rc_ts_data_set) {
6337                                 /* Lets attempt to determine the timestamp granularity. */
6338                                 bbr_make_timestamp_determination(bbr);
6339                         } else {
6340                                 bbr->rc_ts_data_set = 1;
6341                                 bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts;
6342                                 bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time;
6343                         }
6344                 } else {
6345                         /*
6346                          * We have to have consecutive acks
6347                          * reset any "filled" state to none.
6348                          */
6349                         bbr->rc_ts_data_set = 0;
6350                 }
6351         }
6352         /* Round it up */
6353         rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1)));
6354         if (rtt_ticks == 0)
6355                 rtt_ticks = 1;
6356         if (tp->t_srtt != 0) {
6357                 /*
6358                  * srtt is stored as fixed point with 5 bits after the
6359                  * binary point (i.e., scaled by 8).  The following magic is
6360                  * equivalent to the smoothing algorithm in rfc793 with an
6361                  * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
6362                  * Adjust rtt to origin 0.
6363                  */
6364
6365                 delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT)
6366                     - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
6367
6368                 tp->t_srtt += delta;
6369                 if (tp->t_srtt <= 0)
6370                         tp->t_srtt = 1;
6371
6372                 /*
6373                  * We accumulate a smoothed rtt variance (actually, a
6374                  * smoothed mean difference), then set the retransmit timer
6375                  * to smoothed rtt + 4 times the smoothed variance. rttvar
6376                  * is stored as fixed point with 4 bits after the binary
6377                  * point (scaled by 16).  The following is equivalent to
6378                  * rfc793 smoothing with an alpha of .75 (rttvar =
6379                  * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
6380                  * wired-in beta.
6381                  */
6382                 if (delta < 0)
6383                         delta = -delta;
6384                 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
6385                 tp->t_rttvar += delta;
6386                 if (tp->t_rttvar <= 0)
6387                         tp->t_rttvar = 1;
6388                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
6389                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6390         } else {
6391                 /*
6392                  * No rtt measurement yet - use the unsmoothed rtt. Set the
6393                  * variance to half the rtt (so our first retransmit happens
6394                  * at 3*rtt).
6395                  */
6396                 tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT;
6397                 tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1);
6398                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
6399         }
6400         KMOD_TCPSTAT_INC(tcps_rttupdated);
6401         tp->t_rttupdated++;
6402 #ifdef STATS
6403         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks));
6404 #endif
6405         /*
6406          * the retransmit should happen at rtt + 4 * rttvar. Because of the
6407          * way we do the smoothing, srtt and rttvar will each average +1/2
6408          * tick of bias.  When we compute the retransmit timer, we want 1/2
6409          * tick of rounding and 1 extra tick because of +-1/2 tick
6410          * uncertainty in the firing of the timer.  The bias will give us
6411          * exactly the 1.5 tick we need.  But, because the bias is
6412          * statistical, we have to test that we don't drop below the minimum
6413          * feasible timer (which is 2 ticks).
6414          */
6415         TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
6416             max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2),
6417             MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000));
6418
6419         /*
6420          * We received an ack for a packet that wasn't retransmitted; it is
6421          * probably safe to discard any error indications we've received
6422          * recently.  This isn't quite right, but close enough for now (a
6423          * route might have failed after we sent a segment, and the return
6424          * path might not be symmetrical).
6425          */
6426         tp->t_softerror = 0;
6427         rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT);
6428         if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt)
6429                 bbr->r_ctl.bbr_smallest_srtt_this_state = rtt;
6430 }
6431
6432 static void
6433 bbr_earlier_retran(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm,
6434                    uint32_t t, uint32_t cts, int ack_type)
6435 {
6436         /*
6437          * For this RSM, we acknowledged the data from a previous
6438          * transmission, not the last one we made. This means we did a false
6439          * retransmit.
6440          */
6441         if (rsm->r_flags & BBR_HAS_FIN) {
6442                 /*
6443                  * The sending of the FIN often is multiple sent when we
6444                  * have everything outstanding ack'd. We ignore this case
6445                  * since its over now.
6446                  */
6447                 return;
6448         }
6449         if (rsm->r_flags & BBR_TLP) {
6450                 /*
6451                  * We expect TLP's to have this occur often
6452                  */
6453                 bbr->rc_tlp_rtx_out = 0;
6454                 return;
6455         }
6456         if (ack_type != BBR_CUM_ACKED) {
6457                 /*
6458                  * If it was not a cum-ack we
6459                  * don't really know for sure since
6460                  * the timestamp could be from some
6461                  * other transmission.
6462                  */
6463                 return;
6464         }
6465
6466         if (rsm->r_flags & BBR_WAS_SACKPASS) {
6467                 /*
6468                  * We retransmitted based on a sack and the earlier
6469                  * retransmission ack'd it - re-ordering is occuring.
6470                  */
6471                 BBR_STAT_INC(bbr_reorder_seen);
6472                 bbr->r_ctl.rc_reorder_ts = cts;
6473         }
6474         /* Back down the loss count */
6475         if (rsm->r_flags & BBR_MARKED_LOST) {
6476                 bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
6477                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
6478                 rsm->r_flags &= ~BBR_MARKED_LOST;
6479                 if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
6480                         /* LT sampling also needs adjustment */
6481                         bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
6482         }
6483         /***** RRS HERE ************************/
6484         /* Do we need to do this???            */
6485         /* bbr_reset_lt_bw_sampling(bbr, cts); */
6486         /***** RRS HERE ************************/
6487         BBR_STAT_INC(bbr_badfr);
6488         BBR_STAT_ADD(bbr_badfr_bytes, (rsm->r_end - rsm->r_start));
6489 }
6490
6491 static void
6492 bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line)
6493 {
6494         bbr->r_ctl.rc_rtt_shrinks = cts;
6495         if (bbr_can_force_probertt &&
6496             (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
6497             ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
6498                 /*
6499                  * We should enter probe-rtt its been too long
6500                  * since we have been there.
6501                  */
6502                 bbr_enter_probe_rtt(bbr, cts, __LINE__);
6503         } else
6504                 bbr_check_probe_rtt_limits(bbr, cts);
6505 }
6506
6507 static void
6508 tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts)
6509 {
6510         uint64_t orig_bw;
6511
6512         if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) {
6513                 /* We never apply a zero measurment */
6514                 bbr_log_type_bbrupd(bbr, 20, cts, 0, 0,
6515                                     0, 0, 0, 0, 0, 0);
6516                 return;
6517         }
6518         if (bbr->r_ctl.r_measurement_count < 0xffffffff)
6519                 bbr->r_ctl.r_measurement_count++;
6520         orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate);
6521         apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch);
6522         bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw,
6523                             (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate),
6524                             0, 0, 0, 0, 0, 0);
6525         if (orig_bw &&
6526             (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) {
6527                 if (bbr->bbr_hdrw_pacing) {
6528                         /*
6529                          * Apply a new rate to the hardware
6530                          * possibly.
6531                          */
6532                         bbr_update_hardware_pacing_rate(bbr, cts);
6533                 }
6534                 bbr_set_state_target(bbr, __LINE__);
6535                 tcp_bbr_tso_size_check(bbr, cts);
6536                 if (bbr->r_recovery_bw)  {
6537                         bbr_setup_red_bw(bbr, cts);
6538                         bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW);
6539                 }
6540         } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate))
6541                 tcp_bbr_tso_size_check(bbr, cts);
6542 }
6543
6544 static void
6545 bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6546 {
6547         if (bbr->rc_in_persist == 0) {
6548                 /* We log only when not in persist */
6549                 /* Translate to a Bytes Per Second */
6550                 uint64_t tim, bw, ts_diff, ts_bw;
6551                 uint32_t upper, lower, delivered;
6552
6553                 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6554                         tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6555                 else
6556                         tim = 1;
6557                 /*
6558                  * Now that we have processed the tim (skipping the sample
6559                  * or possibly updating the time, go ahead and
6560                  * calculate the cdr.
6561                  */
6562                 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6563                 bw = (uint64_t)delivered;
6564                 bw *= (uint64_t)USECS_IN_SECOND;
6565                 bw /= tim;
6566                 if (bw == 0) {
6567                         /* We must have a calculatable amount */
6568                         return;
6569                 }
6570                 upper = (bw >> 32) & 0x00000000ffffffff;
6571                 lower = bw & 0x00000000ffffffff;
6572                 /*
6573                  * If we are using this b/w shove it in now so we
6574                  * can see in the trace viewer if it gets over-ridden.
6575                  */
6576                 if (rsm->r_ts_valid &&
6577                     bbr->rc_ts_valid &&
6578                     bbr->rc_ts_clock_set &&
6579                     (bbr->rc_ts_cant_be_used == 0) &&
6580                     bbr->rc_use_ts_limit) {
6581                         ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1);
6582                         ts_diff *= bbr->r_ctl.bbr_peer_tsratio;
6583                         if ((delivered == 0) ||
6584                             (rtt < 1000)) {
6585                                 /* Can't use the ts */
6586                                 bbr_log_type_bbrupd(bbr, 61, cts,
6587                                                     ts_diff,
6588                                                     bbr->r_ctl.last_inbound_ts,
6589                                                     rsm->r_del_ack_ts, 0,
6590                                                     0, 0, 0, delivered);
6591                         } else {
6592                                 ts_bw = (uint64_t)delivered;
6593                                 ts_bw *= (uint64_t)USECS_IN_SECOND;
6594                                 ts_bw /= ts_diff;
6595                                 bbr_log_type_bbrupd(bbr, 62, cts,
6596                                                     (ts_bw >> 32),
6597                                                     (ts_bw & 0xffffffff), 0, 0,
6598                                                     0, 0, ts_diff, delivered);
6599                                 if ((bbr->ts_can_raise) &&
6600                                     (ts_bw > bw)) {
6601                                         bbr_log_type_bbrupd(bbr, 8, cts,
6602                                                             delivered,
6603                                                             ts_diff,
6604                                                             (bw >> 32),
6605                                                             (bw & 0x00000000ffffffff),
6606                                                             0, 0, 0, 0);
6607                                         bw = ts_bw;
6608                                 } else if (ts_bw && (ts_bw < bw)) {
6609                                         bbr_log_type_bbrupd(bbr, 7, cts,
6610                                                             delivered,
6611                                                             ts_diff,
6612                                                             (bw >> 32),
6613                                                             (bw & 0x00000000ffffffff),
6614                                                             0, 0, 0, 0);
6615                                         bw = ts_bw;
6616                                 }
6617                         }
6618                 }
6619                 if (rsm->r_first_sent_time &&
6620                     TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6621                         uint64_t sbw, sti;
6622                         /*
6623                          * We use what was in flight at the time of our
6624                          * send  and the size of this send to figure
6625                          * out what we have been sending at (amount).
6626                          * For the time we take from the time of
6627                          * the send of the first send outstanding
6628                          * until this send plus this sends pacing
6629                          * time. This gives us a good calculation
6630                          * as to the rate we have been sending at.
6631                          */
6632
6633                         sbw = (uint64_t)(rsm->r_flight_at_send);
6634                         sbw *= (uint64_t)USECS_IN_SECOND;
6635                         sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6636                         sti += rsm->r_pacing_delay;
6637                         sbw /= sti;
6638                         if (sbw < bw) {
6639                                 bbr_log_type_bbrupd(bbr, 6, cts,
6640                                                     delivered,
6641                                                     (uint32_t)sti,
6642                                                     (bw >> 32),
6643                                                     (uint32_t)bw,
6644                                                     rsm->r_first_sent_time, 0, (sbw >> 32),
6645                                                     (uint32_t)sbw);
6646                                 bw = sbw;
6647                         }
6648                 }
6649                 /* Use the google algorithm for b/w measurements */
6650                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6651                 if ((rsm->r_app_limited == 0) ||
6652                     (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) {
6653                         tcp_bbr_commit_bw(bbr, cts);
6654                         bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6655                                             0, 0, 0, 0,  bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6656                 }
6657         }
6658 }
6659
6660 static void
6661 bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts)
6662 {
6663         if (bbr->rc_in_persist == 0) {
6664                 /* We log only when not in persist */
6665                 /* Translate to a Bytes Per Second */
6666                 uint64_t tim, bw;
6667                 uint32_t upper, lower, delivered;
6668                 int no_apply = 0;
6669
6670                 if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time))
6671                         tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time);
6672                 else
6673                         tim = 1;
6674                 /*
6675                  * Now that we have processed the tim (skipping the sample
6676                  * or possibly updating the time, go ahead and
6677                  * calculate the cdr.
6678                  */
6679                 delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered);
6680                 bw = (uint64_t)delivered;
6681                 bw *= (uint64_t)USECS_IN_SECOND;
6682                 bw /= tim;
6683                 if (tim < bbr->r_ctl.rc_lowest_rtt) {
6684                         bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6685                                             tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6686
6687                         no_apply = 1;
6688                 }
6689                 upper = (bw >> 32) & 0x00000000ffffffff;
6690                 lower = bw & 0x00000000ffffffff;
6691                 /*
6692                  * If we are using this b/w shove it in now so we
6693                  * can see in the trace viewer if it gets over-ridden.
6694                  */
6695                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6696                 /* Gate by the sending rate */
6697                 if (rsm->r_first_sent_time &&
6698                     TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) {
6699                         uint64_t sbw, sti;
6700                         /*
6701                          * We use what was in flight at the time of our
6702                          * send  and the size of this send to figure
6703                          * out what we have been sending at (amount).
6704                          * For the time we take from the time of
6705                          * the send of the first send outstanding
6706                          * until this send plus this sends pacing
6707                          * time. This gives us a good calculation
6708                          * as to the rate we have been sending at.
6709                          */
6710
6711                         sbw = (uint64_t)(rsm->r_flight_at_send);
6712                         sbw *= (uint64_t)USECS_IN_SECOND;
6713                         sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time;
6714                         sti += rsm->r_pacing_delay;
6715                         sbw /= sti;
6716                         if (sbw < bw) {
6717                                 bbr_log_type_bbrupd(bbr, 6, cts,
6718                                                     delivered,
6719                                                     (uint32_t)sti,
6720                                                     (bw >> 32),
6721                                                     (uint32_t)bw,
6722                                                     rsm->r_first_sent_time, 0, (sbw >> 32),
6723                                                     (uint32_t)sbw);
6724                                 bw = sbw;
6725                         }
6726                         if ((sti > tim) &&
6727                             (sti < bbr->r_ctl.rc_lowest_rtt)) {
6728                                 bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered,
6729                                                     (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0);
6730                                 no_apply = 1;
6731                         } else
6732                                 no_apply = 0;
6733                 }
6734                 bbr->r_ctl.rc_bbr_cur_del_rate = bw;
6735                 if ((no_apply == 0) &&
6736                     ((rsm->r_app_limited == 0) ||
6737                      (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) {
6738                         tcp_bbr_commit_bw(bbr, cts);
6739                         bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered,
6740                                             0, 0, 0, 0, bbr->r_ctl.rc_del_time,  rsm->r_del_time);
6741                 }
6742         }
6743 }
6744
6745 static void
6746 bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin,
6747     uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to)
6748 {
6749         uint64_t old_rttprop;
6750
6751         /* Update our delivery time and amount */
6752         bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start);
6753         bbr->r_ctl.rc_del_time = cts;
6754         if (rtt == 0) {
6755                 /*
6756                  * 0 means its a retransmit, for now we don't use these for
6757                  * the rest of BBR.
6758                  */
6759                 return;
6760         }
6761         if ((bbr->rc_use_google == 0) &&
6762             (match != BBR_RTT_BY_EXACTMATCH) &&
6763             (match != BBR_RTT_BY_TIMESTAMP)){
6764                 /*
6765                  * We get a lot of rtt updates, lets not pay attention to
6766                  * any that are not an exact match. That way we don't have
6767                  * to worry about timestamps and the whole nonsense of
6768                  * unsure if its a retransmission etc (if we ever had the
6769                  * timestamp fixed to always have the last thing sent this
6770                  * would not be a issue).
6771                  */
6772                 return;
6773         }
6774         if ((bbr_no_retran && bbr->rc_use_google) &&
6775             (match != BBR_RTT_BY_EXACTMATCH) &&
6776             (match != BBR_RTT_BY_TIMESTAMP)){
6777                 /*
6778                  * We only do measurements in google mode
6779                  * with bbr_no_retran on for sure things.
6780                  */
6781                 return;
6782         }
6783         /* Only update srtt if we know by exact match */
6784         tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin);
6785         if (ack_type == BBR_CUM_ACKED)
6786                 bbr->rc_ack_is_cumack = 1;
6787         else
6788                 bbr->rc_ack_is_cumack = 0;
6789         old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP);
6790         /*
6791          * Note the following code differs to the original
6792          * BBR spec. It calls for <= not <. However after a
6793          * long discussion in email with Neal, he acknowledged
6794          * that it should be < than so that we will have flows
6795          * going into probe-rtt (we were seeing cases where that
6796          * did not happen and caused ugly things to occur). We
6797          * have added this agreed upon fix to our code base.
6798          */
6799         if (rtt < old_rttprop) {
6800                 /* Update when we last saw a rtt drop */
6801                 bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0);
6802                 bbr_set_reduced_rtt(bbr, cts, __LINE__);
6803         }
6804         bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts,
6805             match, rsm->r_start, rsm->r_flags);
6806         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
6807         if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) {
6808                 /*
6809                  * The RTT-prop moved, reset the target (may be a
6810                  * nop for some states).
6811                  */
6812                 bbr_set_state_target(bbr, __LINE__);
6813                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)
6814                         bbr_log_rtt_shrinks(bbr, cts, 0, 0,
6815                                             __LINE__, BBR_RTTS_NEW_TARGET, 0);
6816                 else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP))
6817                         /* It went up */
6818                         bbr_check_probe_rtt_limits(bbr, cts);
6819         }
6820         if ((bbr->rc_use_google == 0) &&
6821             (match == BBR_RTT_BY_TIMESTAMP)) {
6822                 /*
6823                  * We don't do b/w update with
6824                  * these since they are not really
6825                  * reliable.
6826                  */
6827                 return;
6828         }
6829         if (bbr->r_ctl.r_app_limited_until &&
6830             (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) {
6831                 /* We are no longer app-limited */
6832                 bbr->r_ctl.r_app_limited_until = 0;
6833         }
6834         if (bbr->rc_use_google) {
6835                 bbr_google_measurement(bbr, rsm, rtt, cts);
6836         } else {
6837                 bbr_nf_measurement(bbr, rsm, rtt, cts);
6838         }
6839 }
6840
6841 /*
6842  * Convert a timestamp that the main stack
6843  * uses (milliseconds) into one that bbr uses
6844  * (microseconds). Return that converted timestamp.
6845  */
6846 static uint32_t
6847 bbr_ts_convert(uint32_t cts) {
6848         uint32_t sec, msec;
6849
6850         sec = cts / MS_IN_USEC;
6851         msec = cts - (MS_IN_USEC * sec);
6852         return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC));
6853 }
6854
6855 /*
6856  * Return 0 if we did not update the RTT time, return
6857  * 1 if we did.
6858  */
6859 static int
6860 bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr,
6861     struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack)
6862 {
6863         int32_t i;
6864         uint32_t t, uts = 0;
6865
6866         if ((rsm->r_flags & BBR_ACKED) ||
6867             (rsm->r_flags & BBR_WAS_RENEGED) ||
6868             (rsm->r_flags & BBR_RXT_CLEARED)) {
6869                 /* Already done */
6870                 return (0);
6871         }
6872         if (rsm->r_rtr_cnt == 1) {
6873                 /*
6874                  * Only one transmit. Hopefully the normal case.
6875                  */
6876                 if (TSTMP_GT(cts, rsm->r_tim_lastsent[0]))
6877                         t = cts - rsm->r_tim_lastsent[0];
6878                 else
6879                         t = 1;
6880                 if ((int)t <= 0)
6881                         t = 1;
6882                 bbr->r_ctl.rc_last_rtt = t;
6883                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6884                                     BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to);
6885                 return (1);
6886         }
6887         /* Convert to usecs */
6888         if ((bbr_can_use_ts_for_rtt == 1) &&
6889             (bbr->rc_use_google == 1) &&
6890             (ack_type == BBR_CUM_ACKED) &&
6891             (to->to_flags & TOF_TS) &&
6892             (to->to_tsecr != 0)) {
6893                 t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr;
6894                 if (t < 1)
6895                         t = 1;
6896                 t *= MS_IN_USEC;
6897                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0,
6898                                     BBR_RTT_BY_TIMESTAMP,
6899                                     rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)],
6900                                     ack_type, to);
6901                 return (1);
6902         }
6903         uts = bbr_ts_convert(to->to_tsecr);
6904         if ((to->to_flags & TOF_TS) &&
6905             (to->to_tsecr != 0) &&
6906             (ack_type == BBR_CUM_ACKED) &&
6907             ((rsm->r_flags & BBR_OVERMAX) == 0)) {
6908                 /*
6909                  * Now which timestamp does it match? In this block the ACK
6910                  * may be coming from a previous transmission.
6911                  */
6912                 uint32_t fudge;
6913
6914                 fudge = BBR_TIMER_FUDGE;
6915                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
6916                         if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) &&
6917                             (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) {
6918                                 if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6919                                         t = cts - rsm->r_tim_lastsent[i];
6920                                 else
6921                                         t = 1;
6922                                 if ((int)t <= 0)
6923                                         t = 1;
6924                                 bbr->r_ctl.rc_last_rtt = t;
6925                                 bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING,
6926                                                     rsm->r_tim_lastsent[i], ack_type, to);
6927                                 if ((i + 1) < rsm->r_rtr_cnt) {
6928                                         /* Likely */
6929                                         bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type);
6930                                 } else if (rsm->r_flags & BBR_TLP) {
6931                                         bbr->rc_tlp_rtx_out = 0;
6932                                 }
6933                                 return (1);
6934                         }
6935                 }
6936                 /* Fall through if we can't find a matching timestamp */
6937         }
6938         /*
6939          * Ok its a SACK block that we retransmitted. or a windows
6940          * machine without timestamps. We can tell nothing from the
6941          * time-stamp since its not there or the time the peer last
6942          * recieved a segment that moved forward its cum-ack point.
6943          *
6944          * Lets look at the last retransmit and see what we can tell
6945          * (with BBR for space we only keep 2 note we have to keep
6946          * at least 2 so the map can not be condensed more).
6947          */
6948         i = rsm->r_rtr_cnt - 1;
6949         if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6950                 t = cts - rsm->r_tim_lastsent[i];
6951         else
6952                 goto not_sure;
6953         if (t < bbr->r_ctl.rc_lowest_rtt) {
6954                 /*
6955                  * We retransmitted and the ack came back in less
6956                  * than the smallest rtt we have observed in the
6957                  * windowed rtt. We most likey did an improper
6958                  * retransmit as outlined in 4.2 Step 3 point 2 in
6959                  * the rack-draft.
6960                  *
6961                  * Use the prior transmission to update all the
6962                  * information as long as there is only one prior
6963                  * transmission.
6964                  */
6965                 if ((rsm->r_flags & BBR_OVERMAX) == 0) {
6966 #ifdef BBR_INVARIANTS
6967                         if (rsm->r_rtr_cnt == 1)
6968                                 panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags);
6969 #endif
6970                         i = rsm->r_rtr_cnt - 2;
6971                         if (TSTMP_GT(cts, rsm->r_tim_lastsent[i]))
6972                                 t = cts - rsm->r_tim_lastsent[i];
6973                         else
6974                                 t = 1;
6975                         bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET,
6976                                             rsm->r_tim_lastsent[i], ack_type, to);
6977                         bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type);
6978                 } else {
6979                         /*
6980                          * Too many prior transmissions, just
6981                          * updated BBR delivered
6982                          */
6983 not_sure:
6984                         bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6985                                             BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
6986                 }
6987         } else {
6988                 /*
6989                  * We retransmitted it and the retransmit did the
6990                  * job.
6991                  */
6992                 if (rsm->r_flags & BBR_TLP)
6993                         bbr->rc_tlp_rtx_out = 0;
6994                 if ((rsm->r_flags & BBR_OVERMAX) == 0)
6995                         bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts,
6996                                             BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to);
6997                 else
6998                         bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts,
6999                                             BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to);
7000                 return (1);
7001         }
7002         return (0);
7003 }
7004
7005 /*
7006  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7007  */
7008 static void
7009 bbr_log_sack_passed(struct tcpcb *tp,
7010     struct tcp_bbr *bbr, struct bbr_sendmap *rsm)
7011 {
7012         struct bbr_sendmap *nrsm;
7013
7014         nrsm = rsm;
7015         TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap,
7016             bbr_head, r_tnext) {
7017                 if (nrsm == rsm) {
7018                         /* Skip orginal segment he is acked */
7019                         continue;
7020                 }
7021                 if (nrsm->r_flags & BBR_ACKED) {
7022                         /* Skip ack'd segments */
7023                         continue;
7024                 }
7025                 if (nrsm->r_flags & BBR_SACK_PASSED) {
7026                         /*
7027                          * We found one that is already marked
7028                          * passed, we have been here before and
7029                          * so all others below this are marked.
7030                          */
7031                         break;
7032                 }
7033                 BBR_STAT_INC(bbr_sack_passed);
7034                 nrsm->r_flags |= BBR_SACK_PASSED;
7035                 if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) &&
7036                     bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) {
7037                         bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start;
7038                         bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start;
7039                         nrsm->r_flags |= BBR_MARKED_LOST;
7040                 }
7041                 nrsm->r_flags &= ~BBR_WAS_SACKPASS;
7042         }
7043 }
7044
7045 /*
7046  * Returns the number of bytes that were
7047  * newly ack'd by sack blocks.
7048  */
7049 static uint32_t
7050 bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack,
7051     struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts)
7052 {
7053         int32_t times = 0;
7054         uint32_t start, end, maxseg, changed = 0;
7055         struct bbr_sendmap *rsm, *nrsm;
7056         int32_t used_ref = 1;
7057         uint8_t went_back = 0, went_fwd = 0;
7058
7059         maxseg = tp->t_maxseg - bbr->rc_last_options;
7060         start = sack->start;
7061         end = sack->end;
7062         rsm = *prsm;
7063         if (rsm == NULL)
7064                 used_ref = 0;
7065
7066         /* Do we locate the block behind where we last were? */
7067         if (rsm && SEQ_LT(start, rsm->r_start)) {
7068                 went_back = 1;
7069                 TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
7070                         if (SEQ_GEQ(start, rsm->r_start) &&
7071                             SEQ_LT(start, rsm->r_end)) {
7072                                 goto do_rest_ofb;
7073                         }
7074                 }
7075         }
7076 start_at_beginning:
7077         went_fwd = 1;
7078         /*
7079          * Ok lets locate the block where this guy is fwd from rsm (if its
7080          * set)
7081          */
7082         TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) {
7083                 if (SEQ_GEQ(start, rsm->r_start) &&
7084                     SEQ_LT(start, rsm->r_end)) {
7085                         break;
7086                 }
7087         }
7088 do_rest_ofb:
7089         if (rsm == NULL) {
7090                 /*
7091                  * This happens when we get duplicate sack blocks with the
7092                  * same end. For example SACK 4: 100 SACK 3: 100 The sort
7093                  * will not change there location so we would just start at
7094                  * the end of the first one and get lost.
7095                  */
7096                 if (tp->t_flags & TF_SENTFIN) {
7097                         /*
7098                          * Check to see if we have not logged the FIN that
7099                          * went out.
7100                          */
7101                         nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7102                         if (nrsm && (nrsm->r_end + 1) == tp->snd_max) {
7103                                 /*
7104                                  * Ok we did not get the FIN logged.
7105                                  */
7106                                 nrsm->r_end++;
7107                                 rsm = nrsm;
7108                                 goto do_rest_ofb;
7109                         }
7110                 }
7111                 if (times == 1) {
7112 #ifdef BBR_INVARIANTS
7113                         panic("tp:%p bbr:%p sack:%p to:%p prsm:%p",
7114                             tp, bbr, sack, to, prsm);
7115 #else
7116                         goto out;
7117 #endif
7118                 }
7119                 times++;
7120                 BBR_STAT_INC(bbr_sack_proc_restart);
7121                 rsm = NULL;
7122                 goto start_at_beginning;
7123         }
7124         /* Ok we have an ACK for some piece of rsm */
7125         if (rsm->r_start != start) {
7126                 /*
7127                  * Need to split this in two pieces the before and after.
7128                  */
7129                 if (bbr_sack_mergable(rsm, start, end))
7130                         nrsm = bbr_alloc_full_limit(bbr);
7131                 else
7132                         nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7133                 if (nrsm == NULL) {
7134                         /* We could not allocate ignore the sack */
7135                         struct sackblk blk;
7136
7137                         blk.start = start;
7138                         blk.end = end;
7139                         sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7140                         goto out;
7141                 }
7142                 bbr_clone_rsm(bbr, nrsm, rsm, start);
7143                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7144                 if (rsm->r_in_tmap) {
7145                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7146                         nrsm->r_in_tmap = 1;
7147                 }
7148                 rsm->r_flags &= (~BBR_HAS_FIN);
7149                 rsm = nrsm;
7150         }
7151         if (SEQ_GEQ(end, rsm->r_end)) {
7152                 /*
7153                  * The end of this block is either beyond this guy or right
7154                  * at this guy.
7155                  */
7156                 if ((rsm->r_flags & BBR_ACKED) == 0) {
7157                         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7158                         changed += (rsm->r_end - rsm->r_start);
7159                         bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7160                         bbr_log_sack_passed(tp, bbr, rsm);
7161                         if (rsm->r_flags & BBR_MARKED_LOST) {
7162                                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7163                         }
7164                         /* Is Reordering occuring? */
7165                         if (rsm->r_flags & BBR_SACK_PASSED) {
7166                                 BBR_STAT_INC(bbr_reorder_seen);
7167                                 bbr->r_ctl.rc_reorder_ts = cts;
7168                                 if (rsm->r_flags & BBR_MARKED_LOST) {
7169                                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7170                                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7171                                                 /* LT sampling also needs adjustment */
7172                                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7173                                 }
7174                         }
7175                         rsm->r_flags |= BBR_ACKED;
7176                         rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7177                         if (rsm->r_in_tmap) {
7178                                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7179                                 rsm->r_in_tmap = 0;
7180                         }
7181                 }
7182                 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7183                 if (end == rsm->r_end) {
7184                         /* This block only - done */
7185                         goto out;
7186                 }
7187                 /* There is more not coverend by this rsm move on */
7188                 start = rsm->r_end;
7189                 nrsm = TAILQ_NEXT(rsm, r_next);
7190                 rsm = nrsm;
7191                 times = 0;
7192                 goto do_rest_ofb;
7193         }
7194         if (rsm->r_flags & BBR_ACKED) {
7195                 /* Been here done that */
7196                 goto out;
7197         }
7198         /* Ok we need to split off this one at the tail */
7199         if (bbr_sack_mergable(rsm, start, end))
7200                 nrsm = bbr_alloc_full_limit(bbr);
7201         else
7202                 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
7203         if (nrsm == NULL) {
7204                 /* failed XXXrrs what can we do but loose the sack info? */
7205                 struct sackblk blk;
7206
7207                 blk.start = start;
7208                 blk.end = end;
7209                 sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk);
7210                 goto out;
7211         }
7212         /* Clone it */
7213         bbr_clone_rsm(bbr, nrsm, rsm, end);
7214         /* The sack block does not cover this guy fully */
7215         rsm->r_flags &= (~BBR_HAS_FIN);
7216         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
7217         if (rsm->r_in_tmap) {
7218                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7219                 nrsm->r_in_tmap = 1;
7220         }
7221         nrsm->r_dupack = 0;
7222         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0);
7223         bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED);
7224         changed += (rsm->r_end - rsm->r_start);
7225         bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
7226         bbr_log_sack_passed(tp, bbr, rsm);
7227         /* Is Reordering occuring? */
7228         if (rsm->r_flags & BBR_MARKED_LOST) {
7229                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7230         }
7231         if (rsm->r_flags & BBR_SACK_PASSED) {
7232                 BBR_STAT_INC(bbr_reorder_seen);
7233                 bbr->r_ctl.rc_reorder_ts = cts;
7234                 if (rsm->r_flags & BBR_MARKED_LOST) {
7235                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7236                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7237                                 /* LT sampling also needs adjustment */
7238                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7239                 }
7240         }
7241         rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST);
7242         rsm->r_flags |= BBR_ACKED;
7243         if (rsm->r_in_tmap) {
7244                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7245                 rsm->r_in_tmap = 0;
7246         }
7247 out:
7248         if (rsm && (rsm->r_flags & BBR_ACKED)) {
7249                 /*
7250                  * Now can we merge this newly acked
7251                  * block with either the previous or
7252                  * next block?
7253                  */
7254                 nrsm = TAILQ_NEXT(rsm, r_next);
7255                 if (nrsm &&
7256                     (nrsm->r_flags & BBR_ACKED)) {
7257                         /* yep this and next can be merged */
7258                         rsm = bbr_merge_rsm(bbr, rsm, nrsm);
7259                 }
7260                 /* Now what about the previous? */
7261                 nrsm = TAILQ_PREV(rsm, bbr_head, r_next);
7262                 if (nrsm &&
7263                     (nrsm->r_flags & BBR_ACKED)) {
7264                         /* yep the previous and this can be merged */
7265                         rsm = bbr_merge_rsm(bbr, nrsm, rsm);
7266                 }
7267         }
7268         if (used_ref == 0) {
7269                 BBR_STAT_INC(bbr_sack_proc_all);
7270         } else {
7271                 BBR_STAT_INC(bbr_sack_proc_short);
7272         }
7273         if (went_fwd && went_back) {
7274                 BBR_STAT_INC(bbr_sack_search_both);
7275         } else if (went_fwd) {
7276                 BBR_STAT_INC(bbr_sack_search_fwd);
7277         } else if (went_back) {
7278                 BBR_STAT_INC(bbr_sack_search_back);
7279         }
7280         /* Save off where the next seq is */
7281         if (rsm)
7282                 bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next);
7283         else
7284                 bbr->r_ctl.rc_sacklast = NULL;
7285         *prsm = rsm;
7286         return (changed);
7287 }
7288
7289 static void inline
7290 bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack)
7291 {
7292         struct bbr_sendmap *tmap;
7293
7294         BBR_STAT_INC(bbr_reneges_seen);
7295         tmap = NULL;
7296         while (rsm && (rsm->r_flags & BBR_ACKED)) {
7297                 /* Its no longer sacked, mark it so */
7298                 uint32_t oflags;
7299                 bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7300 #ifdef BBR_INVARIANTS
7301                 if (rsm->r_in_tmap) {
7302                         panic("bbr:%p rsm:%p flags:0x%x in tmap?",
7303                             bbr, rsm, rsm->r_flags);
7304                 }
7305 #endif
7306                 oflags = rsm->r_flags;
7307                 if (rsm->r_flags & BBR_MARKED_LOST) {
7308                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7309                         bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7310                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7311                                 /* LT sampling also needs adjustment */
7312                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7313                 }
7314                 rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST);
7315                 rsm->r_flags |= BBR_WAS_RENEGED;
7316                 rsm->r_flags |= BBR_RXT_CLEARED;
7317                 bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__);
7318                 /* Rebuild it into our tmap */
7319                 if (tmap == NULL) {
7320                         TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7321                         tmap = rsm;
7322                 } else {
7323                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext);
7324                         tmap = rsm;
7325                 }
7326                 tmap->r_in_tmap = 1;
7327                 /*
7328                  * XXXrrs Delivered? Should we do anything here?
7329                  *
7330                  * Of course we don't on a rxt timeout so maybe its ok that
7331                  * we don't?
7332                  *
7333                  * For now lets not.
7334                  */
7335                 rsm = TAILQ_NEXT(rsm, r_next);
7336         }
7337         /*
7338          * Now lets possibly clear the sack filter so we start recognizing
7339          * sacks that cover this area.
7340          */
7341         sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack);
7342 }
7343
7344 static void
7345 bbr_log_syn(struct tcpcb *tp, struct tcpopt *to)
7346 {
7347         struct tcp_bbr *bbr;
7348         struct bbr_sendmap *rsm;
7349         uint32_t cts;
7350
7351         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7352         cts = bbr->r_ctl.rc_rcvtime;
7353         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7354         if (rsm && (rsm->r_flags & BBR_HAS_SYN)) {
7355                 if ((rsm->r_end - rsm->r_start) <= 1) {
7356                         /* Log out the SYN completely */
7357                         bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7358                         rsm->r_rtr_bytes = 0;
7359                         TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7360                         if (rsm->r_in_tmap) {
7361                                 TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7362                                 rsm->r_in_tmap = 0;
7363                         }
7364                         if (bbr->r_ctl.rc_next == rsm) {
7365                                 /* scoot along the marker */
7366                                 bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7367                         }
7368                         if (to != NULL)
7369                                 bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0);
7370                         bbr_free(bbr, rsm);
7371                 } else {
7372                         /* There is more (Fast open)? strip out SYN. */
7373                         rsm->r_flags &= ~BBR_HAS_SYN;
7374                         rsm->r_start++;
7375                 }
7376         }
7377 }
7378
7379 /*
7380  * Returns the number of bytes that were
7381  * acknowledged by SACK blocks.
7382  */
7383
7384 static uint32_t
7385 bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th,
7386     uint32_t *prev_acked)
7387 {
7388         uint32_t changed, last_seq, entered_recovery = 0;
7389         struct tcp_bbr *bbr;
7390         struct bbr_sendmap *rsm;
7391         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
7392         register uint32_t th_ack;
7393         int32_t i, j, k, new_sb, num_sack_blks = 0;
7394         uint32_t cts, acked, ack_point, sack_changed = 0;
7395         uint32_t p_maxseg, maxseg, p_acked = 0;
7396
7397         INP_WLOCK_ASSERT(tp->t_inpcb);
7398         if (th->th_flags & TH_RST) {
7399                 /* We don't log resets */
7400                 return (0);
7401         }
7402         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7403         cts = bbr->r_ctl.rc_rcvtime;
7404
7405         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7406         changed = 0;
7407         maxseg = tp->t_maxseg - bbr->rc_last_options;
7408         p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg);
7409         th_ack = th->th_ack;
7410         if (SEQ_GT(th_ack, tp->snd_una)) {
7411                 acked = th_ack - tp->snd_una;
7412                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__);
7413                 bbr->rc_tp->t_acktime = ticks;
7414         } else
7415                 acked = 0;
7416         if (SEQ_LEQ(th_ack, tp->snd_una)) {
7417                 /* Only sent here for sack processing */
7418                 goto proc_sack;
7419         }
7420         if (rsm && SEQ_GT(th_ack, rsm->r_start)) {
7421                 changed = th_ack - rsm->r_start;
7422         } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) {
7423                 /*
7424                  * For the SYN incoming case we will not have called
7425                  * tcp_output for the sending of the SYN, so there will be
7426                  * no map. All other cases should probably be a panic.
7427                  */
7428                 if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) {
7429                         /*
7430                          * We have a timestamp that can be used to generate
7431                          * an initial RTT.
7432                          */
7433                         uint32_t ts, now, rtt;
7434
7435                         ts = bbr_ts_convert(to->to_tsecr);
7436                         now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv));
7437                         rtt = now - ts;
7438                         if (rtt < 1)
7439                                 rtt = 1;
7440                         bbr_log_type_bbrrttprop(bbr, rtt,
7441                                                 tp->iss, 0, cts,
7442                                                 BBR_RTT_BY_TIMESTAMP, tp->iss, 0);
7443                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
7444                         changed = 1;
7445                         bbr->r_wanted_output = 1;
7446                         goto out;
7447                 }
7448                 goto proc_sack;
7449         } else if (rsm == NULL) {
7450                 goto out;
7451         }
7452         if (changed) {
7453                 /*
7454                  * The ACK point is advancing to th_ack, we must drop off
7455                  * the packets in the rack log and calculate any eligble
7456                  * RTT's.
7457                  */
7458                 bbr->r_wanted_output = 1;
7459 more:
7460                 if (rsm == NULL) {
7461                         if (tp->t_flags & TF_SENTFIN) {
7462                                 /* if we send a FIN we will not hav a map */
7463                                 goto proc_sack;
7464                         }
7465 #ifdef BBR_INVARIANTS
7466                         panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n",
7467                             tp,
7468                             th, tp->t_state, bbr,
7469                             tp->snd_una, tp->snd_max, changed);
7470 #endif
7471                         goto proc_sack;
7472                 }
7473         }
7474         if (SEQ_LT(th_ack, rsm->r_start)) {
7475                 /* Huh map is missing this */
7476 #ifdef BBR_INVARIANTS
7477                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n",
7478                     rsm->r_start,
7479                     th_ack, tp->t_state,
7480                     bbr->r_state, bbr);
7481                 panic("th-ack is bad bbr:%p tp:%p", bbr, tp);
7482 #endif
7483                 goto proc_sack;
7484         } else if (th_ack == rsm->r_start) {
7485                 /* None here to ack */
7486                 goto proc_sack;
7487         }
7488         /*
7489          * Clear the dup ack counter, it will
7490          * either be freed or if there is some
7491          * remaining we need to start it at zero.
7492          */
7493         rsm->r_dupack = 0;
7494         /* Now do we consume the whole thing? */
7495         if (SEQ_GEQ(th_ack, rsm->r_end)) {
7496                 /* Its all consumed. */
7497                 uint32_t left;
7498
7499                 if (rsm->r_flags & BBR_ACKED) {
7500                         /*
7501                          * It was acked on the scoreboard -- remove it from
7502                          * total
7503                          */
7504                         p_acked += (rsm->r_end - rsm->r_start);
7505                         bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7506                         if (bbr->r_ctl.rc_sacked == 0)
7507                                 bbr->r_ctl.rc_sacklast = NULL;
7508                 } else {
7509                         bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack);
7510                         if (rsm->r_flags & BBR_MARKED_LOST) {
7511                                 bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start;
7512                         }
7513                         if (rsm->r_flags & BBR_SACK_PASSED) {
7514                                 /*
7515                                  * There are acked segments ACKED on the
7516                                  * scoreboard further up. We are seeing
7517                                  * reordering.
7518                                  */
7519                                 BBR_STAT_INC(bbr_reorder_seen);
7520                                 bbr->r_ctl.rc_reorder_ts = cts;
7521                                 if (rsm->r_flags & BBR_MARKED_LOST) {
7522                                         bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start;
7523                                         if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost))
7524                                                 /* LT sampling also needs adjustment */
7525                                                 bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost;
7526                                 }
7527                         }
7528                         rsm->r_flags &= ~BBR_MARKED_LOST;
7529                 }
7530                 bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7531                 rsm->r_rtr_bytes = 0;
7532                 TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
7533                 if (rsm->r_in_tmap) {
7534                         TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
7535                         rsm->r_in_tmap = 0;
7536                 }
7537                 if (bbr->r_ctl.rc_next == rsm) {
7538                         /* scoot along the marker */
7539                         bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7540                 }
7541                 bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7542                 /* Adjust the packet counts */
7543                 left = th_ack - rsm->r_end;
7544                 /* Free back to zone */
7545                 bbr_free(bbr, rsm);
7546                 if (left) {
7547                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7548                         goto more;
7549                 }
7550                 goto proc_sack;
7551         }
7552         if (rsm->r_flags & BBR_ACKED) {
7553                 /*
7554                  * It was acked on the scoreboard -- remove it from total
7555                  * for the part being cum-acked.
7556                  */
7557                 p_acked += (rsm->r_end - rsm->r_start);
7558                 bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
7559                 if (bbr->r_ctl.rc_sacked == 0)
7560                         bbr->r_ctl.rc_sacklast = NULL;
7561         } else {
7562                 /*
7563                  * It was acked up to th_ack point for the first time
7564                  */
7565                 struct bbr_sendmap lrsm;
7566
7567                 memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap));
7568                 lrsm.r_end = th_ack;
7569                 bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack);
7570         }
7571         if ((rsm->r_flags & BBR_MARKED_LOST) &&
7572             ((rsm->r_flags & BBR_ACKED) == 0)) {
7573                 /*
7574                  * It was marked lost and partly ack'd now
7575                  * for the first time. We lower the rc_lost_bytes
7576                  * and still leave it MARKED.
7577                  */
7578                 bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start;
7579         }
7580         bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED);
7581         bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
7582         rsm->r_rtr_bytes = 0;
7583         /* adjust packet count */
7584         rsm->r_start = th_ack;
7585 proc_sack:
7586         /* Check for reneging */
7587         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
7588         if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) {
7589                 /*
7590                  * The peer has moved snd_una up to the edge of this send,
7591                  * i.e. one that it had previously acked. The only way that
7592                  * can be true if the peer threw away data (space issues)
7593                  * that it had previously sacked (else it would have given
7594                  * us snd_una up to (rsm->r_end). We need to undo the acked
7595                  * markings here.
7596                  *
7597                  * Note we have to look to make sure th_ack is our
7598                  * rsm->r_start in case we get an old ack where th_ack is
7599                  * behind snd_una.
7600                  */
7601                 bbr_peer_reneges(bbr, rsm, th->th_ack);
7602         }
7603         if ((to->to_flags & TOF_SACK) == 0) {
7604                 /* We are done nothing left to log */
7605                 goto out;
7606         }
7607         rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next);
7608         if (rsm) {
7609                 last_seq = rsm->r_end;
7610         } else {
7611                 last_seq = tp->snd_max;
7612         }
7613         /* Sack block processing */
7614         if (SEQ_GT(th_ack, tp->snd_una))
7615                 ack_point = th_ack;
7616         else
7617                 ack_point = tp->snd_una;
7618         for (i = 0; i < to->to_nsacks; i++) {
7619                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
7620                     &sack, sizeof(sack));
7621                 sack.start = ntohl(sack.start);
7622                 sack.end = ntohl(sack.end);
7623                 if (SEQ_GT(sack.end, sack.start) &&
7624                     SEQ_GT(sack.start, ack_point) &&
7625                     SEQ_LT(sack.start, tp->snd_max) &&
7626                     SEQ_GT(sack.end, ack_point) &&
7627                     SEQ_LEQ(sack.end, tp->snd_max)) {
7628                         if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) &&
7629                             (SEQ_LT(sack.end, last_seq)) &&
7630                             ((sack.end - sack.start) < (p_maxseg / 8))) {
7631                                 /*
7632                                  * Not the last piece and its smaller than
7633                                  * 1/8th of a p_maxseg. We ignore this.
7634                                  */
7635                                 BBR_STAT_INC(bbr_runt_sacks);
7636                                 continue;
7637                         }
7638                         sack_blocks[num_sack_blks] = sack;
7639                         num_sack_blks++;
7640 #ifdef NETFLIX_STATS
7641                 } else if (SEQ_LEQ(sack.start, th_ack) &&
7642                     SEQ_LEQ(sack.end, th_ack)) {
7643                         /*
7644                          * Its a D-SACK block.
7645                          */
7646                         tcp_record_dsack(sack.start, sack.end);
7647 #endif
7648                 }
7649         }
7650         if (num_sack_blks == 0)
7651                 goto out;
7652         /*
7653          * Sort the SACK blocks so we can update the rack scoreboard with
7654          * just one pass.
7655          */
7656         new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks,
7657                                   num_sack_blks, th->th_ack);
7658         ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks);
7659         BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks);
7660         BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb));
7661         num_sack_blks = new_sb;
7662         if (num_sack_blks < 2) {
7663                 goto do_sack_work;
7664         }
7665         /* Sort the sacks */
7666         for (i = 0; i < num_sack_blks; i++) {
7667                 for (j = i + 1; j < num_sack_blks; j++) {
7668                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
7669                                 sack = sack_blocks[i];
7670                                 sack_blocks[i] = sack_blocks[j];
7671                                 sack_blocks[j] = sack;
7672                         }
7673                 }
7674         }
7675         /*
7676          * Now are any of the sack block ends the same (yes some
7677          * implememtations send these)?
7678          */
7679 again:
7680         if (num_sack_blks > 1) {
7681                 for (i = 0; i < num_sack_blks; i++) {
7682                         for (j = i + 1; j < num_sack_blks; j++) {
7683                                 if (sack_blocks[i].end == sack_blocks[j].end) {
7684                                         /*
7685                                          * Ok these two have the same end we
7686                                          * want the smallest end and then
7687                                          * throw away the larger and start
7688                                          * again.
7689                                          */
7690                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
7691                                                 /*
7692                                                  * The second block covers
7693                                                  * more area use that
7694                                                  */
7695                                                 sack_blocks[i].start = sack_blocks[j].start;
7696                                         }
7697                                         /*
7698                                          * Now collapse out the dup-sack and
7699                                          * lower the count
7700                                          */
7701                                         for (k = (j + 1); k < num_sack_blks; k++) {
7702                                                 sack_blocks[j].start = sack_blocks[k].start;
7703                                                 sack_blocks[j].end = sack_blocks[k].end;
7704                                                 j++;
7705                                         }
7706                                         num_sack_blks--;
7707                                         goto again;
7708                                 }
7709                         }
7710                 }
7711         }
7712 do_sack_work:
7713         rsm = bbr->r_ctl.rc_sacklast;
7714         for (i = 0; i < num_sack_blks; i++) {
7715                 acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts);
7716                 if (acked) {
7717                         bbr->r_wanted_output = 1;
7718                         changed += acked;
7719                         sack_changed += acked;
7720                 }
7721         }
7722 out:
7723         *prev_acked = p_acked;
7724         if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) {
7725                 /*
7726                  * Ok we have a high probability that we need to go in to
7727                  * recovery since we have data sack'd
7728                  */
7729                 struct bbr_sendmap *rsm;
7730
7731                 rsm = bbr_check_recovery_mode(tp, bbr, cts);
7732                 if (rsm) {
7733                         /* Enter recovery */
7734                         entered_recovery = 1;
7735                         bbr->r_wanted_output = 1;
7736                         /*
7737                          * When we enter recovery we need to assure we send
7738                          * one packet.
7739                          */
7740                         if (bbr->r_ctl.rc_resend == NULL) {
7741                                 bbr->r_ctl.rc_resend = rsm;
7742                         }
7743                 }
7744         }
7745         if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) {
7746                 /*
7747                  * See if we need to rack-retransmit anything if so set it
7748                  * up as the thing to resend assuming something else is not
7749                  * already in that position.
7750                  */
7751                 if (bbr->r_ctl.rc_resend == NULL) {
7752                         bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
7753                 }
7754         }
7755         /*
7756          * We return the amount that changed via sack, this is used by the
7757          * ack-received code to augment what was changed between th_ack <->
7758          * snd_una.
7759          */
7760         return (sack_changed);
7761 }
7762
7763 static void
7764 bbr_strike_dupack(struct tcp_bbr *bbr)
7765 {
7766         struct bbr_sendmap *rsm;
7767
7768         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap);
7769         if (rsm && (rsm->r_dupack < 0xff)) {
7770                 rsm->r_dupack++;
7771                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD)
7772                         bbr->r_wanted_output = 1;
7773         }
7774 }
7775
7776 /*
7777  * Return value of 1, we do not need to call bbr_process_data().
7778  * return value of 0, bbr_process_data can be called.
7779  * For ret_val if its 0 the TCB is locked and valid, if its non-zero
7780  * its unlocked and probably unsafe to touch the TCB.
7781  */
7782 static int
7783 bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
7784     struct tcpcb *tp, struct tcpopt *to,
7785     uint32_t tiwin, int32_t tlen,
7786     int32_t * ofia, int32_t thflags, int32_t * ret_val)
7787 {
7788         int32_t ourfinisacked = 0;
7789         int32_t acked_amount;
7790         uint16_t nsegs;
7791         int32_t acked;
7792         uint32_t lost, sack_changed = 0;
7793         struct mbuf *mfree;
7794         struct tcp_bbr *bbr;
7795         uint32_t prev_acked = 0;
7796
7797         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
7798         lost = bbr->r_ctl.rc_lost;
7799         nsegs = max(1, m->m_pkthdr.lro_nsegs);
7800         if (SEQ_GT(th->th_ack, tp->snd_max)) {
7801                 ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
7802                 bbr->r_wanted_output = 1;
7803                 return (1);
7804         }
7805         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
7806                 /* Process the ack */
7807                 if (bbr->rc_in_persist)
7808                         tp->t_rxtshift = 0;
7809                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd))
7810                         bbr_strike_dupack(bbr);
7811                 sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
7812         }
7813         bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost));
7814         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
7815                 /*
7816                  * Old ack, behind the last one rcv'd or a duplicate ack
7817                  * with SACK info.
7818                  */
7819                 if (th->th_ack == tp->snd_una) {
7820                         bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0);
7821                         if (bbr->r_state == TCPS_SYN_SENT) {
7822                                 /*
7823                                  * Special case on where we sent SYN. When
7824                                  * the SYN-ACK is processed in syn_sent
7825                                  * state it bumps the snd_una. This causes
7826                                  * us to hit here even though we did ack 1
7827                                  * byte.
7828                                  *
7829                                  * Go through the nothing left case so we
7830                                  * send data.
7831                                  */
7832                                 goto nothing_left;
7833                         }
7834                 }
7835                 return (0);
7836         }
7837         /*
7838          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
7839          * something we sent.
7840          */
7841         if (tp->t_flags & TF_NEEDSYN) {
7842                 /*
7843                  * T/TCP: Connection was half-synchronized, and our SYN has
7844                  * been ACK'd (so connection is now fully synchronized).  Go
7845                  * to non-starred state, increment snd_una for ACK of SYN,
7846                  * and check if we can do window scaling.
7847                  */
7848                 tp->t_flags &= ~TF_NEEDSYN;
7849                 tp->snd_una++;
7850                 /* Do window scaling? */
7851                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
7852                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
7853                         tp->rcv_scale = tp->request_r_scale;
7854                         /* Send window already scaled. */
7855                 }
7856         }
7857         INP_WLOCK_ASSERT(tp->t_inpcb);
7858
7859         acked = BYTES_THIS_ACK(tp, th);
7860         KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
7861         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
7862
7863         /*
7864          * If we just performed our first retransmit, and the ACK arrives
7865          * within our recovery window, then it was a mistake to do the
7866          * retransmit in the first place.  Recover our original cwnd and
7867          * ssthresh, and proceed to transmit where we left off.
7868          */
7869         if (tp->t_flags & TF_PREVVALID) {
7870                 tp->t_flags &= ~TF_PREVVALID;
7871                 if (tp->t_rxtshift == 1 &&
7872                     (int)(ticks - tp->t_badrxtwin) < 0)
7873                         bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
7874         }
7875         SOCKBUF_LOCK(&so->so_snd);
7876         acked_amount = min(acked, (int)sbavail(&so->so_snd));
7877         tp->snd_wnd -= acked_amount;
7878         mfree = sbcut_locked(&so->so_snd, acked_amount);
7879         SOCKBUF_UNLOCK(&so->so_snd);
7880         tp->t_flags |= TF_WAKESOW;
7881         m_freem(mfree);
7882         if (SEQ_GT(th->th_ack, tp->snd_una)) {
7883                 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
7884         }
7885         tp->snd_una = th->th_ack;
7886         bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost));
7887         if (IN_RECOVERY(tp->t_flags)) {
7888                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
7889                     (SEQ_LT(th->th_ack, tp->snd_max))) {
7890                         tcp_bbr_partialack(tp);
7891                 } else {
7892                         bbr_post_recovery(tp);
7893                 }
7894         }
7895         if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
7896                 tp->snd_recover = tp->snd_una;
7897         }
7898         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
7899                 tp->snd_nxt = tp->snd_max;
7900         }
7901         if (tp->snd_una == tp->snd_max) {
7902                 /* Nothing left outstanding */
7903 nothing_left:
7904                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
7905                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
7906                         bbr->rc_tp->t_acktime = 0;
7907                 if ((sbused(&so->so_snd) == 0) &&
7908                     (tp->t_flags & TF_SENTFIN)) {
7909                         ourfinisacked = 1;
7910                 }
7911                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
7912                 if (bbr->rc_in_persist == 0) {
7913                         bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
7914                 }
7915                 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
7916                 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
7917                 /*
7918                  * We invalidate the last ack here since we
7919                  * don't want to transfer forward the time
7920                  * for our sum's calculations.
7921                  */
7922                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7923                     (sbavail(&so->so_snd) == 0) &&
7924                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
7925                         /*
7926                          * The socket was gone and the peer sent data, time
7927                          * to reset him.
7928                          */
7929                         *ret_val = 1;
7930                         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
7931                         /* tcp_close will kill the inp pre-log the Reset */
7932                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
7933                         tp = tcp_close(tp);
7934                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
7935                         BBR_STAT_INC(bbr_dropped_af_data);
7936                         return (1);
7937                 }
7938                 /* Set need output so persist might get set */
7939                 bbr->r_wanted_output = 1;
7940         }
7941         if (ofia)
7942                 *ofia = ourfinisacked;
7943         return (0);
7944 }
7945
7946 static void
7947 bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
7948 {
7949         if (bbr->rc_in_persist == 0) {
7950                 bbr_timer_cancel(bbr, __LINE__, cts);
7951                 bbr->r_ctl.rc_last_delay_val = 0;
7952                 tp->t_rxtshift = 0;
7953                 bbr->rc_in_persist = 1;
7954                 bbr->r_ctl.rc_went_idle_time = cts;
7955                 /* We should be capped when rw went to 0 but just in case */
7956                 bbr_log_type_pesist(bbr, cts, 0, line, 1);
7957                 /* Time freezes for the state, so do the accounting now */
7958                 if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
7959                         uint32_t time_in;
7960
7961                         time_in = cts - bbr->r_ctl.rc_bbr_state_time;
7962                         if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
7963                                 int32_t idx;
7964
7965                                 idx = bbr_state_val(bbr);
7966                                 counter_u64_add(bbr_state_time[(idx + 5)], time_in);
7967                         } else {
7968                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
7969                         }
7970                 }
7971                 bbr->r_ctl.rc_bbr_state_time = cts;
7972         }
7973 }
7974
7975 static void
7976 bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time)
7977 {
7978         /*
7979          * Note that if idle time does not exceed our
7980          * threshold, we do nothing continuing the state
7981          * transitions we were last walking through.
7982          */
7983         if (idle_time >= bbr_idle_restart_threshold) {
7984                 if (bbr->rc_use_idle_restart) {
7985                         bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT;
7986                         /*
7987                          * Set our target using BBR_UNIT, so
7988                          * we increase at a dramatic rate but
7989                          * we stop when we get the pipe
7990                          * full again for our current b/w estimate.
7991                          */
7992                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
7993                         bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
7994                         bbr_set_state_target(bbr, __LINE__);
7995                         /* Now setup our gains to ramp up */
7996                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
7997                         bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
7998                         bbr_log_type_statechange(bbr, cts, __LINE__);
7999                 } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
8000                         bbr_substate_change(bbr, cts, __LINE__, 1);
8001                 }
8002         }
8003 }
8004
8005 static void
8006 bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line)
8007 {
8008         uint32_t idle_time;
8009
8010         if (bbr->rc_in_persist == 0)
8011                 return;
8012         idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time);
8013         bbr->rc_in_persist = 0;
8014         bbr->rc_hit_state_1 = 0;
8015         bbr->r_ctl.rc_del_time = cts;
8016         /*
8017          * We invalidate the last ack here since we
8018          * don't want to transfer forward the time
8019          * for our sum's calculations.
8020          */
8021         if (bbr->rc_inp->inp_in_hpts) {
8022                 tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
8023                 bbr->rc_timer_first = 0;
8024                 bbr->r_ctl.rc_hpts_flags = 0;
8025                 bbr->r_ctl.rc_last_delay_val = 0;
8026                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
8027                 bbr->r_agg_early_set = 0;
8028                 bbr->r_ctl.rc_agg_early = 0;
8029         }
8030         bbr_log_type_pesist(bbr, cts, idle_time, line, 0);
8031         if (idle_time >= bbr_rtt_probe_time) {
8032                 /*
8033                  * This qualifies as a RTT_PROBE session since we drop the
8034                  * data outstanding to nothing and waited more than
8035                  * bbr_rtt_probe_time.
8036                  */
8037                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0);
8038                 bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts;
8039         }
8040         tp->t_rxtshift = 0;
8041         /*
8042          * If in probeBW and we have persisted more than an RTT lets do
8043          * special handling.
8044          */
8045         /* Force a time based epoch */
8046         bbr_set_epoch(bbr, cts, __LINE__);
8047         /*
8048          * Setup the lost so we don't count anything against the guy
8049          * we have been stuck with during persists.
8050          */
8051         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
8052         /* Time un-freezes for the state */
8053         bbr->r_ctl.rc_bbr_state_time = cts;
8054         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) ||
8055             (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) {
8056                 /*
8057                  * If we are going back to probe-bw
8058                  * or probe_rtt, we may need to possibly
8059                  * do a fast restart.
8060                  */
8061                 bbr_restart_after_idle(bbr, cts, idle_time);
8062         }
8063 }
8064
8065 static void
8066 bbr_collapsed_window(struct tcp_bbr *bbr)
8067 {
8068         /*
8069          * Now we must walk the
8070          * send map and divide the
8071          * ones left stranded. These
8072          * guys can't cause us to abort
8073          * the connection and are really
8074          * "unsent". However if a buggy
8075          * client actually did keep some
8076          * of the data i.e. collapsed the win
8077          * and refused to ack and then opened
8078          * the win and acked that data. We would
8079          * get into an ack war, the simplier
8080          * method then of just pretending we
8081          * did not send those segments something
8082          * won't work.
8083          */
8084         struct bbr_sendmap *rsm, *nrsm;
8085         tcp_seq max_seq;
8086         uint32_t maxseg;
8087         int can_split = 0;
8088         int fnd = 0;
8089
8090         maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options;
8091         max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd;
8092         bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0);
8093         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
8094                 /* Find the first seq past or at maxseq */
8095                 if (rsm->r_flags & BBR_RWND_COLLAPSED)
8096                         rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8097                 if (SEQ_GEQ(max_seq, rsm->r_start) &&
8098                     SEQ_GEQ(rsm->r_end, max_seq)) {
8099                         fnd = 1;
8100                         break;
8101                 }
8102         }
8103         bbr->rc_has_collapsed = 0;
8104         if (!fnd) {
8105                 /* Nothing to do strange */
8106                 return;
8107         }
8108         /*
8109          * Now can we split?
8110          *
8111          * We don't want to split if splitting
8112          * would generate too many small segments
8113          * less we let an attacker fragment our
8114          * send_map and leave us out of memory.
8115          */
8116         if ((max_seq != rsm->r_start) &&
8117             (max_seq != rsm->r_end)){
8118                 /* can we split? */
8119                 int res1, res2;
8120
8121                 res1 = max_seq - rsm->r_start;
8122                 res2 = rsm->r_end - max_seq;
8123                 if ((res1 >= (maxseg/8)) &&
8124                     (res2 >= (maxseg/8))) {
8125                         /* No small pieces here */
8126                         can_split = 1;
8127                 } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) {
8128                         /* We are under the limit */
8129                         can_split = 1;
8130                 }
8131         }
8132         /* Ok do we need to split this rsm? */
8133         if (max_seq == rsm->r_start) {
8134                 /* It's this guy no split required */
8135                 nrsm = rsm;
8136         } else if (max_seq == rsm->r_end) {
8137                 /* It's the next one no split required. */
8138                 nrsm = TAILQ_NEXT(rsm, r_next);
8139                 if (nrsm == NULL) {
8140                         /* Huh? */
8141                         return;
8142                 }
8143         } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) {
8144                 /* yep we need to split it */
8145                 nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT);
8146                 if (nrsm == NULL) {
8147                         /* failed XXXrrs what can we do mark the whole? */
8148                         nrsm = rsm;
8149                         goto no_split;
8150                 }
8151                 /* Clone it */
8152                 bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0);
8153                 bbr_clone_rsm(bbr, nrsm, rsm, max_seq);
8154                 TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next);
8155                 if (rsm->r_in_tmap) {
8156                         TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8157                         nrsm->r_in_tmap = 1;
8158                 }
8159         } else {
8160                 /*
8161                  * Split not allowed just start here just
8162                  * use this guy.
8163                  */
8164                 nrsm = rsm;
8165         }
8166 no_split:
8167         BBR_STAT_INC(bbr_collapsed_win);
8168         /* reuse fnd as a count */
8169         fnd = 0;
8170         TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) {
8171                 nrsm->r_flags |= BBR_RWND_COLLAPSED;
8172                 fnd++;
8173                 bbr->rc_has_collapsed = 1;
8174         }
8175         bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd);
8176 }
8177
8178 static void
8179 bbr_un_collapse_window(struct tcp_bbr *bbr)
8180 {
8181         struct bbr_sendmap *rsm;
8182         int cleared = 0;
8183
8184         TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) {
8185                 if (rsm->r_flags & BBR_RWND_COLLAPSED) {
8186                         /* Clear the flag */
8187                         rsm->r_flags &= ~BBR_RWND_COLLAPSED;
8188                         cleared++;
8189                 } else
8190                         break;
8191         }
8192         bbr_log_type_rwnd_collapse(bbr,
8193                                    (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared);
8194         bbr->rc_has_collapsed = 0;
8195 }
8196
8197 /*
8198  * Return value of 1, the TCB is unlocked and most
8199  * likely gone, return value of 0, the TCB is still
8200  * locked.
8201  */
8202 static int
8203 bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
8204     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
8205     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
8206 {
8207         /*
8208          * Update window information. Don't look at window if no ACK: TAC's
8209          * send garbage on first SYN.
8210          */
8211         uint16_t nsegs;
8212         int32_t tfo_syn;
8213         struct tcp_bbr *bbr;
8214
8215         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8216         INP_WLOCK_ASSERT(tp->t_inpcb);
8217         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8218         if ((thflags & TH_ACK) &&
8219             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
8220             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
8221             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
8222                 /* keep track of pure window updates */
8223                 if (tlen == 0 &&
8224                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
8225                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
8226                 tp->snd_wnd = tiwin;
8227                 tp->snd_wl1 = th->th_seq;
8228                 tp->snd_wl2 = th->th_ack;
8229                 if (tp->snd_wnd > tp->max_sndwnd)
8230                         tp->max_sndwnd = tp->snd_wnd;
8231                 bbr->r_wanted_output = 1;
8232         } else if (thflags & TH_ACK) {
8233                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
8234                         tp->snd_wnd = tiwin;
8235                         tp->snd_wl1 = th->th_seq;
8236                         tp->snd_wl2 = th->th_ack;
8237                 }
8238         }
8239         if (tp->snd_wnd < ctf_outstanding(tp))
8240                 /* The peer collapsed its window on us */
8241                 bbr_collapsed_window(bbr);
8242         else if (bbr->rc_has_collapsed)
8243                 bbr_un_collapse_window(bbr);
8244         /* Was persist timer active and now we have window space? */
8245         if ((bbr->rc_in_persist != 0) &&
8246             (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8247                                 bbr_minseg(bbr)))) {
8248                 /*
8249                  * Make the rate persist at end of persist mode if idle long
8250                  * enough
8251                  */
8252                 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8253
8254                 /* Make sure we output to start the timer */
8255                 bbr->r_wanted_output = 1;
8256         }
8257         /* Do we need to enter persist? */
8258         if ((bbr->rc_in_persist == 0) &&
8259             (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8260             TCPS_HAVEESTABLISHED(tp->t_state) &&
8261             (tp->snd_max == tp->snd_una) &&
8262             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8263             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8264                 /* No send window.. we must enter persist */
8265                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8266         }
8267         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
8268                 m_freem(m);
8269                 return (0);
8270         }
8271         /*
8272          * We don't support urgent data but
8273          * drag along the up just to make sure
8274          * if there is a stack switch no one
8275          * is surprised.
8276          */
8277         tp->rcv_up = tp->rcv_nxt;
8278         INP_WLOCK_ASSERT(tp->t_inpcb);
8279
8280         /*
8281          * Process the segment text, merging it into the TCP sequencing
8282          * queue, and arranging for acknowledgment of receipt if necessary.
8283          * This process logically involves adjusting tp->rcv_wnd as data is
8284          * presented to the user (this happens in tcp_usrreq.c, case
8285          * PRU_RCVD).  If a FIN has already been received on this connection
8286          * then we just ignore the text.
8287          */
8288         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
8289                    IS_FASTOPEN(tp->t_flags));
8290         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
8291             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8292                 tcp_seq save_start = th->th_seq;
8293                 tcp_seq save_rnxt  = tp->rcv_nxt;
8294                 int     save_tlen  = tlen;
8295
8296                 m_adj(m, drop_hdrlen);  /* delayed header drop */
8297                 /*
8298                  * Insert segment which includes th into TCP reassembly
8299                  * queue with control block tp.  Set thflags to whether
8300                  * reassembly now includes a segment with FIN.  This handles
8301                  * the common case inline (segment is the next to be
8302                  * received on an established connection, and the queue is
8303                  * empty), avoiding linkage into and removal from the queue
8304                  * and repetition of various conversions. Set DELACK for
8305                  * segments received in order, but ack immediately when
8306                  * segments are out of order (so fast retransmit can work).
8307                  */
8308                 if (th->th_seq == tp->rcv_nxt &&
8309                     SEGQ_EMPTY(tp) &&
8310                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
8311                     tfo_syn)) {
8312 #ifdef NETFLIX_SB_LIMITS
8313                         u_int mcnt, appended;
8314
8315                         if (so->so_rcv.sb_shlim) {
8316                                 mcnt = m_memcnt(m);
8317                                 appended = 0;
8318                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8319                                     CFO_NOSLEEP, NULL) == false) {
8320                                         counter_u64_add(tcp_sb_shlim_fails, 1);
8321                                         m_freem(m);
8322                                         return (0);
8323                                 }
8324                         }
8325
8326 #endif
8327                         if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) {
8328                                 bbr->bbr_segs_rcvd += max(1, nsegs);
8329                                 tp->t_flags |= TF_DELACK;
8330                                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8331                         } else {
8332                                 bbr->r_wanted_output = 1;
8333                                 tp->t_flags |= TF_ACKNOW;
8334                         }
8335                         tp->rcv_nxt += tlen;
8336                         if (tlen &&
8337                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8338                             (tp->t_fbyte_in == 0)) {
8339                                 tp->t_fbyte_in = ticks;
8340                                 if (tp->t_fbyte_in == 0)
8341                                         tp->t_fbyte_in = 1;
8342                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
8343                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8344                         }
8345                         thflags = th->th_flags & TH_FIN;
8346                         KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8347                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8348                         SOCKBUF_LOCK(&so->so_rcv);
8349                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
8350                                 m_freem(m);
8351                         else
8352 #ifdef NETFLIX_SB_LIMITS
8353                                 appended =
8354 #endif
8355                                         sbappendstream_locked(&so->so_rcv, m, 0);
8356                         SOCKBUF_UNLOCK(&so->so_rcv);
8357                         tp->t_flags |= TF_WAKESOR;
8358 #ifdef NETFLIX_SB_LIMITS
8359                         if (so->so_rcv.sb_shlim && appended != mcnt)
8360                                 counter_fo_release(so->so_rcv.sb_shlim,
8361                                     mcnt - appended);
8362 #endif
8363                 } else {
8364                         /*
8365                          * XXX: Due to the header drop above "th" is
8366                          * theoretically invalid by now.  Fortunately
8367                          * m_adj() doesn't actually frees any mbufs when
8368                          * trimming from the head.
8369                          */
8370                         tcp_seq temp = save_start;
8371                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
8372                         tp->t_flags |= TF_ACKNOW;
8373                 }
8374                 if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0)) {
8375                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
8376                                 /*
8377                                  * DSACK actually handled in the fastpath
8378                                  * above.
8379                                  */
8380                                 tcp_update_sack_list(tp, save_start,
8381                                     save_start + save_tlen);
8382                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
8383                                 if ((tp->rcv_numsacks >= 1) &&
8384                                     (tp->sackblks[0].end == save_start)) {
8385                                         /*
8386                                          * Partial overlap, recorded at todrop
8387                                          * above.
8388                                          */
8389                                         tcp_update_sack_list(tp,
8390                                             tp->sackblks[0].start,
8391                                             tp->sackblks[0].end);
8392                                 } else {
8393                                         tcp_update_dsack_list(tp, save_start,
8394                                             save_start + save_tlen);
8395                                 }
8396                         } else if (tlen >= save_tlen) {
8397                                 /* Update of sackblks. */
8398                                 tcp_update_dsack_list(tp, save_start,
8399                                     save_start + save_tlen);
8400                         } else if (tlen > 0) {
8401                                 tcp_update_dsack_list(tp, save_start,
8402                                     save_start + tlen);
8403                         }
8404                 }
8405         } else {
8406                 m_freem(m);
8407                 thflags &= ~TH_FIN;
8408         }
8409
8410         /*
8411          * If FIN is received ACK the FIN and let the user know that the
8412          * connection is closing.
8413          */
8414         if (thflags & TH_FIN) {
8415                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
8416                         socantrcvmore(so);
8417                         /* The socket upcall is handled by socantrcvmore. */
8418                         tp->t_flags &= ~TF_WAKESOR;
8419                         /*
8420                          * If connection is half-synchronized (ie NEEDSYN
8421                          * flag on) then delay ACK, so it may be piggybacked
8422                          * when SYN is sent. Otherwise, since we received a
8423                          * FIN then no more input can be expected, send ACK
8424                          * now.
8425                          */
8426                         if (tp->t_flags & TF_NEEDSYN) {
8427                                 tp->t_flags |= TF_DELACK;
8428                                 bbr_timer_cancel(bbr,
8429                                     __LINE__, bbr->r_ctl.rc_rcvtime);
8430                         } else {
8431                                 tp->t_flags |= TF_ACKNOW;
8432                         }
8433                         tp->rcv_nxt++;
8434                 }
8435                 switch (tp->t_state) {
8436                         /*
8437                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
8438                          * CLOSE_WAIT state.
8439                          */
8440                 case TCPS_SYN_RECEIVED:
8441                         tp->t_starttime = ticks;
8442                         /* FALLTHROUGH */
8443                 case TCPS_ESTABLISHED:
8444                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
8445                         break;
8446
8447                         /*
8448                          * If still in FIN_WAIT_1 STATE FIN has not been
8449                          * acked so enter the CLOSING state.
8450                          */
8451                 case TCPS_FIN_WAIT_1:
8452                         tcp_state_change(tp, TCPS_CLOSING);
8453                         break;
8454
8455                         /*
8456                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
8457                          * starting the time-wait timer, turning off the
8458                          * other standard timers.
8459                          */
8460                 case TCPS_FIN_WAIT_2:
8461                         bbr->rc_timer_first = 1;
8462                         bbr_timer_cancel(bbr,
8463                             __LINE__, bbr->r_ctl.rc_rcvtime);
8464                         INP_WLOCK_ASSERT(tp->t_inpcb);
8465                         tcp_twstart(tp);
8466                         return (1);
8467                 }
8468         }
8469         /*
8470          * Return any desired output.
8471          */
8472         if ((tp->t_flags & TF_ACKNOW) ||
8473             (sbavail(&so->so_snd) > ctf_outstanding(tp))) {
8474                 bbr->r_wanted_output = 1;
8475         }
8476         INP_WLOCK_ASSERT(tp->t_inpcb);
8477         return (0);
8478 }
8479
8480 /*
8481  * Here nothing is really faster, its just that we
8482  * have broken out the fast-data path also just like
8483  * the fast-ack. Return 1 if we processed the packet
8484  * return 0 if you need to take the "slow-path".
8485  */
8486 static int
8487 bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
8488     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8489     uint32_t tiwin, int32_t nxt_pkt)
8490 {
8491         uint16_t nsegs;
8492         int32_t newsize = 0;    /* automatic sockbuf scaling */
8493         struct tcp_bbr *bbr;
8494 #ifdef NETFLIX_SB_LIMITS
8495         u_int mcnt, appended;
8496 #endif
8497 #ifdef TCPDEBUG
8498         /*
8499          * The size of tcp_saveipgen must be the size of the max ip header,
8500          * now IPv6.
8501          */
8502         u_char tcp_saveipgen[IP6_HDR_LEN];
8503         struct tcphdr tcp_savetcp;
8504         short ostate = 0;
8505
8506 #endif
8507         /* On the hpts and we would have called output */
8508         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8509
8510         /*
8511          * If last ACK falls within this segment's sequence numbers, record
8512          * the timestamp. NOTE that the test is modified according to the
8513          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8514          */
8515         if (bbr->r_ctl.rc_resend != NULL) {
8516                 return (0);
8517         }
8518         if (tiwin && tiwin != tp->snd_wnd) {
8519                 return (0);
8520         }
8521         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
8522                 return (0);
8523         }
8524         if (__predict_false((to->to_flags & TOF_TS) &&
8525             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
8526                 return (0);
8527         }
8528         if (__predict_false((th->th_ack != tp->snd_una))) {
8529                 return (0);
8530         }
8531         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
8532                 return (0);
8533         }
8534         if ((to->to_flags & TOF_TS) != 0 &&
8535             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8536                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
8537                 tp->ts_recent = to->to_tsval;
8538         }
8539         /*
8540          * This is a pure, in-sequence data packet with nothing on the
8541          * reassembly queue and we have enough buffer space to take it.
8542          */
8543         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8544
8545 #ifdef NETFLIX_SB_LIMITS
8546         if (so->so_rcv.sb_shlim) {
8547                 mcnt = m_memcnt(m);
8548                 appended = 0;
8549                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
8550                     CFO_NOSLEEP, NULL) == false) {
8551                         counter_u64_add(tcp_sb_shlim_fails, 1);
8552                         m_freem(m);
8553                         return (1);
8554                 }
8555         }
8556 #endif
8557         /* Clean receiver SACK report if present */
8558         if (tp->rcv_numsacks)
8559                 tcp_clean_sackreport(tp);
8560         KMOD_TCPSTAT_INC(tcps_preddat);
8561         tp->rcv_nxt += tlen;
8562         if (tlen &&
8563             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
8564             (tp->t_fbyte_in == 0)) {
8565                 tp->t_fbyte_in = ticks;
8566                 if (tp->t_fbyte_in == 0)
8567                         tp->t_fbyte_in = 1;
8568                 if (tp->t_fbyte_out && tp->t_fbyte_in)
8569                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
8570         }
8571         /*
8572          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
8573          */
8574         tp->snd_wl1 = th->th_seq;
8575         /*
8576          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
8577          */
8578         tp->rcv_up = tp->rcv_nxt;
8579         KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs);
8580         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
8581 #ifdef TCPDEBUG
8582         if (so->so_options & SO_DEBUG)
8583                 tcp_trace(TA_INPUT, ostate, tp,
8584                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
8585 #endif
8586         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
8587
8588         /* Add data to socket buffer. */
8589         SOCKBUF_LOCK(&so->so_rcv);
8590         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
8591                 m_freem(m);
8592         } else {
8593                 /*
8594                  * Set new socket buffer size. Give up when limit is
8595                  * reached.
8596                  */
8597                 if (newsize)
8598                         if (!sbreserve_locked(&so->so_rcv,
8599                             newsize, so, NULL))
8600                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
8601                 m_adj(m, drop_hdrlen);  /* delayed header drop */
8602
8603 #ifdef NETFLIX_SB_LIMITS
8604                 appended =
8605 #endif
8606                         sbappendstream_locked(&so->so_rcv, m, 0);
8607                 ctf_calc_rwin(so, tp);
8608         }
8609         SOCKBUF_UNLOCK(&so->so_rcv);
8610         tp->t_flags |= TF_WAKESOR;
8611 #ifdef NETFLIX_SB_LIMITS
8612         if (so->so_rcv.sb_shlim && mcnt != appended)
8613                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
8614 #endif
8615         if (DELAY_ACK(tp, bbr, nsegs)) {
8616                 bbr->bbr_segs_rcvd += max(1, nsegs);
8617                 tp->t_flags |= TF_DELACK;
8618                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8619         } else {
8620                 bbr->r_wanted_output = 1;
8621                 tp->t_flags |= TF_ACKNOW;
8622         }
8623         return (1);
8624 }
8625
8626 /*
8627  * This subfunction is used to try to highly optimize the
8628  * fast path. We again allow window updates that are
8629  * in sequence to remain in the fast-path. We also add
8630  * in the __predict's to attempt to help the compiler.
8631  * Note that if we return a 0, then we can *not* process
8632  * it and the caller should push the packet into the
8633  * slow-path. If we return 1, then all is well and
8634  * the packet is fully processed.
8635  */
8636 static int
8637 bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
8638     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8639     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
8640 {
8641         int32_t acked;
8642         uint16_t nsegs;
8643         uint32_t sack_changed;
8644 #ifdef TCPDEBUG
8645         /*
8646          * The size of tcp_saveipgen must be the size of the max ip header,
8647          * now IPv6.
8648          */
8649         u_char tcp_saveipgen[IP6_HDR_LEN];
8650         struct tcphdr tcp_savetcp;
8651         short ostate = 0;
8652
8653 #endif
8654         uint32_t prev_acked = 0;
8655         struct tcp_bbr *bbr;
8656
8657         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
8658                 /* Old ack, behind (or duplicate to) the last one rcv'd */
8659                 return (0);
8660         }
8661         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
8662                 /* Above what we have sent? */
8663                 return (0);
8664         }
8665         if (__predict_false(tiwin == 0)) {
8666                 /* zero window */
8667                 return (0);
8668         }
8669         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
8670                 /* We need a SYN or a FIN, unlikely.. */
8671                 return (0);
8672         }
8673         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
8674                 /* Timestamp is behind .. old ack with seq wrap? */
8675                 return (0);
8676         }
8677         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
8678                 /* Still recovering */
8679                 return (0);
8680         }
8681         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8682         if (__predict_false(bbr->r_ctl.rc_resend != NULL)) {
8683                 /* We are retransmitting */
8684                 return (0);
8685         }
8686         if (__predict_false(bbr->rc_in_persist != 0)) {
8687                 /* In persist mode */
8688                 return (0);
8689         }
8690         if (bbr->r_ctl.rc_sacked) {
8691                 /* We have sack holes on our scoreboard */
8692                 return (0);
8693         }
8694         /* Ok if we reach here, we can process a fast-ack */
8695         nsegs = max(1, m->m_pkthdr.lro_nsegs);
8696         sack_changed = bbr_log_ack(tp, to, th, &prev_acked);
8697         /*
8698          * We never detect loss in fast ack [we can't
8699          * have a sack and can't be in recovery so
8700          * we always pass 0 (nothing detected)].
8701          */
8702         bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0);
8703         /* Did the window get updated? */
8704         if (tiwin != tp->snd_wnd) {
8705                 tp->snd_wnd = tiwin;
8706                 tp->snd_wl1 = th->th_seq;
8707                 if (tp->snd_wnd > tp->max_sndwnd)
8708                         tp->max_sndwnd = tp->snd_wnd;
8709         }
8710         /* Do we need to exit persists? */
8711         if ((bbr->rc_in_persist != 0) &&
8712             (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2),
8713                                bbr_minseg(bbr)))) {
8714                 bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8715                 bbr->r_wanted_output = 1;
8716         }
8717         /* Do we need to enter persists? */
8718         if ((bbr->rc_in_persist == 0) &&
8719             (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
8720             TCPS_HAVEESTABLISHED(tp->t_state) &&
8721             (tp->snd_max == tp->snd_una) &&
8722             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
8723             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
8724                 /* No send window.. we must enter persist */
8725                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
8726         }
8727         /*
8728          * If last ACK falls within this segment's sequence numbers, record
8729          * the timestamp. NOTE that the test is modified according to the
8730          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
8731          */
8732         if ((to->to_flags & TOF_TS) != 0 &&
8733             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
8734                 tp->ts_recent_age = bbr->r_ctl.rc_rcvtime;
8735                 tp->ts_recent = to->to_tsval;
8736         }
8737         /*
8738          * This is a pure ack for outstanding data.
8739          */
8740         KMOD_TCPSTAT_INC(tcps_predack);
8741
8742         /*
8743          * "bad retransmit" recovery.
8744          */
8745         if (tp->t_flags & TF_PREVVALID) {
8746                 tp->t_flags &= ~TF_PREVVALID;
8747                 if (tp->t_rxtshift == 1 &&
8748                     (int)(ticks - tp->t_badrxtwin) < 0)
8749                         bbr_cong_signal(tp, th, CC_RTO_ERR, NULL);
8750         }
8751         /*
8752          * Recalculate the transmit timer / rtt.
8753          *
8754          * Some boxes send broken timestamp replies during the SYN+ACK
8755          * phase, ignore timestamps of 0 or we could calculate a huge RTT
8756          * and blow up the retransmit timer.
8757          */
8758         acked = BYTES_THIS_ACK(tp, th);
8759
8760 #ifdef TCP_HHOOK
8761         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
8762         hhook_run_tcp_est_in(tp, th, to);
8763 #endif
8764
8765         KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs);
8766         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
8767         sbdrop(&so->so_snd, acked);
8768
8769         if (SEQ_GT(th->th_ack, tp->snd_una))
8770                 bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp));
8771         tp->snd_una = th->th_ack;
8772         if (tp->snd_wnd < ctf_outstanding(tp))
8773                 /* The peer collapsed its window on us */
8774                 bbr_collapsed_window(bbr);
8775         else if (bbr->rc_has_collapsed)
8776                 bbr_un_collapse_window(bbr);
8777
8778         if (SEQ_GT(tp->snd_una, tp->snd_recover)) {
8779                 tp->snd_recover = tp->snd_una;
8780         }
8781         bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0);
8782         /*
8783          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
8784          */
8785         tp->snd_wl2 = th->th_ack;
8786         m_freem(m);
8787         /*
8788          * If all outstanding data are acked, stop retransmit timer,
8789          * otherwise restart timer using current (possibly backed-off)
8790          * value. If process is waiting for space, wakeup/selwakeup/signal.
8791          * If data are ready to send, let tcp_output decide between more
8792          * output or persist.
8793          */
8794 #ifdef TCPDEBUG
8795         if (so->so_options & SO_DEBUG)
8796                 tcp_trace(TA_INPUT, ostate, tp,
8797                     (void *)tcp_saveipgen,
8798                     &tcp_savetcp, 0);
8799 #endif
8800         /* Wake up the socket if we have room to write more */
8801         tp->t_flags |= TF_WAKESOW;
8802         if (tp->snd_una == tp->snd_max) {
8803                 /* Nothing left outstanding */
8804                 bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__);
8805                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
8806                         bbr->rc_tp->t_acktime = 0;
8807                 bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8808                 if (bbr->rc_in_persist == 0) {
8809                         bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime;
8810                 }
8811                 sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
8812                 bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime);
8813                 /*
8814                  * We invalidate the last ack here since we
8815                  * don't want to transfer forward the time
8816                  * for our sum's calculations.
8817                  */
8818                 bbr->r_wanted_output = 1;
8819         }
8820         if (sbavail(&so->so_snd)) {
8821                 bbr->r_wanted_output = 1;
8822         }
8823         return (1);
8824 }
8825
8826 /*
8827  * Return value of 1, the TCB is unlocked and most
8828  * likely gone, return value of 0, the TCB is still
8829  * locked.
8830  */
8831 static int
8832 bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
8833     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
8834     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
8835 {
8836         int32_t todrop;
8837         int32_t ourfinisacked = 0;
8838         struct tcp_bbr *bbr;
8839         int32_t ret_val = 0;
8840
8841         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
8842         ctf_calc_rwin(so, tp);
8843         /*
8844          * If the state is SYN_SENT: if seg contains an ACK, but not for our
8845          * SYN, drop the input. if seg contains a RST, then drop the
8846          * connection. if seg does not contain SYN, then drop it. Otherwise
8847          * this is an acceptable SYN segment initialize tp->rcv_nxt and
8848          * tp->irs if seg contains ack then advance tp->snd_una. BRR does
8849          * not support ECN so we will not say we are capable. if SYN has
8850          * been acked change to ESTABLISHED else SYN_RCVD state arrange for
8851          * segment to be acked (eventually) continue processing rest of
8852          * data/controls, beginning with URG
8853          */
8854         if ((thflags & TH_ACK) &&
8855             (SEQ_LEQ(th->th_ack, tp->iss) ||
8856             SEQ_GT(th->th_ack, tp->snd_max))) {
8857                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
8858                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
8859                 return (1);
8860         }
8861         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
8862                 TCP_PROBE5(connect__refused, NULL, tp,
8863                     mtod(m, const char *), tp, th);
8864                 tp = tcp_drop(tp, ECONNREFUSED);
8865                 ctf_do_drop(m, tp);
8866                 return (1);
8867         }
8868         if (thflags & TH_RST) {
8869                 ctf_do_drop(m, tp);
8870                 return (1);
8871         }
8872         if (!(thflags & TH_SYN)) {
8873                 ctf_do_drop(m, tp);
8874                 return (1);
8875         }
8876         tp->irs = th->th_seq;
8877         tcp_rcvseqinit(tp);
8878         if (thflags & TH_ACK) {
8879                 int tfo_partial = 0;
8880
8881                 KMOD_TCPSTAT_INC(tcps_connects);
8882                 soisconnected(so);
8883 #ifdef MAC
8884                 mac_socketpeer_set_from_mbuf(m, so);
8885 #endif
8886                 /* Do window scaling on this connection? */
8887                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
8888                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
8889                         tp->rcv_scale = tp->request_r_scale;
8890                 }
8891                 tp->rcv_adv += min(tp->rcv_wnd,
8892                     TCP_MAXWIN << tp->rcv_scale);
8893                 /*
8894                  * If not all the data that was sent in the TFO SYN
8895                  * has been acked, resend the remainder right away.
8896                  */
8897                 if (IS_FASTOPEN(tp->t_flags) &&
8898                     (tp->snd_una != tp->snd_max)) {
8899                         tp->snd_nxt = th->th_ack;
8900                         tfo_partial = 1;
8901                 }
8902                 /*
8903                  * If there's data, delay ACK; if there's also a FIN ACKNOW
8904                  * will be turned on later.
8905                  */
8906                 if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) {
8907                         bbr->bbr_segs_rcvd += 1;
8908                         tp->t_flags |= TF_DELACK;
8909                         bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime);
8910                 } else {
8911                         bbr->r_wanted_output = 1;
8912                         tp->t_flags |= TF_ACKNOW;
8913                 }
8914                 if (SEQ_GT(th->th_ack, tp->iss)) {
8915                         /*
8916                          * The SYN is acked
8917                          * handle it specially.
8918                          */
8919                         bbr_log_syn(tp, to);
8920                 }
8921                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
8922                         /*
8923                          * We advance snd_una for the
8924                          * fast open case. If th_ack is
8925                          * acknowledging data beyond
8926                          * snd_una we can't just call
8927                          * ack-processing since the
8928                          * data stream in our send-map
8929                          * will start at snd_una + 1 (one
8930                          * beyond the SYN). If its just
8931                          * equal we don't need to do that
8932                          * and there is no send_map.
8933                          */
8934                         tp->snd_una++;
8935                 }
8936                 /*
8937                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
8938                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
8939                  */
8940                 tp->t_starttime = ticks;
8941                 if (tp->t_flags & TF_NEEDFIN) {
8942                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
8943                         tp->t_flags &= ~TF_NEEDFIN;
8944                         thflags &= ~TH_SYN;
8945                 } else {
8946                         tcp_state_change(tp, TCPS_ESTABLISHED);
8947                         TCP_PROBE5(connect__established, NULL, tp,
8948                             mtod(m, const char *), tp, th);
8949                         cc_conn_init(tp);
8950                 }
8951         } else {
8952                 /*
8953                  * Received initial SYN in SYN-SENT[*] state => simultaneous
8954                  * open.  If segment contains CC option and there is a
8955                  * cached CC, apply TAO test. If it succeeds, connection is *
8956                  * half-synchronized. Otherwise, do 3-way handshake:
8957                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
8958                  * there was no CC option, clear cached CC value.
8959                  */
8960                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
8961                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
8962         }
8963         INP_WLOCK_ASSERT(tp->t_inpcb);
8964         /*
8965          * Advance th->th_seq to correspond to first data byte. If data,
8966          * trim to stay within window, dropping FIN if necessary.
8967          */
8968         th->th_seq++;
8969         if (tlen > tp->rcv_wnd) {
8970                 todrop = tlen - tp->rcv_wnd;
8971                 m_adj(m, -todrop);
8972                 tlen = tp->rcv_wnd;
8973                 thflags &= ~TH_FIN;
8974                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
8975                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
8976         }
8977         tp->snd_wl1 = th->th_seq - 1;
8978         tp->rcv_up = th->th_seq;
8979         /*
8980          * Client side of transaction: already sent SYN and data. If the
8981          * remote host used T/TCP to validate the SYN, our data will be
8982          * ACK'd; if so, enter normal data segment processing in the middle
8983          * of step 5, ack processing. Otherwise, goto step 6.
8984          */
8985         if (thflags & TH_ACK) {
8986                 if ((to->to_flags & TOF_TS) != 0) {
8987                         uint32_t t, rtt;
8988
8989                         t = tcp_tv_to_mssectick(&bbr->rc_tv);
8990                         if (TSTMP_GEQ(t, to->to_tsecr)) {
8991                                 rtt = t - to->to_tsecr;
8992                                 if (rtt == 0) {
8993                                         rtt = 1;
8994                                 }
8995                                 rtt *= MS_IN_USEC;
8996                                 tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
8997                                 apply_filter_min_small(&bbr->r_ctl.rc_rttprop,
8998                                                        rtt, bbr->r_ctl.rc_rcvtime);
8999                         }
9000                 }
9001                 if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
9002                         return (ret_val);
9003                 /* We may have changed to FIN_WAIT_1 above */
9004                 if (tp->t_state == TCPS_FIN_WAIT_1) {
9005                         /*
9006                          * In FIN_WAIT_1 STATE in addition to the processing
9007                          * for the ESTABLISHED state if our FIN is now
9008                          * acknowledged then enter FIN_WAIT_2.
9009                          */
9010                         if (ourfinisacked) {
9011                                 /*
9012                                  * If we can't receive any more data, then
9013                                  * closing user can proceed. Starting the
9014                                  * timer is contrary to the specification,
9015                                  * but if we don't get a FIN we'll hang
9016                                  * forever.
9017                                  *
9018                                  * XXXjl: we should release the tp also, and
9019                                  * use a compressed state.
9020                                  */
9021                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9022                                         soisdisconnected(so);
9023                                         tcp_timer_activate(tp, TT_2MSL,
9024                                             (tcp_fast_finwait2_recycle ?
9025                                             tcp_finwait2_timeout :
9026                                             TP_MAXIDLE(tp)));
9027                                 }
9028                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
9029                         }
9030                 }
9031         }
9032         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9033             tiwin, thflags, nxt_pkt));
9034 }
9035
9036 /*
9037  * Return value of 1, the TCB is unlocked and most
9038  * likely gone, return value of 0, the TCB is still
9039  * locked.
9040  */
9041 static int
9042 bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
9043                 struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9044                 uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9045 {
9046         int32_t ourfinisacked = 0;
9047         int32_t ret_val;
9048         struct tcp_bbr *bbr;
9049
9050         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9051         ctf_calc_rwin(so, tp);
9052         if ((thflags & TH_ACK) &&
9053             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
9054              SEQ_GT(th->th_ack, tp->snd_max))) {
9055                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9056                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9057                 return (1);
9058         }
9059         if (IS_FASTOPEN(tp->t_flags)) {
9060                 /*
9061                  * When a TFO connection is in SYN_RECEIVED, the only valid
9062                  * packets are the initial SYN, a retransmit/copy of the
9063                  * initial SYN (possibly with a subset of the original
9064                  * data), a valid ACK, a FIN, or a RST.
9065                  */
9066                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
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                 } else if (thflags & TH_SYN) {
9071                         /* non-initial SYN is ignored */
9072                         if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
9073                             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
9074                             (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
9075                                 ctf_do_drop(m, NULL);
9076                                 return (0);
9077                         }
9078                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
9079                         ctf_do_drop(m, NULL);
9080                         return (0);
9081                 }
9082         }
9083         if ((thflags & TH_RST) ||
9084             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9085                 return (ctf_process_rst(m, th, so, tp));
9086         /*
9087          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9088          * it's less than ts_recent, drop it.
9089          */
9090         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9091             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9092                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9093                         return (ret_val);
9094         }
9095         /*
9096          * In the SYN-RECEIVED state, validate that the packet belongs to
9097          * this connection before trimming the data to fit the receive
9098          * window.  Check the sequence number versus IRS since we know the
9099          * sequence numbers haven't wrapped.  This is a partial fix for the
9100          * "LAND" DoS attack.
9101          */
9102         if (SEQ_LT(th->th_seq, tp->irs)) {
9103                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
9104                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9105                 return (1);
9106         }
9107         INP_WLOCK_ASSERT(tp->t_inpcb);
9108         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9109                 return (ret_val);
9110         }
9111         /*
9112          * If last ACK falls within this segment's sequence numbers, record
9113          * its timestamp. NOTE: 1) That the test incorporates suggestions
9114          * from the latest proposal of the tcplw@cray.com list (Braden
9115          * 1993/04/26). 2) That updating only on newer timestamps interferes
9116          * with our earlier PAWS tests, so this check should be solely
9117          * predicated on the sequence space of this segment. 3) That we
9118          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9119          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9120          * SEG.Len, This modified check allows us to overcome RFC1323's
9121          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9122          * p.869. In such cases, we can still calculate the RTT correctly
9123          * when RCV.NXT == Last.ACK.Sent.
9124          */
9125         if ((to->to_flags & TOF_TS) != 0 &&
9126             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9127             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9128                     ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9129                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9130                 tp->ts_recent = to->to_tsval;
9131         }
9132         tp->snd_wnd = tiwin;
9133         /*
9134          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9135          * is on (half-synchronized state), then queue data for later
9136          * processing; else drop segment and return.
9137          */
9138         if ((thflags & TH_ACK) == 0) {
9139                 if (IS_FASTOPEN(tp->t_flags)) {
9140                         cc_conn_init(tp);
9141                 }
9142                 return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9143                                          tiwin, thflags, nxt_pkt));
9144         }
9145         KMOD_TCPSTAT_INC(tcps_connects);
9146         soisconnected(so);
9147         /* Do window scaling? */
9148         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9149             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9150                 tp->rcv_scale = tp->request_r_scale;
9151         }
9152         /*
9153          * ok for the first time in lets see if we can use the ts to figure
9154          * out what the initial RTT was.
9155          */
9156         if ((to->to_flags & TOF_TS) != 0) {
9157                 uint32_t t, rtt;
9158
9159                 t = tcp_tv_to_mssectick(&bbr->rc_tv);
9160                 if (TSTMP_GEQ(t, to->to_tsecr)) {
9161                         rtt = t - to->to_tsecr;
9162                         if (rtt == 0) {
9163                                 rtt = 1;
9164                         }
9165                         rtt *= MS_IN_USEC;
9166                         tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0);
9167                         apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime);
9168                 }
9169         }
9170         /* Drop off any SYN in the send map (probably not there)  */
9171         if (thflags & TH_ACK)
9172                 bbr_log_syn(tp, to);
9173         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
9174                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
9175                 tp->t_tfo_pending = NULL;
9176         }
9177         /*
9178          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
9179          * FIN-WAIT-1
9180          */
9181         tp->t_starttime = ticks;
9182         if (tp->t_flags & TF_NEEDFIN) {
9183                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
9184                 tp->t_flags &= ~TF_NEEDFIN;
9185         } else {
9186                 tcp_state_change(tp, TCPS_ESTABLISHED);
9187                 TCP_PROBE5(accept__established, NULL, tp,
9188                            mtod(m, const char *), tp, th);
9189                 /*
9190                  * TFO connections call cc_conn_init() during SYN
9191                  * processing.  Calling it again here for such connections
9192                  * is not harmless as it would undo the snd_cwnd reduction
9193                  * that occurs when a TFO SYN|ACK is retransmitted.
9194                  */
9195                 if (!IS_FASTOPEN(tp->t_flags))
9196                         cc_conn_init(tp);
9197         }
9198         /*
9199          * Account for the ACK of our SYN prior to
9200          * regular ACK processing below, except for
9201          * simultaneous SYN, which is handled later.
9202          */
9203         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
9204                 tp->snd_una++;
9205         /*
9206          * If segment contains data or ACK, will call tcp_reass() later; if
9207          * not, do so now to pass queued data to user.
9208          */
9209         if (tlen == 0 && (thflags & TH_FIN) == 0)
9210                 (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
9211                         (struct mbuf *)0);
9212         tp->snd_wl1 = th->th_seq - 1;
9213         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9214                 return (ret_val);
9215         }
9216         if (tp->t_state == TCPS_FIN_WAIT_1) {
9217                 /* We could have went to FIN_WAIT_1 (or EST) above */
9218                 /*
9219                  * In FIN_WAIT_1 STATE in addition to the processing for the
9220                  * ESTABLISHED state if our FIN is now acknowledged then
9221                  * enter FIN_WAIT_2.
9222                  */
9223                 if (ourfinisacked) {
9224                         /*
9225                          * If we can't receive any more data, then closing
9226                          * user can proceed. Starting the timer is contrary
9227                          * to the specification, but if we don't get a FIN
9228                          * we'll hang forever.
9229                          *
9230                          * XXXjl: we should release the tp also, and use a
9231                          * compressed state.
9232                          */
9233                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9234                                 soisdisconnected(so);
9235                                 tcp_timer_activate(tp, TT_2MSL,
9236                                                    (tcp_fast_finwait2_recycle ?
9237                                                     tcp_finwait2_timeout :
9238                                                     TP_MAXIDLE(tp)));
9239                         }
9240                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
9241                 }
9242         }
9243         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9244                                  tiwin, thflags, nxt_pkt));
9245 }
9246
9247 /*
9248  * Return value of 1, the TCB is unlocked and most
9249  * likely gone, return value of 0, the TCB is still
9250  * locked.
9251  */
9252 static int
9253 bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
9254     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9255     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9256 {
9257         struct tcp_bbr *bbr;
9258         int32_t ret_val;
9259
9260         /*
9261          * Header prediction: check for the two common cases of a
9262          * uni-directional data xfer.  If the packet has no control flags,
9263          * is in-sequence, the window didn't change and we're not
9264          * retransmitting, it's a candidate.  If the length is zero and the
9265          * ack moved forward, we're the sender side of the xfer.  Just free
9266          * the data acked & wake any higher level process that was blocked
9267          * waiting for space.  If the length is non-zero and the ack didn't
9268          * move, we're the receiver side.  If we're getting packets in-order
9269          * (the reassembly queue is empty), add the data toc The socket
9270          * buffer and note that we need a delayed ack. Make sure that the
9271          * hidden state-flags are also off. Since we check for
9272          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
9273          */
9274         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9275         if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) {
9276                 /*
9277                  * If we have delived under 4 segments increase the initial
9278                  * window if raised by the peer. We use this to determine
9279                  * dynamic and static rwnd's at the end of a connection.
9280                  */
9281                 bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd);
9282         }
9283         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
9284             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) &&
9285             __predict_true(SEGQ_EMPTY(tp)) &&
9286             __predict_true(th->th_seq == tp->rcv_nxt)) {
9287                 if (tlen == 0) {
9288                         if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
9289                             tiwin, nxt_pkt, iptos)) {
9290                                 return (0);
9291                         }
9292                 } else {
9293                         if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
9294                             tiwin, nxt_pkt)) {
9295                                 return (0);
9296                         }
9297                 }
9298         }
9299         ctf_calc_rwin(so, tp);
9300
9301         if ((thflags & TH_RST) ||
9302             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9303                 return (ctf_process_rst(m, th, so, tp));
9304         /*
9305          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9306          * synchronized state.
9307          */
9308         if (thflags & TH_SYN) {
9309                 ctf_challenge_ack(m, th, tp, &ret_val);
9310                 return (ret_val);
9311         }
9312         /*
9313          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9314          * it's less than ts_recent, drop it.
9315          */
9316         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9317             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9318                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9319                         return (ret_val);
9320         }
9321         INP_WLOCK_ASSERT(tp->t_inpcb);
9322         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9323                 return (ret_val);
9324         }
9325         /*
9326          * If last ACK falls within this segment's sequence numbers, record
9327          * its timestamp. NOTE: 1) That the test incorporates suggestions
9328          * from the latest proposal of the tcplw@cray.com list (Braden
9329          * 1993/04/26). 2) That updating only on newer timestamps interferes
9330          * with our earlier PAWS tests, so this check should be solely
9331          * predicated on the sequence space of this segment. 3) That we
9332          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9333          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9334          * SEG.Len, This modified check allows us to overcome RFC1323's
9335          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9336          * p.869. In such cases, we can still calculate the RTT correctly
9337          * when RCV.NXT == Last.ACK.Sent.
9338          */
9339         if ((to->to_flags & TOF_TS) != 0 &&
9340             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9341             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9342             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9343                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9344                 tp->ts_recent = to->to_tsval;
9345         }
9346         /*
9347          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9348          * is on (half-synchronized state), then queue data for later
9349          * processing; else drop segment and return.
9350          */
9351         if ((thflags & TH_ACK) == 0) {
9352                 if (tp->t_flags & TF_NEEDSYN) {
9353                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9354                             tiwin, thflags, nxt_pkt));
9355                 } else if (tp->t_flags & TF_ACKNOW) {
9356                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9357                         bbr->r_wanted_output = 1;
9358                         return (ret_val);
9359                 } else {
9360                         ctf_do_drop(m, NULL);
9361                         return (0);
9362                 }
9363         }
9364         /*
9365          * Ack processing.
9366          */
9367         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9368                 return (ret_val);
9369         }
9370         if (sbavail(&so->so_snd)) {
9371                 if (ctf_progress_timeout_check(tp, true)) {
9372                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9373                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9374                         return (1);
9375                 }
9376         }
9377         /* State changes only happen in bbr_process_data() */
9378         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9379             tiwin, thflags, nxt_pkt));
9380 }
9381
9382 /*
9383  * Return value of 1, the TCB is unlocked and most
9384  * likely gone, return value of 0, the TCB is still
9385  * locked.
9386  */
9387 static int
9388 bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
9389     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9390     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9391 {
9392         struct tcp_bbr *bbr;
9393         int32_t ret_val;
9394
9395         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9396         ctf_calc_rwin(so, tp);
9397         if ((thflags & TH_RST) ||
9398             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9399                 return (ctf_process_rst(m, th, so, tp));
9400         /*
9401          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9402          * synchronized state.
9403          */
9404         if (thflags & TH_SYN) {
9405                 ctf_challenge_ack(m, th, tp, &ret_val);
9406                 return (ret_val);
9407         }
9408         /*
9409          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9410          * it's less than ts_recent, drop it.
9411          */
9412         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9413             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9414                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9415                         return (ret_val);
9416         }
9417         INP_WLOCK_ASSERT(tp->t_inpcb);
9418         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9419                 return (ret_val);
9420         }
9421         /*
9422          * If last ACK falls within this segment's sequence numbers, record
9423          * its timestamp. NOTE: 1) That the test incorporates suggestions
9424          * from the latest proposal of the tcplw@cray.com list (Braden
9425          * 1993/04/26). 2) That updating only on newer timestamps interferes
9426          * with our earlier PAWS tests, so this check should be solely
9427          * predicated on the sequence space of this segment. 3) That we
9428          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9429          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9430          * SEG.Len, This modified check allows us to overcome RFC1323's
9431          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9432          * p.869. In such cases, we can still calculate the RTT correctly
9433          * when RCV.NXT == Last.ACK.Sent.
9434          */
9435         if ((to->to_flags & TOF_TS) != 0 &&
9436             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9437             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9438             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9439                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9440                 tp->ts_recent = to->to_tsval;
9441         }
9442         /*
9443          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9444          * is on (half-synchronized state), then queue data for later
9445          * processing; else drop segment and return.
9446          */
9447         if ((thflags & TH_ACK) == 0) {
9448                 if (tp->t_flags & TF_NEEDSYN) {
9449                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9450                             tiwin, thflags, nxt_pkt));
9451                 } else if (tp->t_flags & TF_ACKNOW) {
9452                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9453                         bbr->r_wanted_output = 1;
9454                         return (ret_val);
9455                 } else {
9456                         ctf_do_drop(m, NULL);
9457                         return (0);
9458                 }
9459         }
9460         /*
9461          * Ack processing.
9462          */
9463         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
9464                 return (ret_val);
9465         }
9466         if (sbavail(&so->so_snd)) {
9467                 if (ctf_progress_timeout_check(tp, true)) {
9468                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9469                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9470                         return (1);
9471                 }
9472         }
9473         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9474             tiwin, thflags, nxt_pkt));
9475 }
9476
9477 static int
9478 bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr,
9479     struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so)
9480 {
9481
9482         if (bbr->rc_allow_data_af_clo == 0) {
9483 close_now:
9484                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
9485                 /* tcp_close will kill the inp pre-log the Reset */
9486                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9487                 tp = tcp_close(tp);
9488                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
9489                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
9490                 return (1);
9491         }
9492         if (sbavail(&so->so_snd) == 0)
9493                 goto close_now;
9494         /* Ok we allow data that is ignored and a followup reset */
9495         tp->rcv_nxt = th->th_seq + *tlen;
9496         tp->t_flags2 |= TF2_DROP_AF_DATA;
9497         bbr->r_wanted_output = 1;
9498         *tlen = 0;
9499         return (0);
9500 }
9501
9502 /*
9503  * Return value of 1, the TCB is unlocked and most
9504  * likely gone, return value of 0, the TCB is still
9505  * locked.
9506  */
9507 static int
9508 bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
9509     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9510     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9511 {
9512         int32_t ourfinisacked = 0;
9513         int32_t ret_val;
9514         struct tcp_bbr *bbr;
9515
9516         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9517         ctf_calc_rwin(so, tp);
9518         if ((thflags & TH_RST) ||
9519             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9520                 return (ctf_process_rst(m, th, so, tp));
9521         /*
9522          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9523          * synchronized state.
9524          */
9525         if (thflags & TH_SYN) {
9526                 ctf_challenge_ack(m, th, tp, &ret_val);
9527                 return (ret_val);
9528         }
9529         /*
9530          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9531          * it's less than ts_recent, drop it.
9532          */
9533         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9534             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9535                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9536                         return (ret_val);
9537         }
9538         INP_WLOCK_ASSERT(tp->t_inpcb);
9539         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9540                 return (ret_val);
9541         }
9542         /*
9543          * If new data are received on a connection after the user processes
9544          * are gone, then RST the other end.
9545          */
9546         if ((so->so_state & SS_NOFDREF) && tlen) {
9547                 /*
9548                  * We call a new function now so we might continue and setup
9549                  * to reset at all data being ack'd.
9550                  */
9551                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9552                         return (1);
9553         }
9554         /*
9555          * If last ACK falls within this segment's sequence numbers, record
9556          * its timestamp. NOTE: 1) That the test incorporates suggestions
9557          * from the latest proposal of the tcplw@cray.com list (Braden
9558          * 1993/04/26). 2) That updating only on newer timestamps interferes
9559          * with our earlier PAWS tests, so this check should be solely
9560          * predicated on the sequence space of this segment. 3) That we
9561          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9562          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9563          * SEG.Len, This modified check allows us to overcome RFC1323's
9564          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9565          * p.869. In such cases, we can still calculate the RTT correctly
9566          * when RCV.NXT == Last.ACK.Sent.
9567          */
9568         if ((to->to_flags & TOF_TS) != 0 &&
9569             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9570             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9571             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9572                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9573                 tp->ts_recent = to->to_tsval;
9574         }
9575         /*
9576          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9577          * is on (half-synchronized state), then queue data for later
9578          * processing; else drop segment and return.
9579          */
9580         if ((thflags & TH_ACK) == 0) {
9581                 if (tp->t_flags & TF_NEEDSYN) {
9582                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9583                             tiwin, thflags, nxt_pkt));
9584                 } else if (tp->t_flags & TF_ACKNOW) {
9585                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9586                         bbr->r_wanted_output = 1;
9587                         return (ret_val);
9588                 } else {
9589                         ctf_do_drop(m, NULL);
9590                         return (0);
9591                 }
9592         }
9593         /*
9594          * Ack processing.
9595          */
9596         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9597                 return (ret_val);
9598         }
9599         if (ourfinisacked) {
9600                 /*
9601                  * If we can't receive any more data, then closing user can
9602                  * proceed. Starting the timer is contrary to the
9603                  * specification, but if we don't get a FIN we'll hang
9604                  * forever.
9605                  *
9606                  * XXXjl: we should release the tp also, and use a
9607                  * compressed state.
9608                  */
9609                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
9610                         soisdisconnected(so);
9611                         tcp_timer_activate(tp, TT_2MSL,
9612                             (tcp_fast_finwait2_recycle ?
9613                             tcp_finwait2_timeout :
9614                             TP_MAXIDLE(tp)));
9615                 }
9616                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
9617         }
9618         if (sbavail(&so->so_snd)) {
9619                 if (ctf_progress_timeout_check(tp, true)) {
9620                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9621                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9622                         return (1);
9623                 }
9624         }
9625         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9626             tiwin, thflags, nxt_pkt));
9627 }
9628
9629 /*
9630  * Return value of 1, the TCB is unlocked and most
9631  * likely gone, return value of 0, the TCB is still
9632  * locked.
9633  */
9634 static int
9635 bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
9636     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9637     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9638 {
9639         int32_t ourfinisacked = 0;
9640         int32_t ret_val;
9641         struct tcp_bbr *bbr;
9642
9643         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9644         ctf_calc_rwin(so, tp);
9645         if ((thflags & TH_RST) ||
9646             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9647                 return (ctf_process_rst(m, th, so, tp));
9648         /*
9649          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9650          * synchronized state.
9651          */
9652         if (thflags & TH_SYN) {
9653                 ctf_challenge_ack(m, th, tp, &ret_val);
9654                 return (ret_val);
9655         }
9656         /*
9657          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9658          * it's less than ts_recent, drop it.
9659          */
9660         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9661             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9662                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9663                         return (ret_val);
9664         }
9665         INP_WLOCK_ASSERT(tp->t_inpcb);
9666         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9667                 return (ret_val);
9668         }
9669         /*
9670          * If new data are received on a connection after the user processes
9671          * are gone, then RST the other end.
9672          */
9673         if ((so->so_state & SS_NOFDREF) && tlen) {
9674                 /*
9675                  * We call a new function now so we might continue and setup
9676                  * to reset at all data being ack'd.
9677                  */
9678                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9679                         return (1);
9680         }
9681         /*
9682          * If last ACK falls within this segment's sequence numbers, record
9683          * its timestamp. NOTE: 1) That the test incorporates suggestions
9684          * from the latest proposal of the tcplw@cray.com list (Braden
9685          * 1993/04/26). 2) That updating only on newer timestamps interferes
9686          * with our earlier PAWS tests, so this check should be solely
9687          * predicated on the sequence space of this segment. 3) That we
9688          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9689          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9690          * SEG.Len, This modified check allows us to overcome RFC1323's
9691          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9692          * p.869. In such cases, we can still calculate the RTT correctly
9693          * when RCV.NXT == Last.ACK.Sent.
9694          */
9695         if ((to->to_flags & TOF_TS) != 0 &&
9696             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9697             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9698             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9699                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9700                 tp->ts_recent = to->to_tsval;
9701         }
9702         /*
9703          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9704          * is on (half-synchronized state), then queue data for later
9705          * processing; else drop segment and return.
9706          */
9707         if ((thflags & TH_ACK) == 0) {
9708                 if (tp->t_flags & TF_NEEDSYN) {
9709                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9710                             tiwin, thflags, nxt_pkt));
9711                 } else if (tp->t_flags & TF_ACKNOW) {
9712                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9713                         bbr->r_wanted_output = 1;
9714                         return (ret_val);
9715                 } else {
9716                         ctf_do_drop(m, NULL);
9717                         return (0);
9718                 }
9719         }
9720         /*
9721          * Ack processing.
9722          */
9723         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9724                 return (ret_val);
9725         }
9726         if (ourfinisacked) {
9727                 tcp_twstart(tp);
9728                 m_freem(m);
9729                 return (1);
9730         }
9731         if (sbavail(&so->so_snd)) {
9732                 if (ctf_progress_timeout_check(tp, true)) {
9733                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9734                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9735                         return (1);
9736                 }
9737         }
9738         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9739             tiwin, thflags, nxt_pkt));
9740 }
9741
9742 /*
9743  * Return value of 1, the TCB is unlocked and most
9744  * likely gone, return value of 0, the TCB is still
9745  * locked.
9746  */
9747 static int
9748 bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9749     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9750     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9751 {
9752         int32_t ourfinisacked = 0;
9753         int32_t ret_val;
9754         struct tcp_bbr *bbr;
9755
9756         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9757         ctf_calc_rwin(so, tp);
9758         if ((thflags & TH_RST) ||
9759             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9760                 return (ctf_process_rst(m, th, so, tp));
9761         /*
9762          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9763          * synchronized state.
9764          */
9765         if (thflags & TH_SYN) {
9766                 ctf_challenge_ack(m, th, tp, &ret_val);
9767                 return (ret_val);
9768         }
9769         /*
9770          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9771          * it's less than ts_recent, drop it.
9772          */
9773         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9774             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9775                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9776                         return (ret_val);
9777         }
9778         INP_WLOCK_ASSERT(tp->t_inpcb);
9779         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9780                 return (ret_val);
9781         }
9782         /*
9783          * If new data are received on a connection after the user processes
9784          * are gone, then RST the other end.
9785          */
9786         if ((so->so_state & SS_NOFDREF) && tlen) {
9787                 /*
9788                  * We call a new function now so we might continue and setup
9789                  * to reset at all data being ack'd.
9790                  */
9791                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9792                         return (1);
9793         }
9794         /*
9795          * If last ACK falls within this segment's sequence numbers, record
9796          * its timestamp. NOTE: 1) That the test incorporates suggestions
9797          * from the latest proposal of the tcplw@cray.com list (Braden
9798          * 1993/04/26). 2) That updating only on newer timestamps interferes
9799          * with our earlier PAWS tests, so this check should be solely
9800          * predicated on the sequence space of this segment. 3) That we
9801          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9802          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9803          * SEG.Len, This modified check allows us to overcome RFC1323's
9804          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9805          * p.869. In such cases, we can still calculate the RTT correctly
9806          * when RCV.NXT == Last.ACK.Sent.
9807          */
9808         if ((to->to_flags & TOF_TS) != 0 &&
9809             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9810             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9811             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9812                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9813                 tp->ts_recent = to->to_tsval;
9814         }
9815         /*
9816          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9817          * is on (half-synchronized state), then queue data for later
9818          * processing; else drop segment and return.
9819          */
9820         if ((thflags & TH_ACK) == 0) {
9821                 if (tp->t_flags & TF_NEEDSYN) {
9822                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9823                             tiwin, thflags, nxt_pkt));
9824                 } else if (tp->t_flags & TF_ACKNOW) {
9825                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9826                         bbr->r_wanted_output = 1;
9827                         return (ret_val);
9828                 } else {
9829                         ctf_do_drop(m, NULL);
9830                         return (0);
9831                 }
9832         }
9833         /*
9834          * case TCPS_LAST_ACK: Ack processing.
9835          */
9836         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9837                 return (ret_val);
9838         }
9839         if (ourfinisacked) {
9840                 tp = tcp_close(tp);
9841                 ctf_do_drop(m, tp);
9842                 return (1);
9843         }
9844         if (sbavail(&so->so_snd)) {
9845                 if (ctf_progress_timeout_check(tp, true)) {
9846                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9847                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9848                         return (1);
9849                 }
9850         }
9851         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9852             tiwin, thflags, nxt_pkt));
9853 }
9854
9855 /*
9856  * Return value of 1, the TCB is unlocked and most
9857  * likely gone, return value of 0, the TCB is still
9858  * locked.
9859  */
9860 static int
9861 bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
9862     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
9863     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
9864 {
9865         int32_t ourfinisacked = 0;
9866         int32_t ret_val;
9867         struct tcp_bbr *bbr;
9868
9869         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9870         ctf_calc_rwin(so, tp);
9871         /* Reset receive buffer auto scaling when not in bulk receive mode. */
9872         if ((thflags & TH_RST) ||
9873             (tp->t_fin_is_rst && (thflags & TH_FIN)))
9874                 return (ctf_process_rst(m, th, so, tp));
9875
9876         /*
9877          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
9878          * synchronized state.
9879          */
9880         if (thflags & TH_SYN) {
9881                 ctf_challenge_ack(m, th, tp, &ret_val);
9882                 return (ret_val);
9883         }
9884         INP_WLOCK_ASSERT(tp->t_inpcb);
9885         /*
9886          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
9887          * it's less than ts_recent, drop it.
9888          */
9889         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
9890             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
9891                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
9892                         return (ret_val);
9893         }
9894         INP_WLOCK_ASSERT(tp->t_inpcb);
9895         if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
9896                 return (ret_val);
9897         }
9898         /*
9899          * If new data are received on a connection after the user processes
9900          * are gone, then we may RST the other end depending on the outcome
9901          * of bbr_check_data_after_close.
9902          */
9903         if ((so->so_state & SS_NOFDREF) &&
9904             tlen) {
9905                 /*
9906                  * We call a new function now so we might continue and setup
9907                  * to reset at all data being ack'd.
9908                  */
9909                 if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so))
9910                         return (1);
9911         }
9912         INP_WLOCK_ASSERT(tp->t_inpcb);
9913         /*
9914          * If last ACK falls within this segment's sequence numbers, record
9915          * its timestamp. NOTE: 1) That the test incorporates suggestions
9916          * from the latest proposal of the tcplw@cray.com list (Braden
9917          * 1993/04/26). 2) That updating only on newer timestamps interferes
9918          * with our earlier PAWS tests, so this check should be solely
9919          * predicated on the sequence space of this segment. 3) That we
9920          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
9921          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
9922          * SEG.Len, This modified check allows us to overcome RFC1323's
9923          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
9924          * p.869. In such cases, we can still calculate the RTT correctly
9925          * when RCV.NXT == Last.ACK.Sent.
9926          */
9927         INP_WLOCK_ASSERT(tp->t_inpcb);
9928         if ((to->to_flags & TOF_TS) != 0 &&
9929             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
9930             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
9931             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
9932                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
9933                 tp->ts_recent = to->to_tsval;
9934         }
9935         /*
9936          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
9937          * is on (half-synchronized state), then queue data for later
9938          * processing; else drop segment and return.
9939          */
9940         if ((thflags & TH_ACK) == 0) {
9941                 if (tp->t_flags & TF_NEEDSYN) {
9942                         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9943                             tiwin, thflags, nxt_pkt));
9944                 } else if (tp->t_flags & TF_ACKNOW) {
9945                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
9946                         bbr->r_wanted_output = 1;
9947                         return (ret_val);
9948                 } else {
9949                         ctf_do_drop(m, NULL);
9950                         return (0);
9951                 }
9952         }
9953         /*
9954          * Ack processing.
9955          */
9956         INP_WLOCK_ASSERT(tp->t_inpcb);
9957         if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
9958                 return (ret_val);
9959         }
9960         if (sbavail(&so->so_snd)) {
9961                 if (ctf_progress_timeout_check(tp, true)) {
9962                         bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__);
9963                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
9964                         return (1);
9965                 }
9966         }
9967         INP_WLOCK_ASSERT(tp->t_inpcb);
9968         return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen,
9969             tiwin, thflags, nxt_pkt));
9970 }
9971
9972 static void
9973 bbr_stop_all_timers(struct tcpcb *tp)
9974 {
9975         struct tcp_bbr *bbr;
9976
9977         /*
9978          * Assure no timers are running.
9979          */
9980         if (tcp_timer_active(tp, TT_PERSIST)) {
9981                 /* We enter in persists, set the flag appropriately */
9982                 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
9983                 bbr->rc_in_persist = 1;
9984         }
9985         tcp_timer_suspend(tp, TT_PERSIST);
9986         tcp_timer_suspend(tp, TT_REXMT);
9987         tcp_timer_suspend(tp, TT_KEEP);
9988         tcp_timer_suspend(tp, TT_DELACK);
9989 }
9990
9991 static void
9992 bbr_google_mode_on(struct tcp_bbr *bbr)
9993 {
9994         bbr->rc_use_google = 1;
9995         bbr->rc_no_pacing = 0;
9996         bbr->r_ctl.bbr_google_discount = bbr_google_discount;
9997         bbr->r_use_policer = bbr_policer_detection_enabled;
9998         bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
9999         bbr->bbr_use_rack_cheat = 0;
10000         bbr->r_ctl.rc_incr_tmrs = 0;
10001         bbr->r_ctl.rc_inc_tcp_oh = 0;
10002         bbr->r_ctl.rc_inc_ip_oh = 0;
10003         bbr->r_ctl.rc_inc_enet_oh = 0;
10004         reset_time(&bbr->r_ctl.rc_delrate,
10005                    BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10006         reset_time_small(&bbr->r_ctl.rc_rttprop,
10007                          (11 * USECS_IN_SECOND));
10008         tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10009 }
10010
10011 static void
10012 bbr_google_mode_off(struct tcp_bbr *bbr)
10013 {
10014         bbr->rc_use_google = 0;
10015         bbr->r_ctl.bbr_google_discount = 0;
10016         bbr->no_pacing_until = bbr_no_pacing_until;
10017         bbr->r_use_policer = 0;
10018         if (bbr->no_pacing_until)
10019                 bbr->rc_no_pacing = 1;
10020         else
10021                 bbr->rc_no_pacing = 0;
10022         if (bbr_use_rack_resend_cheat)
10023                 bbr->bbr_use_rack_cheat = 1;
10024         else
10025                 bbr->bbr_use_rack_cheat = 0;
10026         if (bbr_incr_timers)
10027                 bbr->r_ctl.rc_incr_tmrs = 1;
10028         else
10029                 bbr->r_ctl.rc_incr_tmrs = 0;
10030         if (bbr_include_tcp_oh)
10031                 bbr->r_ctl.rc_inc_tcp_oh = 1;
10032         else
10033                 bbr->r_ctl.rc_inc_tcp_oh = 0;
10034         if (bbr_include_ip_oh)
10035                 bbr->r_ctl.rc_inc_ip_oh = 1;
10036         else
10037                 bbr->r_ctl.rc_inc_ip_oh = 0;
10038         if (bbr_include_enet_oh)
10039                 bbr->r_ctl.rc_inc_enet_oh = 1;
10040         else
10041                 bbr->r_ctl.rc_inc_enet_oh = 0;
10042         bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10043         reset_time(&bbr->r_ctl.rc_delrate,
10044                    bbr_num_pktepo_for_del_limit);
10045         reset_time_small(&bbr->r_ctl.rc_rttprop,
10046                          (bbr_filter_len_sec * USECS_IN_SECOND));
10047         tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv));
10048 }
10049 /*
10050  * Return 0 on success, non-zero on failure
10051  * which indicates the error (usually no memory).
10052  */
10053 static int
10054 bbr_init(struct tcpcb *tp)
10055 {
10056         struct tcp_bbr *bbr = NULL;
10057         struct inpcb *inp;
10058         uint32_t cts;
10059
10060         tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO));
10061         if (tp->t_fb_ptr == NULL) {
10062                 /*
10063                  * We need to allocate memory but cant. The INP and INP_INFO
10064                  * locks and they are recusive (happens during setup. So a
10065                  * scheme to drop the locks fails :(
10066                  *
10067                  */
10068                 return (ENOMEM);
10069         }
10070         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10071         bbr->rtt_valid = 0;
10072         inp = tp->t_inpcb;
10073         inp->inp_flags2 |= INP_CANNOT_DO_ECN;
10074         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
10075         TAILQ_INIT(&bbr->r_ctl.rc_map);
10076         TAILQ_INIT(&bbr->r_ctl.rc_free);
10077         TAILQ_INIT(&bbr->r_ctl.rc_tmap);
10078         bbr->rc_tp = tp;
10079         if (tp->t_inpcb) {
10080                 bbr->rc_inp = tp->t_inpcb;
10081         }
10082         cts = tcp_get_usecs(&bbr->rc_tv);
10083         tp->t_acktime = 0;
10084         bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close;
10085         bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade;
10086         bbr->rc_tlp_threshold = bbr_tlp_thresh;
10087         bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh;
10088         bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay;
10089         bbr->r_ctl.rc_min_to = bbr_min_to;
10090         bbr->rc_bbr_state = BBR_STATE_STARTUP;
10091         bbr->r_ctl.bbr_lost_at_state = 0;
10092         bbr->r_ctl.rc_lost_at_startup = 0;
10093         bbr->rc_all_timers_stopped = 0;
10094         bbr->r_ctl.rc_bbr_lastbtlbw = 0;
10095         bbr->r_ctl.rc_pkt_epoch_del = 0;
10096         bbr->r_ctl.rc_pkt_epoch = 0;
10097         bbr->r_ctl.rc_lowest_rtt = 0xffffffff;
10098         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain;
10099         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
10100         bbr->r_ctl.rc_went_idle_time = cts;
10101         bbr->rc_pacer_started = cts;
10102         bbr->r_ctl.rc_pkt_epoch_time = cts;
10103         bbr->r_ctl.rc_rcvtime = cts;
10104         bbr->r_ctl.rc_bbr_state_time = cts;
10105         bbr->r_ctl.rc_del_time = cts;
10106         bbr->r_ctl.rc_tlp_rxt_last_time = cts;
10107         bbr->r_ctl.last_in_probertt = cts;
10108         bbr->skip_gain = 0;
10109         bbr->gain_is_limited = 0;
10110         bbr->no_pacing_until = bbr_no_pacing_until;
10111         if (bbr->no_pacing_until)
10112                 bbr->rc_no_pacing = 1;
10113         if (bbr_use_google_algo) {
10114                 bbr->rc_no_pacing = 0;
10115                 bbr->rc_use_google = 1;
10116                 bbr->r_ctl.bbr_google_discount = bbr_google_discount;
10117                 bbr->r_use_policer = bbr_policer_detection_enabled;
10118         } else {
10119                 bbr->rc_use_google = 0;
10120                 bbr->r_ctl.bbr_google_discount = 0;
10121                 bbr->r_use_policer = 0;
10122         }
10123         if (bbr_ts_limiting)
10124                 bbr->rc_use_ts_limit = 1;
10125         else
10126                 bbr->rc_use_ts_limit = 0;
10127         if (bbr_ts_can_raise)
10128                 bbr->ts_can_raise = 1;
10129         else
10130                 bbr->ts_can_raise = 0;
10131         if (V_tcp_delack_enabled == 1)
10132                 tp->t_delayed_ack = 2;
10133         else if (V_tcp_delack_enabled == 0)
10134                 tp->t_delayed_ack = 0;
10135         else if (V_tcp_delack_enabled < 100)
10136                 tp->t_delayed_ack = V_tcp_delack_enabled;
10137         else
10138                 tp->t_delayed_ack = 2;
10139         if (bbr->rc_use_google == 0)
10140                 bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10141         else
10142                 bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10);
10143         bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms;
10144         bbr->rc_max_rto_sec = bbr_rto_max_sec;
10145         bbr->rc_init_win = bbr_def_init_win;
10146         if (tp->t_flags & TF_REQ_TSTMP)
10147                 bbr->rc_last_options = TCP_TS_OVERHEAD;
10148         bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options;
10149         bbr->r_ctl.rc_high_rwnd = tp->snd_wnd;
10150         bbr->r_init_rtt = 1;
10151
10152         counter_u64_add(bbr_flows_nohdwr_pacing, 1);
10153         if (bbr_allow_hdwr_pacing)
10154                 bbr->bbr_hdw_pace_ena = 1;
10155         else
10156                 bbr->bbr_hdw_pace_ena = 0;
10157         if (bbr_sends_full_iwnd)
10158                 bbr->bbr_init_win_cheat = 1;
10159         else
10160                 bbr->bbr_init_win_cheat = 0;
10161         bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max;
10162         bbr->r_ctl.rc_drain_pg = bbr_drain_gain;
10163         bbr->r_ctl.rc_startup_pg = bbr_high_gain;
10164         bbr->rc_loss_exit = bbr_exit_startup_at_loss;
10165         bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain;
10166         bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second;
10167         bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar;
10168         bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max;
10169         bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor;
10170         bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min;
10171         bbr->r_ctl.bbr_cross_over = bbr_cross_over;
10172         bbr->r_ctl.rc_rtt_shrinks = cts;
10173         if (bbr->rc_use_google) {
10174                 setup_time_filter(&bbr->r_ctl.rc_delrate,
10175                                   FILTER_TYPE_MAX,
10176                                   BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT);
10177                 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10178                                         FILTER_TYPE_MIN, (11 * USECS_IN_SECOND));
10179         } else {
10180                 setup_time_filter(&bbr->r_ctl.rc_delrate,
10181                                   FILTER_TYPE_MAX,
10182                                   bbr_num_pktepo_for_del_limit);
10183                 setup_time_filter_small(&bbr->r_ctl.rc_rttprop,
10184                                         FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND));
10185         }
10186         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0);
10187         if (bbr_uses_idle_restart)
10188                 bbr->rc_use_idle_restart = 1;
10189         else
10190                 bbr->rc_use_idle_restart = 0;
10191         bbr->r_ctl.rc_bbr_cur_del_rate = 0;
10192         bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps;
10193         if (bbr_resends_use_tso)
10194                 bbr->rc_resends_use_tso = 1;
10195 #ifdef NETFLIX_PEAKRATE
10196         tp->t_peakrate_thr = tp->t_maxpeakrate;
10197 #endif
10198         if (tp->snd_una != tp->snd_max) {
10199                 /* Create a send map for the current outstanding data */
10200                 struct bbr_sendmap *rsm;
10201
10202                 rsm = bbr_alloc(bbr);
10203                 if (rsm == NULL) {
10204                         uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10205                         tp->t_fb_ptr = NULL;
10206                         return (ENOMEM);
10207                 }
10208                 rsm->r_flags = BBR_OVERMAX;
10209                 rsm->r_tim_lastsent[0] = cts;
10210                 rsm->r_rtr_cnt = 1;
10211                 rsm->r_rtr_bytes = 0;
10212                 rsm->r_start = tp->snd_una;
10213                 rsm->r_end = tp->snd_max;
10214                 rsm->r_dupack = 0;
10215                 rsm->r_delivered = bbr->r_ctl.rc_delivered;
10216                 rsm->r_ts_valid = 0;
10217                 rsm->r_del_ack_ts = tp->ts_recent;
10218                 rsm->r_del_time = cts;
10219                 if (bbr->r_ctl.r_app_limited_until)
10220                         rsm->r_app_limited = 1;
10221                 else
10222                         rsm->r_app_limited = 0;
10223                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next);
10224                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext);
10225                 rsm->r_in_tmap = 1;
10226                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW)
10227                         rsm->r_bbr_state = bbr_state_val(bbr);
10228                 else
10229                         rsm->r_bbr_state = 8;
10230         }
10231         if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0))
10232                 bbr->bbr_use_rack_cheat = 1;
10233         if (bbr_incr_timers && (bbr->rc_use_google == 0))
10234                 bbr->r_ctl.rc_incr_tmrs = 1;
10235         if (bbr_include_tcp_oh && (bbr->rc_use_google == 0))
10236                 bbr->r_ctl.rc_inc_tcp_oh = 1;
10237         if (bbr_include_ip_oh && (bbr->rc_use_google == 0))
10238                 bbr->r_ctl.rc_inc_ip_oh = 1;
10239         if (bbr_include_enet_oh && (bbr->rc_use_google == 0))
10240                 bbr->r_ctl.rc_inc_enet_oh = 1;
10241
10242         bbr_log_type_statechange(bbr, cts, __LINE__);
10243         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
10244             (tp->t_srtt)) {
10245                 uint32_t rtt;
10246
10247                 rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT);
10248                 apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts);
10249         }
10250         /* announce the settings and state */
10251         bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT);
10252         tcp_bbr_tso_size_check(bbr, cts);
10253         /*
10254          * Now call the generic function to start a timer. This will place
10255          * the TCB on the hptsi wheel if a timer is needed with appropriate
10256          * flags.
10257          */
10258         bbr_stop_all_timers(tp);
10259         bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0);
10260         return (0);
10261 }
10262
10263 /*
10264  * Return 0 if we can accept the connection. Return
10265  * non-zero if we can't handle the connection. A EAGAIN
10266  * means you need to wait until the connection is up.
10267  * a EADDRNOTAVAIL means we can never handle the connection
10268  * (no SACK).
10269  */
10270 static int
10271 bbr_handoff_ok(struct tcpcb *tp)
10272 {
10273         if ((tp->t_state == TCPS_CLOSED) ||
10274             (tp->t_state == TCPS_LISTEN)) {
10275                 /* Sure no problem though it may not stick */
10276                 return (0);
10277         }
10278         if ((tp->t_state == TCPS_SYN_SENT) ||
10279             (tp->t_state == TCPS_SYN_RECEIVED)) {
10280                 /*
10281                  * We really don't know you have to get to ESTAB or beyond
10282                  * to tell.
10283                  */
10284                 return (EAGAIN);
10285         }
10286         if (tp->t_flags & TF_SENTFIN)
10287                 return (EINVAL);
10288         if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) {
10289                 return (0);
10290         }
10291         /*
10292          * If we reach here we don't do SACK on this connection so we can
10293          * never do rack.
10294          */
10295         return (EINVAL);
10296 }
10297
10298 static void
10299 bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged)
10300 {
10301         if (tp->t_fb_ptr) {
10302                 uint32_t calc;
10303                 struct tcp_bbr *bbr;
10304                 struct bbr_sendmap *rsm;
10305
10306                 bbr = (struct tcp_bbr *)tp->t_fb_ptr;
10307                 if (bbr->r_ctl.crte)
10308                         tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
10309                 bbr_log_flowend(bbr);
10310                 bbr->rc_tp = NULL;
10311                 if (tp->t_inpcb) {
10312                         /* Backout any flags2 we applied */
10313                         tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN;
10314                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
10315                         tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
10316                 }
10317                 if (bbr->bbr_hdrw_pacing)
10318                         counter_u64_add(bbr_flows_whdwr_pacing, -1);
10319                 else
10320                         counter_u64_add(bbr_flows_nohdwr_pacing, -1);
10321                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10322                 while (rsm) {
10323                         TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next);
10324                         uma_zfree(bbr_zone, rsm);
10325                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
10326                 }
10327                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10328                 while (rsm) {
10329                         TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next);
10330                         uma_zfree(bbr_zone, rsm);
10331                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free);
10332                 }
10333                 calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd;
10334                 if (calc > (bbr->r_ctl.rc_init_rwnd / 10))
10335                         BBR_STAT_INC(bbr_dynamic_rwnd);
10336                 else
10337                         BBR_STAT_INC(bbr_static_rwnd);
10338                 bbr->r_ctl.rc_free_cnt = 0;
10339                 uma_zfree(bbr_pcb_zone, tp->t_fb_ptr);
10340                 tp->t_fb_ptr = NULL;
10341         }
10342         /* Make sure snd_nxt is correctly set */
10343         tp->snd_nxt = tp->snd_max;
10344 }
10345
10346 static void
10347 bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win)
10348 {
10349         switch (tp->t_state) {
10350         case TCPS_SYN_SENT:
10351                 bbr->r_state = TCPS_SYN_SENT;
10352                 bbr->r_substate = bbr_do_syn_sent;
10353                 break;
10354         case TCPS_SYN_RECEIVED:
10355                 bbr->r_state = TCPS_SYN_RECEIVED;
10356                 bbr->r_substate = bbr_do_syn_recv;
10357                 break;
10358         case TCPS_ESTABLISHED:
10359                 bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd);
10360                 bbr->r_state = TCPS_ESTABLISHED;
10361                 bbr->r_substate = bbr_do_established;
10362                 break;
10363         case TCPS_CLOSE_WAIT:
10364                 bbr->r_state = TCPS_CLOSE_WAIT;
10365                 bbr->r_substate = bbr_do_close_wait;
10366                 break;
10367         case TCPS_FIN_WAIT_1:
10368                 bbr->r_state = TCPS_FIN_WAIT_1;
10369                 bbr->r_substate = bbr_do_fin_wait_1;
10370                 break;
10371         case TCPS_CLOSING:
10372                 bbr->r_state = TCPS_CLOSING;
10373                 bbr->r_substate = bbr_do_closing;
10374                 break;
10375         case TCPS_LAST_ACK:
10376                 bbr->r_state = TCPS_LAST_ACK;
10377                 bbr->r_substate = bbr_do_lastack;
10378                 break;
10379         case TCPS_FIN_WAIT_2:
10380                 bbr->r_state = TCPS_FIN_WAIT_2;
10381                 bbr->r_substate = bbr_do_fin_wait_2;
10382                 break;
10383         case TCPS_LISTEN:
10384         case TCPS_CLOSED:
10385         case TCPS_TIME_WAIT:
10386         default:
10387                 break;
10388         };
10389 }
10390
10391 static void
10392 bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog)
10393 {
10394         /*
10395          * Now what state are we going into now? Is there adjustments
10396          * needed?
10397          */
10398         int32_t old_state, old_gain;
10399
10400         old_state = bbr_state_val(bbr);
10401         old_gain = bbr->r_ctl.rc_bbr_hptsi_gain;
10402         if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) {
10403                 /* Save the lowest srtt we saw in our end of the sub-state */
10404                 bbr->rc_hit_state_1 = 0;
10405                 if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff)
10406                         bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state;
10407         }
10408         bbr->rc_bbr_substate++;
10409         if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) {
10410                 /* Cycle back to first state-> gain */
10411                 bbr->rc_bbr_substate = 0;
10412         }
10413         if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10414                 /*
10415                  * We enter the gain(5/4) cycle (possibly less if
10416                  * shallow buffer detection is enabled)
10417                  */
10418                 if (bbr->skip_gain) {
10419                         /*
10420                          * Hardware pacing has set our rate to
10421                          * the max and limited our b/w just
10422                          * do level i.e. no gain.
10423                          */
10424                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1];
10425                 } else if (bbr->gain_is_limited &&
10426                            bbr->bbr_hdrw_pacing &&
10427                            bbr->r_ctl.crte) {
10428                         /*
10429                          * We can't gain above the hardware pacing
10430                          * rate which is less than our rate + the gain
10431                          * calculate the gain needed to reach the hardware
10432                          * pacing rate..
10433                          */
10434                         uint64_t bw, rate, gain_calc;
10435
10436                         bw = bbr_get_bw(bbr);
10437                         rate = bbr->r_ctl.crte->rate;
10438                         if ((rate > bw) &&
10439                             (((bw *  (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) {
10440                                 gain_calc = (rate * BBR_UNIT) / bw;
10441                                 if (gain_calc < BBR_UNIT)
10442                                         gain_calc = BBR_UNIT;
10443                                 bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc;
10444                         } else {
10445                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10446                         }
10447                 } else
10448                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN];
10449                 if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) {
10450                         bbr->r_ctl.rc_bbr_state_atflight = cts;
10451                 } else
10452                         bbr->r_ctl.rc_bbr_state_atflight = 0;
10453         } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10454                 bbr->rc_hit_state_1 = 1;
10455                 bbr->r_ctl.rc_exta_time_gd = 0;
10456                 bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10457                                                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10458                 if (bbr_state_drain_2_tar) {
10459                         bbr->r_ctl.rc_bbr_state_atflight = 0;
10460                 } else
10461                         bbr->r_ctl.rc_bbr_state_atflight = cts;
10462                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN];
10463         } else {
10464                 /* All other cycles hit here 2-7 */
10465                 if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) {
10466                         if (bbr_sub_drain_slam_cwnd &&
10467                             (bbr->rc_use_google == 0) &&
10468                             (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10469                                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10470                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10471                         }
10472                         if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP))
10473                                 bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) -
10474                                                                bbr_get_rtt(bbr, BBR_RTT_PROP));
10475                         else
10476                                 bbr->r_ctl.rc_exta_time_gd = 0;
10477                         if (bbr->r_ctl.rc_exta_time_gd) {
10478                                 bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd;
10479                                 /* Now chop up the time for each state (div by 7) */
10480                                 bbr->r_ctl.rc_level_state_extra /= 7;
10481                                 if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) {
10482                                         /* Add a randomization */
10483                                         bbr_randomize_extra_state_time(bbr);
10484                                 }
10485                         }
10486                 }
10487                 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10488                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)];
10489         }
10490         if (bbr->rc_use_google) {
10491                 bbr->r_ctl.rc_bbr_state_atflight = max(1, cts);
10492         }
10493         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10494         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10495         if (dolog)
10496                 bbr_log_type_statechange(bbr, cts, line);
10497
10498         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10499                 uint32_t time_in;
10500
10501                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10502                 if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
10503                         counter_u64_add(bbr_state_time[(old_state + 5)], time_in);
10504                 } else {
10505                         counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10506                 }
10507         }
10508         bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
10509         bbr_set_state_target(bbr, __LINE__);
10510         if (bbr_sub_drain_slam_cwnd &&
10511             (bbr->rc_use_google == 0) &&
10512             (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10513                 /* Slam down the cwnd */
10514                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10515                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10516                 if (bbr_sub_drain_app_limit) {
10517                         /* Go app limited if we are on a long drain */
10518                         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered +
10519                                                           ctf_flight_size(bbr->rc_tp,
10520                                                               (bbr->r_ctl.rc_sacked +
10521                                                                bbr->r_ctl.rc_lost_bytes)));
10522                 }
10523                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10524         }
10525         if (bbr->rc_lt_use_bw) {
10526                 /* In policed mode we clamp pacing_gain to BBR_UNIT */
10527                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10528         }
10529         /* Google changes TSO size every cycle */
10530         if (bbr->rc_use_google)
10531                 tcp_bbr_tso_size_check(bbr, cts);
10532         bbr->r_ctl.gain_epoch = cts;
10533         bbr->r_ctl.rc_bbr_state_time = cts;
10534         bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch;
10535 }
10536
10537 static void
10538 bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10539 {
10540         if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) &&
10541             (google_allow_early_out == 1) &&
10542             (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) {
10543                 /* We have reached out target flight size possibly early */
10544                 goto change_state;
10545         }
10546         if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10547                 return;
10548         }
10549         if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) {
10550                 /*
10551                  * Must be a rttProp movement forward before
10552                  * we can change states.
10553                  */
10554                 return;
10555         }
10556         if (bbr_state_val(bbr) == BBR_SUB_GAIN) {
10557                 /*
10558                  * The needed time has passed but for
10559                  * the gain cycle extra rules apply:
10560                  * 1) If we have seen loss, we exit
10561                  * 2) If we have not reached the target
10562                  *    we stay in GAIN (gain-to-target).
10563                  */
10564                 if (google_consider_lost && losses)
10565                         goto change_state;
10566                 if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) {
10567                         return;
10568                 }
10569         }
10570 change_state:
10571         /* For gain we must reach our target, all others last 1 rttProp */
10572         bbr_substate_change(bbr, cts, __LINE__, 1);
10573 }
10574
10575 static void
10576 bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses)
10577 {
10578         uint32_t flight, bbr_cur_cycle_time;
10579
10580         if (bbr->rc_use_google) {
10581                 bbr_set_probebw_google_gains(bbr, cts, losses);
10582                 return;
10583         }
10584         if (cts == 0) {
10585                 /*
10586                  * Never alow cts to be 0 we
10587                  * do this so we can judge if
10588                  * we have set a timestamp.
10589                  */
10590                 cts = 1;
10591         }
10592         if (bbr_state_is_pkt_epoch)
10593                 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT);
10594         else
10595                 bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP);
10596
10597         if (bbr->r_ctl.rc_bbr_state_atflight == 0) {
10598                 if (bbr_state_val(bbr) == BBR_SUB_DRAIN) {
10599                         flight = ctf_flight_size(bbr->rc_tp,
10600                                      (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10601                         if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) {
10602                                 /* Keep it slam down */
10603                                 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) {
10604                                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10605                                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10606                                 }
10607                                 if (bbr_sub_drain_app_limit) {
10608                                         /* Go app limited if we are on a long drain */
10609                                         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight);
10610                                 }
10611                         }
10612                         if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) &&
10613                             (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) ||
10614                              (flight >= bbr->r_ctl.flightsize_at_drain))) {
10615                                 /*
10616                                  * Still here after the same time as
10617                                  * the gain. We need to drain harder
10618                                  * for the next srtt. Reduce by a set amount
10619                                  * the gain drop is capped at DRAIN states
10620                                  * value (88).
10621                                  */
10622                                 bbr->r_ctl.flightsize_at_drain = flight;
10623                                 if (bbr_drain_drop_mul &&
10624                                     bbr_drain_drop_div &&
10625                                     (bbr_drain_drop_mul < bbr_drain_drop_div)) {
10626                                         /* Use your specific drop value (def 4/5 = 20%) */
10627                                         bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul;
10628                                         bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div;
10629                                 } else {
10630                                         /* You get drop of 20% */
10631                                         bbr->r_ctl.rc_bbr_hptsi_gain *= 4;
10632                                         bbr->r_ctl.rc_bbr_hptsi_gain /= 5;
10633                                 }
10634                                 if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) {
10635                                         /* Reduce our gain again to the bottom  */
10636                                         bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
10637                                 }
10638                                 bbr_log_exit_gain(bbr, cts, 4);
10639                                 /*
10640                                  * Extend out so we wait another
10641                                  * epoch before dropping again.
10642                                  */
10643                                 bbr->r_ctl.gain_epoch = cts;
10644                         }
10645                         if (flight <= bbr->r_ctl.rc_target_at_state) {
10646                                 if (bbr_sub_drain_slam_cwnd &&
10647                                     (bbr->rc_use_google == 0) &&
10648                                     (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
10649                                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10650                                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10651                                 }
10652                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10653                                 bbr_log_exit_gain(bbr, cts, 3);
10654                         }
10655                 } else {
10656                         /* Its a gain  */
10657                         if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) {
10658                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10659                                 goto change_state;
10660                         }
10661                         if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) ||
10662                             ((ctf_outstanding(bbr->rc_tp) +  bbr->rc_tp->t_maxseg - 1) >=
10663                              bbr->rc_tp->snd_wnd)) {
10664                                 bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1);
10665                                 bbr_log_exit_gain(bbr, cts, 2);
10666                         }
10667                 }
10668                 /**
10669                  * We fall through and return always one of two things has
10670                  * occured.
10671                  * 1) We are still not at target
10672                  *    <or>
10673                  * 2) We reached the target and set rc_bbr_state_atflight
10674                  *    which means we no longer hit this block
10675                  *    next time we are called.
10676                  */
10677                 return;
10678         }
10679 change_state:
10680         if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time))
10681                 return;
10682         if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) {
10683                 /* Less than a full time-period has passed */
10684                 return;
10685         }
10686         if (bbr->r_ctl.rc_level_state_extra &&
10687             (bbr_state_val(bbr) > BBR_SUB_DRAIN) &&
10688             ((cts - bbr->r_ctl.rc_bbr_state_time) <
10689              (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10690                 /* Less than a full time-period + extra has passed */
10691                 return;
10692         }
10693         if (bbr_gain_gets_extra_too &&
10694             bbr->r_ctl.rc_level_state_extra &&
10695             (bbr_state_val(bbr) == BBR_SUB_GAIN) &&
10696             ((cts - bbr->r_ctl.rc_bbr_state_time) <
10697              (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) {
10698                 /* Less than a full time-period + extra has passed */
10699                 return;
10700         }
10701         bbr_substate_change(bbr, cts, __LINE__, 1);
10702 }
10703
10704 static uint32_t
10705 bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain)
10706 {
10707         uint32_t mss, tar;
10708
10709         if (bbr->rc_use_google) {
10710                 /* Google just uses the cwnd target */
10711                 tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain);
10712         } else {
10713                 mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options),
10714                           bbr->r_ctl.rc_pace_max_segs);
10715                 /* Get the base cwnd with gain rounded to a mss */
10716                 tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr),
10717                                                       gain), mss);
10718                 /* Make sure it is within our min */
10719                 if (tar < get_min_cwnd(bbr))
10720                         return (get_min_cwnd(bbr));
10721         }
10722         return (tar);
10723 }
10724
10725 static void
10726 bbr_set_state_target(struct tcp_bbr *bbr, int line)
10727 {
10728         uint32_t tar, meth;
10729
10730         if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) &&
10731             ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) {
10732                 /* Special case using old probe-rtt method */
10733                 tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10734                 meth = 1;
10735         } else {
10736                 /* Non-probe-rtt case and reduced probe-rtt  */
10737                 if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) &&
10738                     (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) {
10739                         /* For gain cycle we use the hptsi gain */
10740                         tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10741                         meth = 2;
10742                 } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) {
10743                         /*
10744                          * If configured, or for google all other states
10745                          * get BBR_UNIT.
10746                          */
10747                         tar = bbr_get_a_state_target(bbr, BBR_UNIT);
10748                         meth = 3;
10749                 } else {
10750                         /*
10751                          * Or we set a target based on the pacing gain
10752                          * for non-google mode and default (non-configured).
10753                          * Note we don't set a target goal below drain (192).
10754                          */
10755                         if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN])  {
10756                                 tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]);
10757                                 meth = 4;
10758                         } else {
10759                                 tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain);
10760                                 meth = 5;
10761                         }
10762                 }
10763         }
10764         bbr_log_set_of_state_target(bbr, tar, line, meth);
10765         bbr->r_ctl.rc_target_at_state = tar;
10766 }
10767
10768 static void
10769 bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line)
10770 {
10771         /* Change to probe_rtt */
10772         uint32_t time_in;
10773
10774         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10775         bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp,
10776                                              (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
10777         bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain
10778                                           + bbr->r_ctl.rc_delivered);
10779         /* Setup so we force feed the filter */
10780         if (bbr->rc_use_google || bbr_probertt_sets_rtt)
10781                 bbr->rc_prtt_set_ts = 1;
10782         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10783                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10784                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10785         }
10786         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0);
10787         bbr->r_ctl.rc_rtt_shrinks = cts;
10788         bbr->r_ctl.last_in_probertt = cts;
10789         bbr->r_ctl.rc_probertt_srttchktim = cts;
10790         bbr->r_ctl.rc_bbr_state_time = cts;
10791         bbr->rc_bbr_state = BBR_STATE_PROBE_RTT;
10792         /* We need to force the filter to update */
10793
10794         if ((bbr_sub_drain_slam_cwnd) &&
10795             bbr->rc_hit_state_1 &&
10796             (bbr->rc_use_google == 0) &&
10797             (bbr_state_val(bbr) == BBR_SUB_DRAIN)) {
10798                 if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd)
10799                         bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10800         } else
10801                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
10802         /* Update the lost */
10803         bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10804         if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){
10805                 /* Set to the non-configurable default of 4 (PROBE_RTT_MIN)  */
10806                 bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
10807                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10808                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
10809                 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10810                 bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6);
10811                 bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd;
10812         } else {
10813                 /*
10814                  * We bring it down slowly by using a hptsi gain that is
10815                  * probably 75%. This will slowly float down our outstanding
10816                  * without tampering with the cwnd.
10817                  */
10818                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
10819                 bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
10820                 bbr_set_state_target(bbr, __LINE__);
10821                 if (bbr_prtt_slam_cwnd &&
10822                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
10823                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
10824                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10825                 }
10826         }
10827         if (ctf_flight_size(bbr->rc_tp,
10828                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
10829             bbr->r_ctl.rc_target_at_state) {
10830                 /* We are at target */
10831                 bbr->r_ctl.rc_bbr_enters_probertt = cts;
10832         } else {
10833                 /* We need to come down to reach target before our time begins */
10834                 bbr->r_ctl.rc_bbr_enters_probertt = 0;
10835         }
10836         bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch;
10837         BBR_STAT_INC(bbr_enter_probertt);
10838         bbr_log_exit_gain(bbr, cts, 0);
10839         bbr_log_type_statechange(bbr, cts, line);
10840 }
10841
10842 static void
10843 bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts)
10844 {
10845         /*
10846          * Sanity check on probe-rtt intervals.
10847          * In crazy situations where we are competing
10848          * against new-reno flows with huge buffers
10849          * our rtt-prop interval could come to dominate
10850          * things if we can't get through a full set
10851          * of cycles, we need to adjust it.
10852          */
10853         if (bbr_can_adjust_probertt &&
10854             (bbr->rc_use_google == 0)) {
10855                 uint16_t val = 0;
10856                 uint32_t cur_rttp, fval, newval, baseval;
10857
10858                 /* Are we to small and go into probe-rtt to often? */
10859                 baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1));
10860                 cur_rttp = roundup(baseval, USECS_IN_SECOND);
10861                 fval = bbr_filter_len_sec * USECS_IN_SECOND;
10862                 if (bbr_is_ratio == 0) {
10863                         if (fval > bbr_rtt_probe_limit)
10864                                 newval = cur_rttp + (fval - bbr_rtt_probe_limit);
10865                         else
10866                                 newval = cur_rttp;
10867                 } else {
10868                         int mul;
10869
10870                         mul = fval / bbr_rtt_probe_limit;
10871                         newval = cur_rttp * mul;
10872                 }
10873                 if (cur_rttp >  bbr->r_ctl.rc_probertt_int) {
10874                         bbr->r_ctl.rc_probertt_int = cur_rttp;
10875                         reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10876                         val = 1;
10877                 } else {
10878                         /*
10879                          * No adjustments were made
10880                          * do we need to shrink it?
10881                          */
10882                         if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) {
10883                                 if (cur_rttp <= bbr_rtt_probe_limit) {
10884                                         /*
10885                                          * Things have calmed down lets
10886                                          * shrink all the way to default
10887                                          */
10888                                         bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit;
10889                                         reset_time_small(&bbr->r_ctl.rc_rttprop,
10890                                                          (bbr_filter_len_sec * USECS_IN_SECOND));
10891                                         cur_rttp = bbr_rtt_probe_limit;
10892                                         newval = (bbr_filter_len_sec * USECS_IN_SECOND);
10893                                         val = 2;
10894                                 } else {
10895                                         /*
10896                                          * Well does some adjustment make sense?
10897                                          */
10898                                         if (cur_rttp < bbr->r_ctl.rc_probertt_int) {
10899                                                 /* We can reduce interval time some */
10900                                                 bbr->r_ctl.rc_probertt_int = cur_rttp;
10901                                                 reset_time_small(&bbr->r_ctl.rc_rttprop, newval);
10902                                                 val = 3;
10903                                         }
10904                                 }
10905                         }
10906                 }
10907                 if (val)
10908                         bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val);
10909         }
10910 }
10911
10912 static void
10913 bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts)
10914 {
10915         /* Exit probe-rtt */
10916
10917         if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) {
10918                 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
10919                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
10920         }
10921         bbr_log_exit_gain(bbr, cts, 1);
10922         bbr->rc_hit_state_1 = 0;
10923         bbr->r_ctl.rc_rtt_shrinks = cts;
10924         bbr->r_ctl.last_in_probertt = cts;
10925         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0);
10926         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
10927         bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp,
10928                                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
10929                                           bbr->r_ctl.rc_delivered);
10930         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
10931                 uint32_t time_in;
10932
10933                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
10934                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
10935         }
10936         if (bbr->rc_filled_pipe) {
10937                 /* Switch to probe_bw */
10938                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
10939                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
10940                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain;
10941                 bbr_substate_change(bbr, cts, __LINE__, 0);
10942                 bbr_log_type_statechange(bbr, cts, __LINE__);
10943         } else {
10944                 /* Back to startup */
10945                 bbr->rc_bbr_state = BBR_STATE_STARTUP;
10946                 bbr->r_ctl.rc_bbr_state_time = cts;
10947                 /*
10948                  * We don't want to give a complete free 3
10949                  * measurements until we exit, so we use
10950                  * the number of pe's we were in probe-rtt
10951                  * to add to the startup_epoch. That way
10952                  * we will still retain the old state.
10953                  */
10954                 bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt);
10955                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
10956                 /* Make sure to use the lower pg when shifting back in */
10957                 if (bbr->r_ctl.rc_lost &&
10958                     bbr_use_lower_gain_in_startup &&
10959                     (bbr->rc_use_google == 0))
10960                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
10961                 else
10962                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg;
10963                 bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg;
10964                 /* Probably not needed but set it anyway */
10965                 bbr_set_state_target(bbr, __LINE__);
10966                 bbr_log_type_statechange(bbr, cts, __LINE__);
10967                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
10968                     bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0);
10969         }
10970         bbr_check_probe_rtt_limits(bbr, cts);
10971 }
10972
10973 static int32_t inline
10974 bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts)
10975 {
10976         if ((bbr->rc_past_init_win == 1) &&
10977             (bbr->rc_in_persist == 0) &&
10978             (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) {
10979                 return (1);
10980         }
10981         if (bbr_can_force_probertt &&
10982             (bbr->rc_in_persist == 0) &&
10983             (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) &&
10984             ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) {
10985                 return (1);
10986         }
10987         return (0);
10988 }
10989
10990 static int32_t
10991 bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t  pkt_epoch)
10992 {
10993         uint64_t btlbw, gain;
10994         if (pkt_epoch == 0) {
10995                 /*
10996                  * Need to be on a pkt-epoch to continue.
10997                  */
10998                 return (0);
10999         }
11000         btlbw = bbr_get_full_bw(bbr);
11001         gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11002                  (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11003         if (btlbw >= gain) {
11004                 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11005                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11006                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11007                 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11008         }
11009         if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)
11010                 return (1);
11011         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11012                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8);
11013         return(0);
11014 }
11015
11016 static int32_t inline
11017 bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch)
11018 {
11019         /* Have we gained 25% in the last 3 packet based epoch's? */
11020         uint64_t btlbw, gain;
11021         int do_exit;
11022         int delta, rtt_gain;
11023
11024         if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11025             (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11026                 /*
11027                  * This qualifies as a RTT_PROBE session since we drop the
11028                  * data outstanding to nothing and waited more than
11029                  * bbr_rtt_probe_time.
11030                  */
11031                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11032                 bbr_set_reduced_rtt(bbr, cts, __LINE__);
11033         }
11034         if (bbr_should_enter_probe_rtt(bbr, cts)) {
11035                 bbr_enter_probe_rtt(bbr, cts, __LINE__);
11036                 return (0);
11037         }
11038         if (bbr->rc_use_google)
11039                 return (bbr_google_startup(bbr, cts,  pkt_epoch));
11040
11041         if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11042             (bbr_use_lower_gain_in_startup)) {
11043                 /* Drop to a lower gain 1.5 x since we saw loss */
11044                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower;
11045         }
11046         if (pkt_epoch == 0) {
11047                 /*
11048                  * Need to be on a pkt-epoch to continue.
11049                  */
11050                 return (0);
11051         }
11052         if (bbr_rtt_gain_thresh) {
11053                 /*
11054                  * Do we allow a flow to stay
11055                  * in startup with no loss and no
11056                  * gain in rtt over a set threshold?
11057                  */
11058                 if (bbr->r_ctl.rc_pkt_epoch_rtt &&
11059                     bbr->r_ctl.startup_last_srtt &&
11060                     (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) {
11061                         delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt;
11062                         rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt;
11063                 } else
11064                         rtt_gain = 0;
11065                 if ((bbr->r_ctl.startup_last_srtt == 0)  ||
11066                     (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt))
11067                         /* First time or new lower value */
11068                         bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt;
11069
11070                 if ((bbr->r_ctl.rc_lost == 0) &&
11071                     (rtt_gain < bbr_rtt_gain_thresh)) {
11072                         /*
11073                          * No loss, and we are under
11074                          * our gain threhold for
11075                          * increasing RTT.
11076                          */
11077                         if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11078                                 bbr->r_ctl.rc_bbr_last_startup_epoch++;
11079                         bbr_log_startup_event(bbr, cts, rtt_gain,
11080                                               delta, bbr->r_ctl.startup_last_srtt, 10);
11081                         return (0);
11082                 }
11083         }
11084         if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) &&
11085             (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) &&
11086             (!IN_RECOVERY(bbr->rc_tp->t_flags))) {
11087                 /*
11088                  * We only assess if we have a new measurment when
11089                  * we have no loss and are not in recovery.
11090                  * Drag up by one our last_startup epoch so we will hold
11091                  * the number of non-gain we have already accumulated.
11092                  */
11093                 if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch)
11094                         bbr->r_ctl.rc_bbr_last_startup_epoch++;
11095                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11096                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9);
11097                 return (0);
11098         }
11099         /* Case where we reduced the lost (bad retransmit) */
11100         if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost)
11101                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11102         bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count;
11103         btlbw = bbr_get_full_bw(bbr);
11104         if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower)
11105                 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11106                          (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11107         else
11108                 gain = ((bbr->r_ctl.rc_bbr_lastbtlbw *
11109                          (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw;
11110         do_exit = 0;
11111         if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw)
11112                 bbr->r_ctl.rc_bbr_lastbtlbw = btlbw;
11113         if (btlbw >= gain) {
11114                 bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch;
11115                 /* Update the lost so we won't exit in next set of tests */
11116                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11117                 bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11118                                       bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3);
11119         }
11120         if ((bbr->rc_loss_exit &&
11121              (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) &&
11122              (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) &&
11123             ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) {
11124                 /*
11125                  * If we had no gain,  we had loss and that loss was above
11126                  * our threshould, the rwnd is not constrained, and we have
11127                  * had at least 3 packet epochs exit. Note that this is
11128                  * switched off by sysctl. Google does not do this by the
11129                  * way.
11130                  */
11131                 if ((ctf_flight_size(bbr->rc_tp,
11132                          (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) +
11133                      (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) {
11134                         do_exit = 1;
11135                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11136                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4);
11137                 } else {
11138                         /* Just record an updated loss value */
11139                         bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11140                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11141                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5);
11142                 }
11143         } else
11144                 bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost;
11145         if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) ||
11146             do_exit) {
11147                 /* Return 1 to exit the startup state. */
11148                 return (1);
11149         }
11150         /* Stay in startup */
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, 8);
11153         return (0);
11154 }
11155
11156 static void
11157 bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses)
11158 {
11159         /*
11160          * A tick occured in the rtt epoch do we need to do anything?
11161          */
11162 #ifdef BBR_INVARIANTS
11163         if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
11164             (bbr->rc_bbr_state != BBR_STATE_DRAIN) &&
11165             (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) &&
11166             (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
11167             (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) {
11168                 /* Debug code? */
11169                 panic("Unknown BBR state %d?\n", bbr->rc_bbr_state);
11170         }
11171 #endif
11172         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
11173                 /* Do we exit the startup state? */
11174                 if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) {
11175                         uint32_t time_in;
11176
11177                         bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch,
11178                                               bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6);
11179                         bbr->rc_filled_pipe = 1;
11180                         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11181                         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11182                                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11183                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11184                         } else
11185                                 time_in = 0;
11186                         if (bbr->rc_no_pacing)
11187                                 bbr->rc_no_pacing = 0;
11188                         bbr->r_ctl.rc_bbr_state_time = cts;
11189                         bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg;
11190                         bbr->rc_bbr_state = BBR_STATE_DRAIN;
11191                         bbr_set_state_target(bbr, __LINE__);
11192                         if ((bbr->rc_use_google == 0) &&
11193                             bbr_slam_cwnd_in_main_drain) {
11194                                 /* Here we don't have to worry about probe-rtt */
11195                                 bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd;
11196                                 bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11197                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11198                         }
11199                         bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain;
11200                         bbr_log_type_statechange(bbr, cts, __LINE__);
11201                         if (ctf_flight_size(bbr->rc_tp,
11202                                 (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <=
11203                             bbr->r_ctl.rc_target_at_state) {
11204                                 /*
11205                                  * Switch to probe_bw if we are already
11206                                  * there
11207                                  */
11208                                 bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11209                                 bbr_substate_change(bbr, cts, __LINE__, 0);
11210                                 bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11211                                 bbr_log_type_statechange(bbr, cts, __LINE__);
11212                         }
11213                 }
11214         } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) {
11215                 uint32_t inflight;
11216                 struct tcpcb *tp;
11217
11218                 tp = bbr->rc_tp;
11219                 inflight = ctf_flight_size(tp,
11220                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11221                 if (inflight >= bbr->r_ctl.rc_target_at_state) {
11222                         /* We have reached a flight of the cwnd target */
11223                         bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11224                         bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11225                         bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT;
11226                         bbr_set_state_target(bbr, __LINE__);
11227                         /*
11228                          * Rig it so we don't do anything crazy and
11229                          * start fresh with a new randomization.
11230                          */
11231                         bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff;
11232                         bbr->rc_bbr_substate = BBR_SUB_LEVEL6;
11233                         bbr_substate_change(bbr, cts, __LINE__, 1);
11234                 }
11235         } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) {
11236                 /* Has in-flight reached the bdp (or less)? */
11237                 uint32_t inflight;
11238                 struct tcpcb *tp;
11239
11240                 tp = bbr->rc_tp;
11241                 inflight = ctf_flight_size(tp,
11242                               (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11243                 if ((bbr->rc_use_google == 0) &&
11244                     bbr_slam_cwnd_in_main_drain &&
11245                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11246                         /*
11247                          * Here we don't have to worry about probe-rtt
11248                          * re-slam it, but keep it slammed down.
11249                          */
11250                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11251                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11252                 }
11253                 if (inflight <= bbr->r_ctl.rc_target_at_state) {
11254                         /* We have drained */
11255                         bbr->rc_bbr_state = BBR_STATE_PROBE_BW;
11256                         bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost;
11257                         if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) {
11258                                 uint32_t time_in;
11259
11260                                 time_in = cts - bbr->r_ctl.rc_bbr_state_time;
11261                                 counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in);
11262                         }
11263                         if ((bbr->rc_use_google == 0) &&
11264                             bbr_slam_cwnd_in_main_drain &&
11265                             (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) {
11266                                 /* Restore the cwnd */
11267                                 tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd;
11268                                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11269                         }
11270                         /* Setup probe-rtt has being done now RRS-HERE */
11271                         bbr->r_ctl.rc_rtt_shrinks = cts;
11272                         bbr->r_ctl.last_in_probertt = cts;
11273                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0);
11274                         /* Randomly pick a sub-state */
11275                         bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts);
11276                         bbr_substate_change(bbr, cts, __LINE__, 0);
11277                         bbr_log_type_statechange(bbr, cts, __LINE__);
11278                 }
11279         } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) {
11280                 uint32_t flight;
11281
11282                 flight = ctf_flight_size(bbr->rc_tp,
11283                              (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11284                 bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered);
11285                 if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) &&
11286                     (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11287                         /*
11288                          * We must keep cwnd at the desired MSS.
11289                          */
11290                         bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options);
11291                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11292                 } else if ((bbr_prtt_slam_cwnd) &&
11293                            (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) {
11294                         /* Re-slam it */
11295                         bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state;
11296                         bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__);
11297                 }
11298                 if (bbr->r_ctl.rc_bbr_enters_probertt == 0) {
11299                         /* Has outstanding reached our target? */
11300                         if (flight <= bbr->r_ctl.rc_target_at_state) {
11301                                 bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0);
11302                                 bbr->r_ctl.rc_bbr_enters_probertt = cts;
11303                                 /* If time is exactly 0, be 1usec off */
11304                                 if (bbr->r_ctl.rc_bbr_enters_probertt == 0)
11305                                         bbr->r_ctl.rc_bbr_enters_probertt = 1;
11306                                 if (bbr->rc_use_google == 0) {
11307                                         /*
11308                                          * Restore any lowering that as occured to
11309                                          * reach here
11310                                          */
11311                                         if (bbr->r_ctl.bbr_rttprobe_gain_val)
11312                                                 bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val;
11313                                         else
11314                                                 bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT;
11315                                 }
11316                         }
11317                         if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) &&
11318                             (bbr->rc_use_google == 0) &&
11319                             bbr->r_ctl.bbr_rttprobe_gain_val &&
11320                             (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) ||
11321                              (flight >= bbr->r_ctl.flightsize_at_drain))) {
11322                                 /*
11323                                  * We have doddled with our current hptsi
11324                                  * gain an srtt and have still not made it
11325                                  * to target, or we have increased our flight.
11326                                  * Lets reduce the gain by xx%
11327                                  * flooring the reduce at DRAIN (based on
11328                                  * mul/div)
11329                                  */
11330                                 int red;
11331
11332                                 bbr->r_ctl.flightsize_at_drain = flight;
11333                                 bbr->r_ctl.rc_probertt_srttchktim = cts;
11334                                 red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1);
11335                                 if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) {
11336                                         /* Reduce our gain again */
11337                                         bbr->r_ctl.rc_bbr_hptsi_gain -= red;
11338                                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0);
11339                                 } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) {
11340                                         /* one more chance before we give up */
11341                                         bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1);
11342                                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0);
11343                                 } else {
11344                                         /* At the very bottom */
11345                                         bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1);
11346                                 }
11347                         }
11348                 }
11349                 if (bbr->r_ctl.rc_bbr_enters_probertt &&
11350                     (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) &&
11351                     ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) {
11352                         /* Time to exit probe RTT normally */
11353                         bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts);
11354                 }
11355         } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) {
11356                 if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) &&
11357                     (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
11358                         /*
11359                          * This qualifies as a RTT_PROBE session since we
11360                          * drop the data outstanding to nothing and waited
11361                          * more than bbr_rtt_probe_time.
11362                          */
11363                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
11364                         bbr_set_reduced_rtt(bbr, cts, __LINE__);
11365                 }
11366                 if (bbr_should_enter_probe_rtt(bbr, cts)) {
11367                         bbr_enter_probe_rtt(bbr, cts, __LINE__);
11368                 } else {
11369                         bbr_set_probebw_gains(bbr, cts, losses);
11370                 }
11371         }
11372 }
11373
11374 static void
11375 bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses)
11376 {
11377         int32_t epoch = 0;
11378
11379         if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) {
11380                 bbr_set_epoch(bbr, cts, line);
11381                 /* At each epoch doe lt bw sampling */
11382                 epoch = 1;
11383         }
11384         bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses);
11385 }
11386
11387 static int
11388 bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
11389     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
11390     int32_t nxt_pkt, struct timeval *tv)
11391 {
11392         int32_t thflags, retval;
11393         uint32_t cts, lcts;
11394         uint32_t tiwin;
11395         struct tcpopt to;
11396         struct tcp_bbr *bbr;
11397         struct bbr_sendmap *rsm;
11398         struct timeval ltv;
11399         int32_t did_out = 0;
11400         int32_t in_recovery;
11401         uint16_t nsegs;
11402         int32_t prev_state;
11403         uint32_t lost;
11404
11405         nsegs = max(1, m->m_pkthdr.lro_nsegs);
11406         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
11407         /* add in our stats */
11408         kern_prefetch(bbr, &prev_state);
11409         prev_state = 0;
11410         thflags = th->th_flags;
11411         /*
11412          * If this is either a state-changing packet or current state isn't
11413          * established, we require a write lock on tcbinfo.  Otherwise, we
11414          * allow the tcbinfo to be in either alocked or unlocked, as the
11415          * caller may have unnecessarily acquired a write lock due to a
11416          * race.
11417          */
11418         INP_WLOCK_ASSERT(tp->t_inpcb);
11419         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
11420             __func__));
11421         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
11422             __func__));
11423
11424         tp->t_rcvtime = ticks;
11425         /*
11426          * Unscale the window into a 32-bit value. For the SYN_SENT state
11427          * the scale is zero.
11428          */
11429         tiwin = th->th_win << tp->snd_scale;
11430 #ifdef STATS
11431         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
11432 #endif
11433
11434         if (m->m_flags & M_TSTMP) {
11435                 /* Prefer the hardware timestamp if present */
11436                 struct timespec ts;
11437
11438                 mbuf_tstmp2timespec(m, &ts);
11439                 bbr->rc_tv.tv_sec = ts.tv_sec;
11440                 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11441                 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11442         } else if (m->m_flags & M_TSTMP_LRO) {
11443                 /* Next the arrival timestamp */
11444                 struct timespec ts;
11445
11446                 mbuf_tstmp2timespec(m, &ts);
11447                 bbr->rc_tv.tv_sec = ts.tv_sec;
11448                 bbr->rc_tv.tv_usec = ts.tv_nsec / 1000;
11449                 bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv);
11450         } else {
11451                 /*
11452                  * Ok just get the current time.
11453                  */
11454                 bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv);
11455         }
11456         /*
11457          * Parse options on any incoming segment.
11458          */
11459         tcp_dooptions(&to, (u_char *)(th + 1),
11460             (th->th_off << 2) - sizeof(struct tcphdr),
11461             (thflags & TH_SYN) ? TO_SYN : 0);
11462
11463         /*
11464          * If timestamps were negotiated during SYN/ACK and a
11465          * segment without a timestamp is received, silently drop
11466          * the segment.
11467          * See section 3.2 of RFC 7323.
11468          */
11469         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS)) {
11470                 retval = 0;
11471                 goto done_with_input;
11472         }
11473         /*
11474          * If echoed timestamp is later than the current time, fall back to
11475          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
11476          * were used when this connection was established.
11477          */
11478         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
11479                 to.to_tsecr -= tp->ts_offset;
11480                 if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv)))
11481                         to.to_tsecr = 0;
11482         }
11483         /*
11484          * If its the first time in we need to take care of options and
11485          * verify we can do SACK for rack!
11486          */
11487         if (bbr->r_state == 0) {
11488                 /*
11489                  * Process options only when we get SYN/ACK back. The SYN
11490                  * case for incoming connections is handled in tcp_syncache.
11491                  * According to RFC1323 the window field in a SYN (i.e., a
11492                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
11493                  * this is traditional behavior, may need to be cleaned up.
11494                  */
11495                 if (bbr->rc_inp == NULL) {
11496                         bbr->rc_inp = tp->t_inpcb;
11497                 }
11498                 /*
11499                  * We need to init rc_inp here since its not init'd when
11500                  * bbr_init is called
11501                  */
11502                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
11503                         if ((to.to_flags & TOF_SCALE) &&
11504                             (tp->t_flags & TF_REQ_SCALE)) {
11505                                 tp->t_flags |= TF_RCVD_SCALE;
11506                                 tp->snd_scale = to.to_wscale;
11507                         } else
11508                                 tp->t_flags &= ~TF_REQ_SCALE;
11509                         /*
11510                          * Initial send window.  It will be updated with the
11511                          * next incoming segment to the scaled value.
11512                          */
11513                         tp->snd_wnd = th->th_win;
11514                         if ((to.to_flags & TOF_TS) &&
11515                             (tp->t_flags & TF_REQ_TSTMP)) {
11516                                 tp->t_flags |= TF_RCVD_TSTMP;
11517                                 tp->ts_recent = to.to_tsval;
11518                                 tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv);
11519                         } else
11520                             tp->t_flags &= ~TF_REQ_TSTMP;
11521                         if (to.to_flags & TOF_MSS)
11522                                 tcp_mss(tp, to.to_mss);
11523                         if ((tp->t_flags & TF_SACK_PERMIT) &&
11524                             (to.to_flags & TOF_SACKPERM) == 0)
11525                                 tp->t_flags &= ~TF_SACK_PERMIT;
11526                         if (IS_FASTOPEN(tp->t_flags)) {
11527                                 if (to.to_flags & TOF_FASTOPEN) {
11528                                         uint16_t mss;
11529
11530                                         if (to.to_flags & TOF_MSS)
11531                                                 mss = to.to_mss;
11532                                         else
11533                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
11534                                                         mss = TCP6_MSS;
11535                                                 else
11536                                                         mss = TCP_MSS;
11537                                         tcp_fastopen_update_cache(tp, mss,
11538                                             to.to_tfo_len, to.to_tfo_cookie);
11539                                 } else
11540                                         tcp_fastopen_disable_path(tp);
11541                         }
11542                 }
11543                 /*
11544                  * At this point we are at the initial call. Here we decide
11545                  * if we are doing RACK or not. We do this by seeing if
11546                  * TF_SACK_PERMIT is set, if not rack is *not* possible and
11547                  * we switch to the default code.
11548                  */
11549                 if ((tp->t_flags & TF_SACK_PERMIT) == 0) {
11550                         /* Bail */
11551                         tcp_switch_back_to_default(tp);
11552                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
11553                             tlen, iptos);
11554                         return (1);
11555                 }
11556                 /* Set the flag */
11557                 bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
11558                 tcp_set_hpts(tp->t_inpcb);
11559                 sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack);
11560         }
11561         if (thflags & TH_ACK) {
11562                 /* Track ack types */
11563                 if (to.to_flags & TOF_SACK)
11564                         BBR_STAT_INC(bbr_acks_with_sacks);
11565                 else
11566                         BBR_STAT_INC(bbr_plain_acks);
11567         }
11568         /*
11569          * This is the one exception case where we set the rack state
11570          * always. All other times (timers etc) we must have a rack-state
11571          * set (so we assure we have done the checks above for SACK).
11572          */
11573         if (thflags & TH_FIN)
11574                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
11575         if (bbr->r_state != tp->t_state)
11576                 bbr_set_state(tp, bbr, tiwin);
11577
11578         if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL)
11579                 kern_prefetch(rsm, &prev_state);
11580         prev_state = bbr->r_state;
11581         bbr->rc_ack_was_delayed = 0;
11582         lost = bbr->r_ctl.rc_lost;
11583         bbr->rc_is_pkt_epoch_now = 0;
11584         if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) {
11585                 /* Get the real time into lcts and figure the real delay */
11586                 lcts = tcp_get_usecs(&ltv);
11587                 if (TSTMP_GT(lcts, cts)) {
11588                         bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts;
11589                         bbr->rc_ack_was_delayed = 1;
11590                         if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay,
11591                                      bbr->r_ctl.highest_hdwr_delay))
11592                                 bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay;
11593                 } else {
11594                         bbr->r_ctl.rc_ack_hdwr_delay = 0;
11595                         bbr->rc_ack_was_delayed = 0;
11596                 }
11597         } else {
11598                 bbr->r_ctl.rc_ack_hdwr_delay = 0;
11599                 bbr->rc_ack_was_delayed = 0;
11600         }
11601         bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m);
11602         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
11603                 retval = 0;
11604                 m_freem(m);
11605                 goto done_with_input;
11606         }
11607         /*
11608          * If a segment with the ACK-bit set arrives in the SYN-SENT state
11609          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
11610          */
11611         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
11612             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
11613                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11614                 ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11615                 return (1);
11616         }
11617         in_recovery = IN_RECOVERY(tp->t_flags);
11618         if (tiwin > bbr->r_ctl.rc_high_rwnd)
11619                 bbr->r_ctl.rc_high_rwnd = tiwin;
11620 #ifdef BBR_INVARIANTS
11621         if ((tp->t_inpcb->inp_flags & INP_DROPPED) ||
11622             (tp->t_inpcb->inp_flags2 & INP_FREED)) {
11623                 panic("tp:%p bbr:%p given a dropped inp:%p",
11624                     tp, bbr, tp->t_inpcb);
11625         }
11626 #endif
11627         bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp,
11628                                             (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11629         bbr->rtt_valid = 0;
11630         if (to.to_flags & TOF_TS) {
11631                 bbr->rc_ts_valid = 1;
11632                 bbr->r_ctl.last_inbound_ts = to.to_tsval;
11633         } else {
11634                 bbr->rc_ts_valid = 0;
11635                 bbr->r_ctl.last_inbound_ts = 0;
11636         }
11637         retval = (*bbr->r_substate) (m, th, so,
11638             tp, &to, drop_hdrlen,
11639             tlen, tiwin, thflags, nxt_pkt, iptos);
11640 #ifdef BBR_INVARIANTS
11641         if ((retval == 0) &&
11642             (tp->t_inpcb == NULL)) {
11643                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
11644                     retval, tp, prev_state);
11645         }
11646 #endif
11647         if (nxt_pkt == 0)
11648                 BBR_STAT_INC(bbr_rlock_left_ret0);
11649         else
11650                 BBR_STAT_INC(bbr_rlock_left_ret1);
11651         if (retval == 0) {
11652                 /*
11653                  * If retval is 1 the tcb is unlocked and most likely the tp
11654                  * is gone.
11655                  */
11656                 INP_WLOCK_ASSERT(tp->t_inpcb);
11657                 tcp_bbr_xmit_timer_commit(bbr, tp, cts);
11658                 if (bbr->rc_is_pkt_epoch_now)
11659                         bbr_set_pktepoch(bbr, cts, __LINE__);
11660                 bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost));
11661                 if (nxt_pkt == 0) {
11662                         if (bbr->r_wanted_output != 0) {
11663                                 bbr->rc_output_starts_timer = 0;
11664                                 did_out = 1;
11665                                 (void)tp->t_fb->tfb_tcp_output(tp);
11666                         } else
11667                                 bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0);
11668                 }
11669                 if ((nxt_pkt == 0) &&
11670                     ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
11671                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
11672                      (tp->t_flags & TF_DELACK) ||
11673                      ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
11674                       (tp->t_state <= TCPS_CLOSING)))) {
11675                         /*
11676                          * We could not send (probably in the hpts but
11677                          * stopped the timer)?
11678                          */
11679                         if ((tp->snd_max == tp->snd_una) &&
11680                             ((tp->t_flags & TF_DELACK) == 0) &&
11681                             (bbr->rc_inp->inp_in_hpts) &&
11682                             (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11683                                 /*
11684                                  * keep alive not needed if we are hptsi
11685                                  * output yet
11686                                  */
11687                                 ;
11688                         } else {
11689                                 if (bbr->rc_inp->inp_in_hpts) {
11690                                         tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT);
11691                                         if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
11692                                             (TSTMP_GT(lcts, bbr->rc_pacer_started))) {
11693                                                 uint32_t del;
11694
11695                                                 del = lcts - bbr->rc_pacer_started;
11696                                                 if (bbr->r_ctl.rc_last_delay_val > del) {
11697                                                         BBR_STAT_INC(bbr_force_timer_start);
11698                                                         bbr->r_ctl.rc_last_delay_val -= del;
11699                                                         bbr->rc_pacer_started = lcts;
11700                                                 } else {
11701                                                         /* We are late */
11702                                                         bbr->r_ctl.rc_last_delay_val = 0;
11703                                                         BBR_STAT_INC(bbr_force_output);
11704                                                         (void)tp->t_fb->tfb_tcp_output(tp);
11705                                                 }
11706                                         }
11707                                 }
11708                                 bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val,
11709                                     0);
11710                         }
11711                 } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) {
11712                         /* Do we have the correct timer running? */
11713                         bbr_timer_audit(tp, bbr, lcts, &so->so_snd);
11714                 }
11715                 /* Do we have a new state */
11716                 if (bbr->r_state != tp->t_state)
11717                         bbr_set_state(tp, bbr, tiwin);
11718 done_with_input:
11719                 bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out);
11720                 if (did_out)
11721                         bbr->r_wanted_output = 0;
11722 #ifdef BBR_INVARIANTS
11723                 if (tp->t_inpcb == NULL) {
11724                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
11725                             did_out,
11726                             retval, tp, prev_state);
11727                 }
11728 #endif
11729         }
11730         return (retval);
11731 }
11732
11733 static void
11734 bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
11735     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
11736 {
11737         struct timeval tv;
11738         int retval;
11739
11740         /* First lets see if we have old packets */
11741         if (tp->t_in_pkt) {
11742                 if (ctf_do_queued_segments(so, tp, 1)) {
11743                         m_freem(m);
11744                         return;
11745                 }
11746         }
11747         if (m->m_flags & M_TSTMP_LRO) {
11748                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
11749                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
11750         } else {
11751                 /* Should not be should we kassert instead? */
11752                 tcp_get_usecs(&tv);
11753         }
11754         retval = bbr_do_segment_nounlock(m, th, so, tp,
11755                                          drop_hdrlen, tlen, iptos, 0, &tv);
11756         if (retval == 0) {
11757                 tcp_handle_wakeup(tp, so);
11758                 INP_WUNLOCK(tp->t_inpcb);
11759         }
11760 }
11761
11762 /*
11763  * Return how much data can be sent without violating the
11764  * cwnd or rwnd.
11765  */
11766
11767 static inline uint32_t
11768 bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin,
11769     uint32_t avail, int32_t sb_offset, uint32_t cts)
11770 {
11771         uint32_t len;
11772
11773         if (ctf_outstanding(tp) >= tp->snd_wnd) {
11774                 /* We never want to go over our peers rcv-window */
11775                 len = 0;
11776         } else {
11777                 uint32_t flight;
11778
11779                 flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes));
11780                 if (flight >= sendwin) {
11781                         /*
11782                          * We have in flight what we are allowed by cwnd (if
11783                          * it was rwnd blocking it would have hit above out
11784                          * >= tp->snd_wnd).
11785                          */
11786                         return (0);
11787                 }
11788                 len = sendwin - flight;
11789                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
11790                         /* We would send too much (beyond the rwnd) */
11791                         len = tp->snd_wnd - ctf_outstanding(tp);
11792                 }
11793                 if ((len + sb_offset) > avail) {
11794                         /*
11795                          * We don't have that much in the SB, how much is
11796                          * there?
11797                          */
11798                         len = avail - sb_offset;
11799                 }
11800         }
11801         return (len);
11802 }
11803
11804 static inline void
11805 bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11806 {
11807 #ifdef NETFLIX_STATS
11808         KMOD_TCPSTAT_INC(tcps_sndpack_error);
11809         KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len);
11810 #endif
11811 }
11812
11813 static inline void
11814 bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error)
11815 {
11816         if (error) {
11817                 bbr_do_error_accounting(tp, bbr, rsm, len, error);
11818                 return;
11819         }
11820         if (rsm) {
11821                 if (rsm->r_flags & BBR_TLP) {
11822                         /*
11823                          * TLP should not count in retran count, but in its
11824                          * own bin
11825                          */
11826 #ifdef NETFLIX_STATS
11827                         tp->t_sndtlppack++;
11828                         tp->t_sndtlpbyte += len;
11829                         KMOD_TCPSTAT_INC(tcps_tlpresends);
11830                         KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len);
11831 #endif
11832                 } else {
11833                         /* Retransmit */
11834                         tp->t_sndrexmitpack++;
11835                         KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
11836                         KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
11837 #ifdef STATS
11838                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
11839                             len);
11840 #endif
11841                 }
11842                 /*
11843                  * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is
11844                  * sub-state
11845                  */
11846                 counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len);
11847                 if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) {
11848                         /* Non probe_bw log in 1, 2, or 4. */
11849                         counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len);
11850                 } else {
11851                         /*
11852                          * Log our probe state 3, and log also 5-13 to show
11853                          * us the recovery sub-state for the send. This
11854                          * means that 3 == (5+6+7+8+9+10+11+12+13)
11855                          */
11856                         counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len);
11857                         counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len);
11858                 }
11859                 /* Place in both 16's the totals of retransmitted */
11860                 counter_u64_add(bbr_state_lost[16], len);
11861                 counter_u64_add(bbr_state_resend[16], len);
11862                 /* Place in 17's the total sent */
11863                 counter_u64_add(bbr_state_resend[17], len);
11864                 counter_u64_add(bbr_state_lost[17], len);
11865
11866         } else {
11867                 /* New sends */
11868                 KMOD_TCPSTAT_INC(tcps_sndpack);
11869                 KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
11870                 /* Place in 17's the total sent */
11871                 counter_u64_add(bbr_state_resend[17], len);
11872                 counter_u64_add(bbr_state_lost[17], len);
11873 #ifdef STATS
11874                 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
11875                     len);
11876 #endif
11877         }
11878 }
11879
11880 static void
11881 bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level)
11882 {
11883         if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) {
11884                 /*
11885                  * Limit the cwnd to not be above N x the target plus whats
11886                  * is outstanding. The target is based on the current b/w
11887                  * estimate.
11888                  */
11889                 uint32_t target;
11890
11891                 target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT);
11892                 target += ctf_outstanding(tp);
11893                 target *= bbr_target_cwnd_mult_limit;
11894                 if (tp->snd_cwnd > target)
11895                         tp->snd_cwnd = target;
11896                 bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__);
11897         }
11898 }
11899
11900 static int
11901 bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg)
11902 {
11903         /*
11904          * "adv" is the amount we could increase the window, taking into
11905          * account that we are limited by TCP_MAXWIN << tp->rcv_scale.
11906          */
11907         int32_t adv;
11908         int32_t oldwin;
11909
11910         adv = recwin;
11911         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
11912                 oldwin = (tp->rcv_adv - tp->rcv_nxt);
11913                 if (adv > oldwin)
11914                         adv -= oldwin;
11915                 else {
11916                         /* We can't increase the window */
11917                         adv = 0;
11918                 }
11919         } else
11920                 oldwin = 0;
11921
11922         /*
11923          * If the new window size ends up being the same as or less
11924          * than the old size when it is scaled, then don't force
11925          * a window update.
11926          */
11927         if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
11928                 return (0);
11929
11930         if (adv >= (2 * maxseg) &&
11931             (adv >= (so->so_rcv.sb_hiwat / 4) ||
11932             recwin <= (so->so_rcv.sb_hiwat / 8) ||
11933             so->so_rcv.sb_hiwat <= 8 * maxseg)) {
11934                 return (1);
11935         }
11936         if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat)
11937                 return (1);
11938         return (0);
11939 }
11940
11941 /*
11942  * Return 0 on success and a errno on failure to send.
11943  * Note that a 0 return may not mean we sent anything
11944  * if the TCB was on the hpts. A non-zero return
11945  * does indicate the error we got from ip[6]_output.
11946  */
11947 static int
11948 bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv)
11949 {
11950         struct socket *so;
11951         int32_t len;
11952         uint32_t cts;
11953         uint32_t recwin, sendwin;
11954         int32_t sb_offset;
11955         int32_t flags, abandon, error = 0;
11956         struct tcp_log_buffer *lgb = NULL;
11957         struct mbuf *m;
11958         struct mbuf *mb;
11959         uint32_t if_hw_tsomaxsegcount = 0;
11960         uint32_t if_hw_tsomaxsegsize = 0;
11961         uint32_t if_hw_tsomax = 0;
11962         struct ip *ip = NULL;
11963 #ifdef TCPDEBUG
11964         struct ipovly *ipov = NULL;
11965 #endif
11966         struct tcp_bbr *bbr;
11967         struct tcphdr *th;
11968 #ifdef NETFLIX_TCPOUDP
11969         struct udphdr *udp = NULL;
11970 #endif
11971         u_char opt[TCP_MAXOLEN];
11972         unsigned ipoptlen, optlen, hdrlen;
11973 #ifdef NETFLIX_TCPOUDP
11974         unsigned ulen;
11975 #endif
11976         uint32_t bbr_seq;
11977         uint32_t delay_calc=0;
11978         uint8_t doing_tlp = 0;
11979         uint8_t local_options;
11980 #ifdef BBR_INVARIANTS
11981         uint8_t doing_retran_from = 0;
11982         uint8_t picked_up_retran = 0;
11983 #endif
11984         uint8_t wanted_cookie = 0;
11985         uint8_t more_to_rxt=0;
11986         int32_t prefetch_so_done = 0;
11987         int32_t prefetch_rsm = 0;
11988         uint32_t what_we_can = 0;
11989         uint32_t tot_len = 0;
11990         uint32_t rtr_cnt = 0;
11991         uint32_t maxseg, pace_max_segs, p_maxseg;
11992         int32_t csum_flags;
11993         int32_t hw_tls;
11994 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
11995         unsigned ipsec_optlen = 0;
11996
11997 #endif
11998         volatile int32_t sack_rxmit;
11999         struct bbr_sendmap *rsm = NULL;
12000         int32_t tso, mtu;
12001         struct tcpopt to;
12002         int32_t slot = 0;
12003         struct inpcb *inp;
12004         struct sockbuf *sb;
12005         uint32_t hpts_calling;
12006 #ifdef INET6
12007         struct ip6_hdr *ip6 = NULL;
12008         int32_t isipv6;
12009 #endif
12010         uint8_t app_limited = BBR_JR_SENT_DATA;
12011         uint8_t filled_all = 0;
12012         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
12013         /* We take a cache hit here */
12014         memcpy(&bbr->rc_tv, tv, sizeof(struct timeval));
12015         cts = tcp_tv_to_usectick(&bbr->rc_tv);
12016         inp = bbr->rc_inp;
12017         so = inp->inp_socket;
12018         sb = &so->so_snd;
12019         if (sb->sb_flags & SB_TLS_IFNET)
12020                 hw_tls = 1;
12021         else
12022                 hw_tls = 0;
12023         kern_prefetch(sb, &maxseg);
12024         maxseg = tp->t_maxseg - bbr->rc_last_options;
12025         if (bbr_minseg(bbr) < maxseg) {
12026                 tcp_bbr_tso_size_check(bbr, cts);
12027         }
12028         /* Remove any flags that indicate we are pacing on the inp  */
12029         pace_max_segs = bbr->r_ctl.rc_pace_max_segs;
12030         p_maxseg = min(maxseg, pace_max_segs);
12031         INP_WLOCK_ASSERT(inp);
12032 #ifdef TCP_OFFLOAD
12033         if (tp->t_flags & TF_TOE)
12034                 return (tcp_offload_output(tp));
12035 #endif
12036
12037 #ifdef INET6
12038         if (bbr->r_state) {
12039                 /* Use the cache line loaded if possible */
12040                 isipv6 = bbr->r_is_v6;
12041         } else {
12042                 isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
12043         }
12044 #endif
12045         if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
12046             inp->inp_in_hpts) {
12047                 /*
12048                  * We are on the hpts for some timer but not hptsi output.
12049                  * Possibly remove from the hpts so we can send/recv etc.
12050                  */
12051                 if ((tp->t_flags & TF_ACKNOW) == 0) {
12052                         /*
12053                          * No immediate demand right now to send an ack, but
12054                          * the user may have read, making room for new data
12055                          * (a window update). If so we may want to cancel
12056                          * whatever timer is running (KEEP/DEL-ACK?) and
12057                          * continue to send out a window update. Or we may
12058                          * have gotten more data into the socket buffer to
12059                          * send.
12060                          */
12061                         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12062                                       (long)TCP_MAXWIN << tp->rcv_scale);
12063                         if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) &&
12064                             ((tcp_outflags[tp->t_state] & TH_RST) == 0) &&
12065                             ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <=
12066                             (tp->snd_max - tp->snd_una))) {
12067                                 /*
12068                                  * Nothing new to send and no window update
12069                                  * is needed to send. Lets just return and
12070                                  * let the timer-run off.
12071                                  */
12072                                 return (0);
12073                         }
12074                 }
12075                 tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12076                 bbr_timer_cancel(bbr, __LINE__, cts);
12077         }
12078         if (bbr->r_ctl.rc_last_delay_val) {
12079                 /* Calculate a rough delay for early escape to sending  */
12080                 if (SEQ_GT(cts, bbr->rc_pacer_started))
12081                         delay_calc = cts - bbr->rc_pacer_started;
12082                 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12083                         delay_calc -= bbr->r_ctl.rc_last_delay_val;
12084                 else
12085                         delay_calc = 0;
12086         }
12087         /* Mark that we have called bbr_output(). */
12088         if ((bbr->r_timer_override) ||
12089             (tp->t_state < TCPS_ESTABLISHED)) {
12090                 /* Timeouts or early states are exempt */
12091                 if (inp->inp_in_hpts)
12092                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12093         } else if (inp->inp_in_hpts) {
12094                 if ((bbr->r_ctl.rc_last_delay_val) &&
12095                     (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
12096                     delay_calc) {
12097                         /*
12098                          * We were being paced for output and the delay has
12099                          * already exceeded when we were supposed to be
12100                          * called, lets go ahead and pull out of the hpts
12101                          * and call output.
12102                          */
12103                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1);
12104                         bbr->r_ctl.rc_last_delay_val = 0;
12105                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12106                 } else if (tp->t_state == TCPS_CLOSED) {
12107                         bbr->r_ctl.rc_last_delay_val = 0;
12108                         tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT);
12109                 } else {
12110                         /*
12111                          * On the hpts, you shall not pass! even if ACKNOW
12112                          * is on, we will when the hpts fires, unless of
12113                          * course we are overdue.
12114                          */
12115                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1);
12116                         return (0);
12117                 }
12118         }
12119         bbr->rc_cwnd_limited = 0;
12120         if (bbr->r_ctl.rc_last_delay_val) {
12121                 /* recalculate the real delay and deal with over/under  */
12122                 if (SEQ_GT(cts, bbr->rc_pacer_started))
12123                         delay_calc = cts - bbr->rc_pacer_started;
12124                 else
12125                         delay_calc = 0;
12126                 if (delay_calc >= bbr->r_ctl.rc_last_delay_val)
12127                         /* Setup the delay which will be added in */
12128                         delay_calc -= bbr->r_ctl.rc_last_delay_val;
12129                 else {
12130                         /*
12131                          * We are early setup to adjust
12132                          * our slot time.
12133                          */
12134                         uint64_t merged_val;
12135
12136                         bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc);
12137                         bbr->r_agg_early_set = 1;
12138                         if (bbr->r_ctl.rc_hptsi_agg_delay) {
12139                                 if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) {
12140                                         /* Nope our previous late cancels out the early */
12141                                         bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early;
12142                                         bbr->r_agg_early_set = 0;
12143                                         bbr->r_ctl.rc_agg_early = 0;
12144                                 } else {
12145                                         bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay;
12146                                         bbr->r_ctl.rc_hptsi_agg_delay = 0;
12147                                 }
12148                         }
12149                         merged_val = bbr->rc_pacer_started;
12150                         merged_val <<= 32;
12151                         merged_val |= bbr->r_ctl.rc_last_delay_val;
12152                         bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls,
12153                                                  bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val,
12154                                                  bbr->r_agg_early_set, 3);
12155                         bbr->r_ctl.rc_last_delay_val = 0;
12156                         BBR_STAT_INC(bbr_early);
12157                         delay_calc = 0;
12158                 }
12159         } else {
12160                 /* We were not delayed due to hptsi */
12161                 if (bbr->r_agg_early_set)
12162                         bbr->r_ctl.rc_agg_early = 0;
12163                 bbr->r_agg_early_set = 0;
12164                 delay_calc = 0;
12165         }
12166         if (delay_calc) {
12167                 /*
12168                  * We had a hptsi delay which means we are falling behind on
12169                  * sending at the expected rate. Calculate an extra amount
12170                  * of data we can send, if any, to put us back on track.
12171                  */
12172                 if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay)
12173                         bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff;
12174                 else
12175                         bbr->r_ctl.rc_hptsi_agg_delay += delay_calc;
12176         }
12177         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12178         if ((tp->snd_una == tp->snd_max) &&
12179             (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) &&
12180             (sbavail(sb))) {
12181                 /*
12182                  * Ok we have been idle with nothing outstanding
12183                  * we possibly need to start fresh with either a new
12184                  * suite of states or a fast-ramp up.
12185                  */
12186                 bbr_restart_after_idle(bbr,
12187                                        cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time));
12188         }
12189         /*
12190          * Now was there a hptsi delay where we are behind? We only count
12191          * being behind if: a) We are not in recovery. b) There was a delay.
12192          * <and> c) We had room to send something.
12193          *
12194          */
12195         hpts_calling = inp->inp_hpts_calls;
12196         inp->inp_hpts_calls = 0;
12197         if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
12198                 if (bbr_process_timers(tp, bbr, cts, hpts_calling)) {
12199                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1);
12200                         return (0);
12201                 }
12202         }
12203         bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12204         if (hpts_calling &&
12205             (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12206                 bbr->r_ctl.rc_last_delay_val = 0;
12207         }
12208         bbr->r_timer_override = 0;
12209         bbr->r_wanted_output = 0;
12210         /*
12211          * For TFO connections in SYN_RECEIVED, only allow the initial
12212          * SYN|ACK and those sent by the retransmit timer.
12213          */
12214         if (IS_FASTOPEN(tp->t_flags) &&
12215             ((tp->t_state == TCPS_SYN_RECEIVED) ||
12216              (tp->t_state == TCPS_SYN_SENT)) &&
12217             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
12218             (tp->t_rxtshift == 0)) {    /* not a retransmit */
12219                 len = 0;
12220                 goto just_return_nolock;
12221         }
12222         /*
12223          * Before sending anything check for a state update. For hpts
12224          * calling without input this is important. If its input calling
12225          * then this was already done.
12226          */
12227         if (bbr->rc_use_google == 0)
12228                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12229 again:
12230         /*
12231          * If we've recently taken a timeout, snd_max will be greater than
12232          * snd_max. BBR in general does not pay much attention to snd_nxt
12233          * for historic reasons the persist timer still uses it. This means
12234          * we have to look at it. All retransmissions that are not persits
12235          * use the rsm that needs to be sent so snd_nxt is ignored. At the
12236          * end of this routine we pull snd_nxt always up to snd_max.
12237          */
12238         doing_tlp = 0;
12239 #ifdef BBR_INVARIANTS
12240         doing_retran_from = picked_up_retran = 0;
12241 #endif
12242         error = 0;
12243         tso = 0;
12244         slot = 0;
12245         mtu = 0;
12246         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
12247         sb_offset = tp->snd_max - tp->snd_una;
12248         flags = tcp_outflags[tp->t_state];
12249         sack_rxmit = 0;
12250         len = 0;
12251         rsm = NULL;
12252         if (flags & TH_RST) {
12253                 SOCKBUF_LOCK(sb);
12254                 goto send;
12255         }
12256 recheck_resend:
12257         while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) {
12258                 /* We need to always have one in reserve */
12259                 rsm = bbr_alloc(bbr);
12260                 if (rsm == NULL) {
12261                         error = ENOMEM;
12262                         /* Lie to get on the hpts */
12263                         tot_len = tp->t_maxseg;
12264                         if (hpts_calling)
12265                                 /* Retry in a ms */
12266                                 slot = 1001;
12267                         goto just_return_nolock;
12268                 }
12269                 TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next);
12270                 bbr->r_ctl.rc_free_cnt++;
12271                 rsm = NULL;
12272         }
12273         /* What do we send, a resend? */
12274         if (bbr->r_ctl.rc_resend == NULL) {
12275                 /* Check for rack timeout */
12276                 bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts);
12277                 if (bbr->r_ctl.rc_resend) {
12278 #ifdef BBR_INVARIANTS
12279                         picked_up_retran = 1;
12280 #endif
12281                         bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend);
12282                 }
12283         }
12284         if (bbr->r_ctl.rc_resend) {
12285                 rsm = bbr->r_ctl.rc_resend;
12286 #ifdef BBR_INVARIANTS
12287                 doing_retran_from = 1;
12288 #endif
12289                 /* Remove any TLP flags its a RACK or T-O */
12290                 rsm->r_flags &= ~BBR_TLP;
12291                 bbr->r_ctl.rc_resend = NULL;
12292                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
12293 #ifdef BBR_INVARIANTS
12294                         panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n",
12295                             tp, bbr, rsm, rsm->r_start, tp->snd_una);
12296                         goto recheck_resend;
12297 #else
12298                         /* TSNH */
12299                         rsm = NULL;
12300                         goto recheck_resend;
12301 #endif
12302                 }
12303                 rtr_cnt++;
12304                 if (rsm->r_flags & BBR_HAS_SYN) {
12305                         /* Only retransmit a SYN by itself */
12306                         len = 0;
12307                         if ((flags & TH_SYN) == 0) {
12308                                 /* Huh something is wrong */
12309                                 rsm->r_start++;
12310                                 if (rsm->r_start == rsm->r_end) {
12311                                         /* Clean it up, somehow we missed the ack? */
12312                                         bbr_log_syn(tp, NULL);
12313                                 } else {
12314                                         /* TFO with data? */
12315                                         rsm->r_flags &= ~BBR_HAS_SYN;
12316                                         len = rsm->r_end - rsm->r_start;
12317                                 }
12318                         } else {
12319                                 /* Retransmitting SYN */
12320                                 rsm = NULL;
12321                                 SOCKBUF_LOCK(sb);
12322                                 goto send;
12323                         }
12324                 } else
12325                         len = rsm->r_end - rsm->r_start;
12326                 if ((bbr->rc_resends_use_tso == 0) &&
12327                     (len > maxseg)) {
12328                         len = maxseg;
12329                         more_to_rxt = 1;
12330                 }
12331                 sb_offset = rsm->r_start - tp->snd_una;
12332                 if (len > 0) {
12333                         sack_rxmit = 1;
12334                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
12335                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
12336                             min(len, maxseg));
12337                 } else {
12338                         /* I dont think this can happen */
12339                         rsm = NULL;
12340                         goto recheck_resend;
12341                 }
12342                 BBR_STAT_INC(bbr_resends_set);
12343         } else if (bbr->r_ctl.rc_tlp_send) {
12344                 /*
12345                  * Tail loss probe
12346                  */
12347                 doing_tlp = 1;
12348                 rsm = bbr->r_ctl.rc_tlp_send;
12349                 bbr->r_ctl.rc_tlp_send = NULL;
12350                 sack_rxmit = 1;
12351                 len = rsm->r_end - rsm->r_start;
12352                 rtr_cnt++;
12353                 if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12354                         len = maxseg;
12355
12356                 if (SEQ_GT(tp->snd_una, rsm->r_start)) {
12357 #ifdef BBR_INVARIANTS
12358                         panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u",
12359                             tp, bbr, tp->snd_una, rsm, rsm->r_start);
12360 #else
12361                         /* TSNH */
12362                         rsm = NULL;
12363                         goto recheck_resend;
12364 #endif
12365                 }
12366                 sb_offset = rsm->r_start - tp->snd_una;
12367                 BBR_STAT_INC(bbr_tlp_set);
12368         }
12369         /*
12370          * Enforce a connection sendmap count limit if set
12371          * as long as we are not retransmiting.
12372          */
12373         if ((rsm == NULL) &&
12374             (V_tcp_map_entries_limit > 0) &&
12375             (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
12376                 BBR_STAT_INC(bbr_alloc_limited);
12377                 if (!bbr->alloc_limit_reported) {
12378                         bbr->alloc_limit_reported = 1;
12379                         BBR_STAT_INC(bbr_alloc_limited_conns);
12380                 }
12381                 goto just_return_nolock;
12382         }
12383 #ifdef BBR_INVARIANTS
12384         if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) {
12385                 panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u",
12386                     tp, bbr, rsm, sb_offset, len);
12387         }
12388 #endif
12389         /*
12390          * Get standard flags, and add SYN or FIN if requested by 'hidden'
12391          * state flags.
12392          */
12393         if (tp->t_flags & TF_NEEDFIN && (rsm == NULL))
12394                 flags |= TH_FIN;
12395         if (tp->t_flags & TF_NEEDSYN)
12396                 flags |= TH_SYN;
12397
12398         if (rsm && (rsm->r_flags & BBR_HAS_FIN)) {
12399                 /* we are retransmitting the fin */
12400                 len--;
12401                 if (len) {
12402                         /*
12403                          * When retransmitting data do *not* include the
12404                          * FIN. This could happen from a TLP probe if we
12405                          * allowed data with a FIN.
12406                          */
12407                         flags &= ~TH_FIN;
12408                 }
12409         } else if (rsm) {
12410                 if (flags & TH_FIN)
12411                         flags &= ~TH_FIN;
12412         }
12413         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
12414                 void *end_rsm;
12415
12416                 end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext);
12417                 if (end_rsm)
12418                         kern_prefetch(end_rsm, &prefetch_rsm);
12419                 prefetch_rsm = 1;
12420         }
12421         SOCKBUF_LOCK(sb);
12422         /*
12423          * If snd_nxt == snd_max and we have transmitted a FIN, the
12424          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
12425          * negative length.  This can also occur when TCP opens up its
12426          * congestion window while receiving additional duplicate acks after
12427          * fast-retransmit because TCP will reset snd_nxt to snd_max after
12428          * the fast-retransmit.
12429          *
12430          * In the normal retransmit-FIN-only case, however, snd_nxt will be
12431          * set to snd_una, the sb_offset will be 0, and the length may wind
12432          * up 0.
12433          *
12434          * If sack_rxmit is true we are retransmitting from the scoreboard
12435          * in which case len is already set.
12436          */
12437         if (sack_rxmit == 0) {
12438                 uint32_t avail;
12439
12440                 avail = sbavail(sb);
12441                 if (SEQ_GT(tp->snd_max, tp->snd_una))
12442                         sb_offset = tp->snd_max - tp->snd_una;
12443                 else
12444                         sb_offset = 0;
12445                 if (bbr->rc_tlp_new_data) {
12446                         /* TLP is forcing out new data */
12447                         uint32_t tlplen;
12448
12449                         doing_tlp = 1;
12450                         tlplen = maxseg;
12451
12452                         if (tlplen > (uint32_t)(avail - sb_offset)) {
12453                                 tlplen = (uint32_t)(avail - sb_offset);
12454                         }
12455                         if (tlplen > tp->snd_wnd) {
12456                                 len = tp->snd_wnd;
12457                         } else {
12458                                 len = tlplen;
12459                         }
12460                         bbr->rc_tlp_new_data = 0;
12461                 } else {
12462                         what_we_can = len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts);
12463                         if ((len < p_maxseg) &&
12464                             (bbr->rc_in_persist == 0) &&
12465                             (ctf_outstanding(tp) >= (2 * p_maxseg)) &&
12466                             ((avail - sb_offset) >= p_maxseg)) {
12467                                 /*
12468                                  * We are not completing whats in the socket
12469                                  * buffer (i.e. there is at least a segment
12470                                  * waiting to send) and we have 2 or more
12471                                  * segments outstanding. There is no sense
12472                                  * of sending a little piece. Lets defer and
12473                                  * and wait until we can send a whole
12474                                  * segment.
12475                                  */
12476                                 len = 0;
12477                         }
12478                         if (bbr->rc_in_persist) {
12479                                 /*
12480                                  * We are in persists, figure out if
12481                                  * a retransmit is available (maybe the previous
12482                                  * persists we sent) or if we have to send new
12483                                  * data.
12484                                  */
12485                                 rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
12486                                 if (rsm) {
12487                                         len = rsm->r_end - rsm->r_start;
12488                                         if (rsm->r_flags & BBR_HAS_FIN)
12489                                                 len--;
12490                                         if ((bbr->rc_resends_use_tso == 0) && (len > maxseg))
12491                                                 len = maxseg;
12492                                         if (len > 1)
12493                                                 BBR_STAT_INC(bbr_persist_reneg);
12494                                         /*
12495                                          * XXXrrs we could force the len to
12496                                          * 1 byte here to cause the chunk to
12497                                          * split apart.. but that would then
12498                                          * mean we always retransmit it as
12499                                          * one byte even after the window
12500                                          * opens.
12501                                          */
12502                                         sack_rxmit = 1;
12503                                         sb_offset = rsm->r_start - tp->snd_una;
12504                                 } else {
12505                                         /*
12506                                          * First time through in persists or peer
12507                                          * acked our one byte. Though we do have
12508                                          * to have something in the sb.
12509                                          */
12510                                         len = 1;
12511                                         sb_offset = 0;
12512                                         if (avail == 0)
12513                                             len = 0;
12514                                 }
12515                         }
12516                 }
12517         }
12518         if (prefetch_so_done == 0) {
12519                 kern_prefetch(so, &prefetch_so_done);
12520                 prefetch_so_done = 1;
12521         }
12522         /*
12523          * Lop off SYN bit if it has already been sent.  However, if this is
12524          * SYN-SENT state and if segment contains data and if we don't know
12525          * that foreign host supports TAO, suppress sending segment.
12526          */
12527         if ((flags & TH_SYN) && (rsm == NULL) &&
12528             SEQ_GT(tp->snd_max, tp->snd_una)) {
12529                 if (tp->t_state != TCPS_SYN_RECEIVED)
12530                         flags &= ~TH_SYN;
12531                 /*
12532                  * When sending additional segments following a TFO SYN|ACK,
12533                  * do not include the SYN bit.
12534                  */
12535                 if (IS_FASTOPEN(tp->t_flags) &&
12536                     (tp->t_state == TCPS_SYN_RECEIVED))
12537                         flags &= ~TH_SYN;
12538                 sb_offset--, len++;
12539                 if (sbavail(sb) == 0)
12540                         len = 0;
12541         } else if ((flags & TH_SYN) && rsm) {
12542                 /*
12543                  * Subtract one from the len for the SYN being
12544                  * retransmitted.
12545                  */
12546                 len--;
12547         }
12548         /*
12549          * Be careful not to send data and/or FIN on SYN segments. This
12550          * measure is needed to prevent interoperability problems with not
12551          * fully conformant TCP implementations.
12552          */
12553         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
12554                 len = 0;
12555                 flags &= ~TH_FIN;
12556         }
12557         /*
12558          * On TFO sockets, ensure no data is sent in the following cases:
12559          *
12560          *  - When retransmitting SYN|ACK on a passively-created socket
12561          *  - When retransmitting SYN on an actively created socket
12562          *  - When sending a zero-length cookie (cookie request) on an
12563          *    actively created socket
12564          *  - When the socket is in the CLOSED state (RST is being sent)
12565          */
12566         if (IS_FASTOPEN(tp->t_flags) &&
12567             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
12568              ((tp->t_state == TCPS_SYN_SENT) &&
12569               (tp->t_tfo_client_cookie_len == 0)) ||
12570              (flags & TH_RST))) {
12571                 len = 0;
12572                 sack_rxmit = 0;
12573                 rsm = NULL;
12574         }
12575         /* Without fast-open there should never be data sent on a SYN */
12576         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags)))
12577                 len = 0;
12578         if (len <= 0) {
12579                 /*
12580                  * If FIN has been sent but not acked, but we haven't been
12581                  * called to retransmit, len will be < 0.  Otherwise, window
12582                  * shrank after we sent into it.  If window shrank to 0,
12583                  * cancel pending retransmit, pull snd_nxt back to (closed)
12584                  * window, and set the persist timer if it isn't already
12585                  * going.  If the window didn't close completely, just wait
12586                  * for an ACK.
12587                  *
12588                  * We also do a general check here to ensure that we will
12589                  * set the persist timer when we have data to send, but a
12590                  * 0-byte window. This makes sure the persist timer is set
12591                  * even if the packet hits one of the "goto send" lines
12592                  * below.
12593                  */
12594                 len = 0;
12595                 if ((tp->snd_wnd == 0) &&
12596                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12597                     (tp->snd_una == tp->snd_max) &&
12598                     (sb_offset < (int)sbavail(sb))) {
12599                         /*
12600                          * Not enough room in the rwnd to send
12601                          * a paced segment out.
12602                          */
12603                         bbr_enter_persist(tp, bbr, cts, __LINE__);
12604                 }
12605         } else if ((rsm == NULL) &&
12606                    (doing_tlp == 0) &&
12607                    (len < bbr->r_ctl.rc_pace_max_segs)) {
12608                 /*
12609                  * We are not sending a full segment for
12610                  * some reason. Should we not send anything (think
12611                  * sws or persists)?
12612                  */
12613                 if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12614                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
12615                     (len < (int)(sbavail(sb) - sb_offset))) {
12616                         /*
12617                          * Here the rwnd is less than
12618                          * the pacing size, this is not a retransmit,
12619                          * we are established and
12620                          * the send is not the last in the socket buffer
12621                          * lets not send, and possibly enter persists.
12622                          */
12623                         len = 0;
12624                         if (tp->snd_max == tp->snd_una)
12625                                 bbr_enter_persist(tp, bbr, cts, __LINE__);
12626                 } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) &&
12627                            (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12628                                                  bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12629                            (len < (int)(sbavail(sb) - sb_offset)) &&
12630                            (len < bbr_minseg(bbr))) {
12631                         /*
12632                          * Here we are not retransmitting, and
12633                          * the cwnd is not so small that we could
12634                          * not send at least a min size (rxt timer
12635                          * not having gone off), We have 2 segments or
12636                          * more already in flight, its not the tail end
12637                          * of the socket buffer  and the cwnd is blocking
12638                          * us from sending out minimum pacing segment size.
12639                          * Lets not send anything.
12640                          */
12641                         bbr->rc_cwnd_limited = 1;
12642                         len = 0;
12643                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
12644                             min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) &&
12645                            (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12646                                                  bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) &&
12647                            (len < (int)(sbavail(sb) - sb_offset)) &&
12648                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
12649                         /*
12650                          * Here we have a send window but we have
12651                          * filled it up and we can't send another pacing segment.
12652                          * We also have in flight more than 2 segments
12653                          * and we are not completing the sb i.e. we allow
12654                          * the last bytes of the sb to go out even if
12655                          * its not a full pacing segment.
12656                          */
12657                         len = 0;
12658                 }
12659         }
12660         /* len will be >= 0 after this point. */
12661         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
12662         tcp_sndbuf_autoscale(tp, so, sendwin);
12663         /*
12664          *
12665          */
12666         if (bbr->rc_in_persist &&
12667             len &&
12668             (rsm == NULL) &&
12669             (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) {
12670                 /*
12671                  * We are in persist, not doing a retransmit and don't have enough space
12672                  * yet to send a full TSO. So is it at the end of the sb
12673                  * if so we need to send else nuke to 0 and don't send.
12674                  */
12675                 int sbleft;
12676                 if (sbavail(sb) > sb_offset)
12677                         sbleft = sbavail(sb) - sb_offset;
12678                 else
12679                         sbleft = 0;
12680                 if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) {
12681                         /* not at end of sb lets not send */
12682                         len = 0;
12683                 }
12684         }
12685         /*
12686          * Decide if we can use TCP Segmentation Offloading (if supported by
12687          * hardware).
12688          *
12689          * TSO may only be used if we are in a pure bulk sending state.  The
12690          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
12691          * options prevent using TSO.  With TSO the TCP header is the same
12692          * (except for the sequence number) for all generated packets.  This
12693          * makes it impossible to transmit any options which vary per
12694          * generated segment or packet.
12695          *
12696          * IPv4 handling has a clear separation of ip options and ip header
12697          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen()
12698          * does the right thing below to provide length of just ip options
12699          * and thus checking for ipoptlen is enough to decide if ip options
12700          * are present.
12701          */
12702 #ifdef INET6
12703         if (isipv6)
12704                 ipoptlen = ip6_optlen(inp);
12705         else
12706 #endif
12707         if (inp->inp_options)
12708                 ipoptlen = inp->inp_options->m_len -
12709                     offsetof(struct ipoption, ipopt_list);
12710         else
12711                 ipoptlen = 0;
12712 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12713         /*
12714          * Pre-calculate here as we save another lookup into the darknesses
12715          * of IPsec that way and can actually decide if TSO is ok.
12716          */
12717 #ifdef INET6
12718         if (isipv6 && IPSEC_ENABLED(ipv6))
12719                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
12720 #ifdef INET
12721         else
12722 #endif
12723 #endif                          /* INET6 */
12724 #ifdef INET
12725         if (IPSEC_ENABLED(ipv4))
12726                 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
12727 #endif                          /* INET */
12728 #endif                          /* IPSEC */
12729 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
12730         ipoptlen += ipsec_optlen;
12731 #endif
12732         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
12733             (len > maxseg) &&
12734             (tp->t_port == 0) &&
12735             ((tp->t_flags & TF_SIGNATURE) == 0) &&
12736             tp->rcv_numsacks == 0 &&
12737             ipoptlen == 0)
12738                 tso = 1;
12739
12740         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
12741             (long)TCP_MAXWIN << tp->rcv_scale);
12742         /*
12743          * Sender silly window avoidance.   We transmit under the following
12744          * conditions when len is non-zero:
12745          *
12746          * - We have a full segment (or more with TSO) - This is the last
12747          * buffer in a write()/send() and we are either idle or running
12748          * NODELAY - we've timed out (e.g. persist timer) - we have more
12749          * then 1/2 the maximum send window's worth of data (receiver may be
12750          * limited the window size) - we need to retransmit
12751          */
12752         if (rsm)
12753                 goto send;
12754         if (len) {
12755                 if (sack_rxmit)
12756                         goto send;
12757                 if (len >= p_maxseg)
12758                         goto send;
12759                 /*
12760                  * NOTE! on localhost connections an 'ack' from the remote
12761                  * end may occur synchronously with the output and cause us
12762                  * to flush a buffer queued with moretocome.  XXX
12763                  *
12764                  */
12765                 if (((tp->t_flags & TF_MORETOCOME) == 0) &&     /* normal case */
12766                     ((tp->t_flags & TF_NODELAY) ||
12767                     ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) &&
12768                     (tp->t_flags & TF_NOPUSH) == 0) {
12769                         goto send;
12770                 }
12771                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
12772                         goto send;
12773                 }
12774                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
12775                         goto send;
12776                 }
12777         }
12778         /*
12779          * Sending of standalone window updates.
12780          *
12781          * Window updates are important when we close our window due to a
12782          * full socket buffer and are opening it again after the application
12783          * reads data from it.  Once the window has opened again and the
12784          * remote end starts to send again the ACK clock takes over and
12785          * provides the most current window information.
12786          *
12787          * We must avoid the silly window syndrome whereas every read from
12788          * the receive buffer, no matter how small, causes a window update
12789          * to be sent.  We also should avoid sending a flurry of window
12790          * updates when the socket buffer had queued a lot of data and the
12791          * application is doing small reads.
12792          *
12793          * Prevent a flurry of pointless window updates by only sending an
12794          * update when we can increase the advertized window by more than
12795          * 1/4th of the socket buffer capacity.  When the buffer is getting
12796          * full or is very small be more aggressive and send an update
12797          * whenever we can increase by two mss sized segments. In all other
12798          * situations the ACK's to new incoming data will carry further
12799          * window increases.
12800          *
12801          * Don't send an independent window update if a delayed ACK is
12802          * pending (it will get piggy-backed on it) or the remote side
12803          * already has done a half-close and won't send more data.  Skip
12804          * this if the connection is in T/TCP half-open state.
12805          */
12806         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
12807             !(tp->t_flags & TF_DELACK) &&
12808             !TCPS_HAVERCVDFIN(tp->t_state)) {
12809                 /* Check to see if we should do a window update */
12810                 if (bbr_window_update_needed(tp, so, recwin, maxseg))
12811                         goto send;
12812         }
12813         /*
12814          * Send if we owe the peer an ACK, RST, SYN.  ACKNOW
12815          * is also a catch-all for the retransmit timer timeout case.
12816          */
12817         if (tp->t_flags & TF_ACKNOW) {
12818                 goto send;
12819         }
12820         if (flags & TH_RST) {
12821                 /* Always send a RST if one is due */
12822                 goto send;
12823         }
12824         if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) {
12825                 goto send;
12826         }
12827         /*
12828          * If our state indicates that FIN should be sent and we have not
12829          * yet done so, then we need to send.
12830          */
12831         if (flags & TH_FIN &&
12832             ((tp->t_flags & TF_SENTFIN) == 0)) {
12833                 goto send;
12834         }
12835         /*
12836          * No reason to send a segment, just return.
12837          */
12838 just_return:
12839         SOCKBUF_UNLOCK(sb);
12840 just_return_nolock:
12841         if (tot_len)
12842                 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
12843         if (bbr->rc_no_pacing)
12844                 slot = 0;
12845         if (tot_len == 0) {
12846                 if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >=
12847                     tp->snd_wnd) {
12848                         BBR_STAT_INC(bbr_rwnd_limited);
12849                         app_limited = BBR_JR_RWND_LIMITED;
12850                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12851                         if ((bbr->rc_in_persist == 0) &&
12852                             TCPS_HAVEESTABLISHED(tp->t_state) &&
12853                             (tp->snd_max == tp->snd_una) &&
12854                             sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
12855                                 /* No send window.. we must enter persist */
12856                                 bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__);
12857                         }
12858                 } else if (ctf_outstanding(tp) >= sbavail(sb)) {
12859                         BBR_STAT_INC(bbr_app_limited);
12860                         app_limited = BBR_JR_APP_LIMITED;
12861                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12862                 } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12863                                                  bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) {
12864                         BBR_STAT_INC(bbr_cwnd_limited);
12865                         app_limited = BBR_JR_CWND_LIMITED;
12866                         bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12867                                                                         bbr->r_ctl.rc_lost_bytes)));
12868                         bbr->rc_cwnd_limited = 1;
12869                 } else {
12870                         BBR_STAT_INC(bbr_app_limited);
12871                         app_limited = BBR_JR_APP_LIMITED;
12872                         bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp));
12873                 }
12874                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
12875                 bbr->r_agg_early_set = 0;
12876                 bbr->r_ctl.rc_agg_early = 0;
12877                 bbr->r_ctl.rc_last_delay_val = 0;
12878         } else if (bbr->rc_use_google == 0)
12879                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
12880         /* Are we app limited? */
12881         if ((app_limited == BBR_JR_APP_LIMITED) ||
12882             (app_limited == BBR_JR_RWND_LIMITED)) {
12883                 /**
12884                  * We are application limited.
12885                  */
12886                 bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
12887                                                                        bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered);
12888         }
12889         if (tot_len == 0)
12890                 counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1);
12891         /* Dont update the time if we did not send */
12892         bbr->r_ctl.rc_last_delay_val = 0;
12893         bbr->rc_output_starts_timer = 1;
12894         bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len);
12895         bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len);
12896         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12897                 /* Make sure snd_nxt is drug up */
12898                 tp->snd_nxt = tp->snd_max;
12899         }
12900         return (error);
12901
12902 send:
12903         if (doing_tlp == 0) {
12904                 /*
12905                  * Data not a TLP, and its not the rxt firing. If it is the
12906                  * rxt firing, we want to leave the tlp_in_progress flag on
12907                  * so we don't send another TLP. It has to be a rack timer
12908                  * or normal send (response to acked data) to clear the tlp
12909                  * in progress flag.
12910                  */
12911                 bbr->rc_tlp_in_progress = 0;
12912                 bbr->rc_tlp_rtx_out = 0;
12913         } else {
12914                 /*
12915                  * Its a TLP.
12916                  */
12917                 bbr->rc_tlp_in_progress = 1;
12918         }
12919         bbr_timer_cancel(bbr, __LINE__, cts);
12920         if (rsm == NULL) {
12921                 if (sbused(sb) > 0) {
12922                         /*
12923                          * This is sub-optimal. We only send a stand alone
12924                          * FIN on its own segment.
12925                          */
12926                         if (flags & TH_FIN) {
12927                                 flags &= ~TH_FIN;
12928                                 if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) {
12929                                         /* Lets not send this */
12930                                         slot = 0;
12931                                         goto just_return;
12932                                 }
12933                         }
12934                 }
12935         } else {
12936                 /*
12937                  * We do *not* send a FIN on a retransmit if it has data.
12938                  * The if clause here where len > 1 should never come true.
12939                  */
12940                 if ((len > 0) &&
12941                     (((rsm->r_flags & BBR_HAS_FIN) == 0) &&
12942                     (flags & TH_FIN))) {
12943                         flags &= ~TH_FIN;
12944                         len--;
12945                 }
12946         }
12947         SOCKBUF_LOCK_ASSERT(sb);
12948         if (len > 0) {
12949                 if ((tp->snd_una == tp->snd_max) &&
12950                     (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) {
12951                         /*
12952                          * This qualifies as a RTT_PROBE session since we
12953                          * drop the data outstanding to nothing and waited
12954                          * more than bbr_rtt_probe_time.
12955                          */
12956                         bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0);
12957                         bbr_set_reduced_rtt(bbr, cts, __LINE__);
12958                 }
12959                 if (len >= maxseg)
12960                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
12961                 else
12962                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
12963         }
12964         /*
12965          * Before ESTABLISHED, force sending of initial options unless TCP
12966          * set not to do any options. NOTE: we assume that the IP/TCP header
12967          * plus TCP options always fit in a single mbuf, leaving room for a
12968          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
12969          * + optlen <= MCLBYTES
12970          */
12971         optlen = 0;
12972 #ifdef INET6
12973         if (isipv6)
12974                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12975         else
12976 #endif
12977                 hdrlen = sizeof(struct tcpiphdr);
12978
12979         /*
12980          * Compute options for segment. We only have to care about SYN and
12981          * established connection segments.  Options for SYN-ACK segments
12982          * are handled in TCP syncache.
12983          */
12984         to.to_flags = 0;
12985         local_options = 0;
12986         if ((tp->t_flags & TF_NOOPT) == 0) {
12987                 /* Maximum segment size. */
12988                 if (flags & TH_SYN) {
12989                         to.to_mss = tcp_mssopt(&inp->inp_inc);
12990 #ifdef NETFLIX_TCPOUDP
12991                         if (tp->t_port)
12992                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
12993 #endif
12994                         to.to_flags |= TOF_MSS;
12995                         /*
12996                          * On SYN or SYN|ACK transmits on TFO connections,
12997                          * only include the TFO option if it is not a
12998                          * retransmit, as the presence of the TFO option may
12999                          * have caused the original SYN or SYN|ACK to have
13000                          * been dropped by a middlebox.
13001                          */
13002                         if (IS_FASTOPEN(tp->t_flags) &&
13003                             (tp->t_rxtshift == 0)) {
13004                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
13005                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
13006                                         to.to_tfo_cookie =
13007                                             (u_int8_t *)&tp->t_tfo_cookie.server;
13008                                         to.to_flags |= TOF_FASTOPEN;
13009                                         wanted_cookie = 1;
13010                                 } else if (tp->t_state == TCPS_SYN_SENT) {
13011                                         to.to_tfo_len =
13012                                             tp->t_tfo_client_cookie_len;
13013                                         to.to_tfo_cookie =
13014                                             tp->t_tfo_cookie.client;
13015                                         to.to_flags |= TOF_FASTOPEN;
13016                                         wanted_cookie = 1;
13017                                 }
13018                         }
13019                 }
13020                 /* Window scaling. */
13021                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
13022                         to.to_wscale = tp->request_r_scale;
13023                         to.to_flags |= TOF_SCALE;
13024                 }
13025                 /* Timestamps. */
13026                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
13027                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
13028                         to.to_tsval =   tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset;
13029                         to.to_tsecr = tp->ts_recent;
13030                         to.to_flags |= TOF_TS;
13031                         local_options += TCPOLEN_TIMESTAMP + 2;
13032                 }
13033                 /* Set receive buffer autosizing timestamp. */
13034                 if (tp->rfbuf_ts == 0 &&
13035                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
13036                         tp->rfbuf_ts =  tcp_tv_to_mssectick(&bbr->rc_tv);
13037                 /* Selective ACK's. */
13038                 if (flags & TH_SYN)
13039                         to.to_flags |= TOF_SACKPERM;
13040                 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13041                     tp->rcv_numsacks > 0) {
13042                         to.to_flags |= TOF_SACK;
13043                         to.to_nsacks = tp->rcv_numsacks;
13044                         to.to_sacks = (u_char *)tp->sackblks;
13045                 }
13046 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13047                 /* TCP-MD5 (RFC2385). */
13048                 if (tp->t_flags & TF_SIGNATURE)
13049                         to.to_flags |= TOF_SIGNATURE;
13050 #endif                          /* TCP_SIGNATURE */
13051
13052                 /* Processing the options. */
13053                 hdrlen += (optlen = tcp_addoptions(&to, opt));
13054                 /*
13055                  * If we wanted a TFO option to be added, but it was unable
13056                  * to fit, ensure no data is sent.
13057                  */
13058                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
13059                     !(to.to_flags & TOF_FASTOPEN))
13060                         len = 0;
13061         }
13062 #ifdef NETFLIX_TCPOUDP
13063         if (tp->t_port) {
13064                 if (V_tcp_udp_tunneling_port == 0) {
13065                         /* The port was removed?? */
13066                         SOCKBUF_UNLOCK(&so->so_snd);
13067                         return (EHOSTUNREACH);
13068                 }
13069                 hdrlen += sizeof(struct udphdr);
13070         }
13071 #endif
13072 #ifdef INET6
13073         if (isipv6)
13074                 ipoptlen = ip6_optlen(tp->t_inpcb);
13075         else
13076 #endif
13077         if (tp->t_inpcb->inp_options)
13078                 ipoptlen = tp->t_inpcb->inp_options->m_len -
13079                     offsetof(struct ipoption, ipopt_list);
13080         else
13081                 ipoptlen = 0;
13082         ipoptlen = 0;
13083 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
13084         ipoptlen += ipsec_optlen;
13085 #endif
13086         if (bbr->rc_last_options != local_options) {
13087                 /*
13088                  * Cache the options length this generally does not change
13089                  * on a connection. We use this to calculate TSO.
13090                  */
13091                 bbr->rc_last_options = local_options;
13092         }
13093         maxseg = tp->t_maxseg - (ipoptlen + optlen);
13094         p_maxseg = min(maxseg, pace_max_segs);
13095         /*
13096          * Adjust data length if insertion of options will bump the packet
13097          * length beyond the t_maxseg length. Clear the FIN bit because we
13098          * cut off the tail of the segment.
13099          */
13100         if (len > maxseg) {
13101                 if (len != 0 && (flags & TH_FIN)) {
13102                         flags &= ~TH_FIN;
13103                 }
13104                 if (tso) {
13105                         uint32_t moff;
13106                         int32_t max_len;
13107
13108                         /* extract TSO information */
13109                         if_hw_tsomax = tp->t_tsomax;
13110                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
13111                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
13112                         KASSERT(ipoptlen == 0,
13113                             ("%s: TSO can't do IP options", __func__));
13114
13115                         /*
13116                          * Check if we should limit by maximum payload
13117                          * length:
13118                          */
13119                         if (if_hw_tsomax != 0) {
13120                                 /* compute maximum TSO length */
13121                                 max_len = (if_hw_tsomax - hdrlen -
13122                                     max_linkhdr);
13123                                 if (max_len <= 0) {
13124                                         len = 0;
13125                                 } else if (len > max_len) {
13126                                         len = max_len;
13127                                 }
13128                         }
13129                         /*
13130                          * Prevent the last segment from being fractional
13131                          * unless the send sockbuf can be emptied:
13132                          */
13133                         if ((sb_offset + len) < sbavail(sb)) {
13134                                 moff = len % (uint32_t)maxseg;
13135                                 if (moff != 0) {
13136                                         len -= moff;
13137                                 }
13138                         }
13139                         /*
13140                          * In case there are too many small fragments don't
13141                          * use TSO:
13142                          */
13143                         if (len <= maxseg) {
13144                                 len = maxseg;
13145                                 tso = 0;
13146                         }
13147                 } else {
13148                         /* Not doing TSO */
13149                         if (optlen + ipoptlen >= tp->t_maxseg) {
13150                                 /*
13151                                  * Since we don't have enough space to put
13152                                  * the IP header chain and the TCP header in
13153                                  * one packet as required by RFC 7112, don't
13154                                  * send it. Also ensure that at least one
13155                                  * byte of the payload can be put into the
13156                                  * TCP segment.
13157                                  */
13158                                 SOCKBUF_UNLOCK(&so->so_snd);
13159                                 error = EMSGSIZE;
13160                                 sack_rxmit = 0;
13161                                 goto out;
13162                         }
13163                         len = maxseg;
13164                 }
13165         } else {
13166                 /* Not doing TSO */
13167                 if_hw_tsomaxsegcount = 0;
13168                 tso = 0;
13169         }
13170         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
13171             ("%s: len > IP_MAXPACKET", __func__));
13172 #ifdef DIAGNOSTIC
13173 #ifdef INET6
13174         if (max_linkhdr + hdrlen > MCLBYTES)
13175 #else
13176         if (max_linkhdr + hdrlen > MHLEN)
13177 #endif
13178                 panic("tcphdr too big");
13179 #endif
13180         /*
13181          * This KASSERT is here to catch edge cases at a well defined place.
13182          * Before, those had triggered (random) panic conditions further
13183          * down.
13184          */
13185 #ifdef BBR_INVARIANTS
13186         if (sack_rxmit) {
13187                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
13188                         panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u",
13189                             rsm, tp, bbr, rsm->r_start, tp->snd_una);
13190                 }
13191         }
13192 #endif
13193         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
13194         if ((len == 0) &&
13195             (flags & TH_FIN) &&
13196             (sbused(sb))) {
13197                 /*
13198                  * We have outstanding data, don't send a fin by itself!.
13199                  */
13200                 slot = 0;
13201                 goto just_return;
13202         }
13203         /*
13204          * Grab a header mbuf, attaching a copy of data to be transmitted,
13205          * and initialize the header from the template for sends on this
13206          * connection.
13207          */
13208         if (len) {
13209                 uint32_t moff;
13210                 uint32_t orig_len;
13211
13212                 /*
13213                  * We place a limit on sending with hptsi.
13214                  */
13215                 if ((rsm == NULL) && len > pace_max_segs)
13216                         len = pace_max_segs;
13217                 if (len <= maxseg)
13218                         tso = 0;
13219 #ifdef INET6
13220                 if (MHLEN < hdrlen + max_linkhdr)
13221                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
13222                 else
13223 #endif
13224                         m = m_gethdr(M_NOWAIT, MT_DATA);
13225
13226                 if (m == NULL) {
13227                         BBR_STAT_INC(bbr_failed_mbuf_aloc);
13228                         bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13229                         SOCKBUF_UNLOCK(sb);
13230                         error = ENOBUFS;
13231                         sack_rxmit = 0;
13232                         goto out;
13233                 }
13234                 m->m_data += max_linkhdr;
13235                 m->m_len = hdrlen;
13236                 /*
13237                  * Start the m_copy functions from the closest mbuf to the
13238                  * sb_offset in the socket buffer chain.
13239                  */
13240                 if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) {
13241 #ifdef BBR_INVARIANTS
13242                         if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0)))
13243                                 panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u",
13244                                     tp, bbr, len, sb_offset, sbavail(sb), rsm,
13245                                     doing_retran_from,
13246                                     picked_up_retran,
13247                                     doing_tlp);
13248
13249 #endif
13250                         /*
13251                          * In this messed up situation we have two choices,
13252                          * a) pretend the send worked, and just start timers
13253                          * and what not (not good since that may lead us
13254                          * back here a lot). <or> b) Send the lowest segment
13255                          * in the map. <or> c) Drop the connection. Lets do
13256                          * <b> which if it continues to happen will lead to
13257                          * <c> via timeouts.
13258                          */
13259                         BBR_STAT_INC(bbr_offset_recovery);
13260                         rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map);
13261                         sb_offset = 0;
13262                         if (rsm == NULL) {
13263                                 sack_rxmit = 0;
13264                                 len = sbavail(sb);
13265                         } else {
13266                                 sack_rxmit = 1;
13267                                 if (rsm->r_start != tp->snd_una) {
13268                                         /*
13269                                          * Things are really messed up, <c>
13270                                          * is the only thing to do.
13271                                          */
13272                                         BBR_STAT_INC(bbr_offset_drop);
13273                                         tcp_set_inp_to_drop(inp, EFAULT);
13274                                         SOCKBUF_UNLOCK(sb);
13275                                         (void)m_free(m);
13276                                         return (0);
13277                                 }
13278                                 len = rsm->r_end - rsm->r_start;
13279                         }
13280                         if (len > sbavail(sb))
13281                                 len = sbavail(sb);
13282                         if (len > maxseg)
13283                                 len = maxseg;
13284                 }
13285                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
13286                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
13287                         m_copydata(mb, moff, (int)len,
13288                             mtod(m, caddr_t)+hdrlen);
13289                         if (rsm == NULL)
13290                                 sbsndptr_adv(sb, mb, len);
13291                         m->m_len += len;
13292                 } else {
13293                         struct sockbuf *msb;
13294
13295                         if (rsm)
13296                                 msb = NULL;
13297                         else
13298                                 msb = sb;
13299 #ifdef BBR_INVARIANTS
13300                         if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) {
13301                                 if (rsm) {
13302                                         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 ",
13303                                             tp, bbr, len, moff,
13304                                             sbavail(sb), rsm,
13305                                             tp->snd_una, rsm->r_flags, rsm->r_start,
13306                                             doing_retran_from,
13307                                             picked_up_retran,
13308                                             doing_tlp, sack_rxmit);
13309                                 } else {
13310                                         panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u",
13311                                             tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una);
13312                                 }
13313                         }
13314 #endif
13315                         orig_len = len;
13316                         m->m_next = tcp_m_copym(
13317                                 mb, moff, &len,
13318                                 if_hw_tsomaxsegcount,
13319                                 if_hw_tsomaxsegsize, msb,
13320                                 ((rsm == NULL) ? hw_tls : 0)
13321 #ifdef NETFLIX_COPY_ARGS
13322                                 , &filled_all
13323 #endif
13324                                 );
13325                         if (len <= maxseg) {
13326                                 /*
13327                                  * Must have ran out of mbufs for the copy
13328                                  * shorten it to no longer need tso. Lets
13329                                  * not put on sendalot since we are low on
13330                                  * mbufs.
13331                                  */
13332                                 tso = 0;
13333                         }
13334                         if (m->m_next == NULL) {
13335                                 SOCKBUF_UNLOCK(sb);
13336                                 (void)m_free(m);
13337                                 error = ENOBUFS;
13338                                 sack_rxmit = 0;
13339                                 goto out;
13340                         }
13341                 }
13342 #ifdef BBR_INVARIANTS
13343                 if (tso && len < maxseg) {
13344                         panic("tp:%p tso on, but len:%d < maxseg:%d",
13345                             tp, len, maxseg);
13346                 }
13347                 if (tso && if_hw_tsomaxsegcount) {
13348                         int32_t seg_cnt = 0;
13349                         struct mbuf *foo;
13350
13351                         foo = m;
13352                         while (foo) {
13353                                 seg_cnt++;
13354                                 foo = foo->m_next;
13355                         }
13356                         if (seg_cnt > if_hw_tsomaxsegcount) {
13357                                 panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount);
13358                         }
13359                 }
13360 #endif
13361                 /*
13362                  * If we're sending everything we've got, set PUSH. (This
13363                  * will keep happy those implementations which only give
13364                  * data to the user when a buffer fills or a PUSH comes in.)
13365                  */
13366                 if (sb_offset + len == sbused(sb) &&
13367                     sbused(sb) &&
13368                     !(flags & TH_SYN)) {
13369                         flags |= TH_PUSH;
13370                 }
13371                 SOCKBUF_UNLOCK(sb);
13372         } else {
13373                 SOCKBUF_UNLOCK(sb);
13374                 if (tp->t_flags & TF_ACKNOW)
13375                         KMOD_TCPSTAT_INC(tcps_sndacks);
13376                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
13377                         KMOD_TCPSTAT_INC(tcps_sndctrl);
13378                 else
13379                         KMOD_TCPSTAT_INC(tcps_sndwinup);
13380
13381                 m = m_gethdr(M_NOWAIT, MT_DATA);
13382                 if (m == NULL) {
13383                         BBR_STAT_INC(bbr_failed_mbuf_aloc);
13384                         bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0);
13385                         error = ENOBUFS;
13386                         /* Fudge the send time since we could not send */
13387                         sack_rxmit = 0;
13388                         goto out;
13389                 }
13390 #ifdef INET6
13391                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
13392                     MHLEN >= hdrlen) {
13393                         M_ALIGN(m, hdrlen);
13394                 } else
13395 #endif
13396                         m->m_data += max_linkhdr;
13397                 m->m_len = hdrlen;
13398         }
13399         SOCKBUF_UNLOCK_ASSERT(sb);
13400         m->m_pkthdr.rcvif = (struct ifnet *)0;
13401 #ifdef MAC
13402         mac_inpcb_create_mbuf(inp, m);
13403 #endif
13404 #ifdef INET6
13405         if (isipv6) {
13406                 ip6 = mtod(m, struct ip6_hdr *);
13407 #ifdef NETFLIX_TCPOUDP
13408                 if (tp->t_port) {
13409                         udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr));
13410                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13411                         udp->uh_dport = tp->t_port;
13412                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
13413                         udp->uh_ulen = htons(ulen);
13414                         th = (struct tcphdr *)(udp + 1);
13415                 } else {
13416 #endif
13417                         th = (struct tcphdr *)(ip6 + 1);
13418
13419 #ifdef NETFLIX_TCPOUDP
13420                 }
13421 #endif
13422                 tcpip_fillheaders(inp,
13423 #ifdef NETFLIX_TCPOUDP
13424                                   tp->t_port,
13425 #endif
13426                                   ip6, th);
13427         } else
13428 #endif                          /* INET6 */
13429         {
13430                 ip = mtod(m, struct ip *);
13431 #ifdef TCPDEBUG
13432                 ipov = (struct ipovly *)ip;
13433 #endif
13434 #ifdef NETFLIX_TCPOUDP
13435                 if (tp->t_port) {
13436                         udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip));
13437                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
13438                         udp->uh_dport = tp->t_port;
13439                         ulen = hdrlen + len - sizeof(struct ip);
13440                         udp->uh_ulen = htons(ulen);
13441                         th = (struct tcphdr *)(udp + 1);
13442                 } else
13443 #endif
13444                         th = (struct tcphdr *)(ip + 1);
13445                 tcpip_fillheaders(inp,
13446 #ifdef NETFLIX_TCPOUDP
13447                                   tp->t_port,
13448 #endif
13449                                   ip, th);
13450         }
13451         /*
13452          * If we are doing retransmissions, then snd_nxt will not reflect
13453          * the first unsent octet.  For ACK only packets, we do not want the
13454          * sequence number of the retransmitted packet, we want the sequence
13455          * number of the next unsent octet.  So, if there is no data (and no
13456          * SYN or FIN), use snd_max instead of snd_nxt when filling in
13457          * ti_seq.  But if we are in persist state, snd_max might reflect
13458          * one byte beyond the right edge of the window, so use snd_nxt in
13459          * that case, since we know we aren't doing a retransmission.
13460          * (retransmit and persist are mutually exclusive...)
13461          */
13462         if (sack_rxmit == 0) {
13463                 if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) {
13464                         /* New data (including new persists) */
13465                         th->th_seq = htonl(tp->snd_max);
13466                         bbr_seq = tp->snd_max;
13467                 } else if (flags & TH_SYN) {
13468                         /* Syn's always send from iss */
13469                         th->th_seq = htonl(tp->iss);
13470                         bbr_seq = tp->iss;
13471                 } else if (flags & TH_FIN) {
13472                         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) {
13473                                 /*
13474                                  * If we sent the fin already its 1 minus
13475                                  * snd_max
13476                                  */
13477                                 th->th_seq = (htonl(tp->snd_max - 1));
13478                                 bbr_seq = (tp->snd_max - 1);
13479                         } else {
13480                                 /* First time FIN use snd_max */
13481                                 th->th_seq = htonl(tp->snd_max);
13482                                 bbr_seq = tp->snd_max;
13483                         }
13484                 } else if (flags & TH_RST) {
13485                         /*
13486                          * For a Reset send the last cum ack in sequence
13487                          * (this like any other choice may still generate a
13488                          * challenge ack, if a ack-update packet is in
13489                          * flight).
13490                          */
13491                         th->th_seq = htonl(tp->snd_una);
13492                         bbr_seq = tp->snd_una;
13493                 } else {
13494                         /*
13495                          * len == 0 and not persist we use snd_max, sending
13496                          * an ack unless we have sent the fin then its 1
13497                          * minus.
13498                          */
13499                         /*
13500                          * XXXRRS Question if we are in persists and we have
13501                          * nothing outstanding to send and we have not sent
13502                          * a FIN, we will send an ACK. In such a case it
13503                          * might be better to send (tp->snd_una - 1) which
13504                          * would force the peer to ack.
13505                          */
13506                         if (tp->t_flags & TF_SENTFIN) {
13507                                 th->th_seq = htonl(tp->snd_max - 1);
13508                                 bbr_seq = (tp->snd_max - 1);
13509                         } else {
13510                                 th->th_seq = htonl(tp->snd_max);
13511                                 bbr_seq = tp->snd_max;
13512                         }
13513                 }
13514         } else {
13515                 /* All retransmits use the rsm to guide the send */
13516                 th->th_seq = htonl(rsm->r_start);
13517                 bbr_seq = rsm->r_start;
13518         }
13519         th->th_ack = htonl(tp->rcv_nxt);
13520         if (optlen) {
13521                 bcopy(opt, th + 1, optlen);
13522                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
13523         }
13524         th->th_flags = flags;
13525         /*
13526          * Calculate receive window.  Don't shrink window, but avoid silly
13527          * window syndrome.
13528          */
13529         if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) &&
13530                                   recwin < maxseg)))
13531                 recwin = 0;
13532         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
13533             recwin < (tp->rcv_adv - tp->rcv_nxt))
13534                 recwin = (tp->rcv_adv - tp->rcv_nxt);
13535         if (recwin > TCP_MAXWIN << tp->rcv_scale)
13536                 recwin = TCP_MAXWIN << tp->rcv_scale;
13537
13538         /*
13539          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
13540          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
13541          * handled in syncache.
13542          */
13543         if (flags & TH_SYN)
13544                 th->th_win = htons((u_short)
13545                     (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
13546         else {
13547                 /* Avoid shrinking window with window scaling. */
13548                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
13549                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
13550         }
13551         /*
13552          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
13553          * window.  This may cause the remote transmitter to stall.  This
13554          * flag tells soreceive() to disable delayed acknowledgements when
13555          * draining the buffer.  This can occur if the receiver is
13556          * attempting to read more data than can be buffered prior to
13557          * transmitting on the connection.
13558          */
13559         if (th->th_win == 0) {
13560                 tp->t_sndzerowin++;
13561                 tp->t_flags |= TF_RXWIN0SENT;
13562         } else
13563                 tp->t_flags &= ~TF_RXWIN0SENT;
13564         /*
13565          * We don't support urgent data, but drag along
13566          * the pointer in case of a stack switch.
13567          */
13568         tp->snd_up = tp->snd_una;
13569
13570 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
13571         if (to.to_flags & TOF_SIGNATURE) {
13572                 /*
13573                  * Calculate MD5 signature and put it into the place
13574                  * determined before. NOTE: since TCP options buffer doesn't
13575                  * point into mbuf's data, calculate offset and use it.
13576                  */
13577                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
13578                     (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
13579                         /*
13580                          * Do not send segment if the calculation of MD5
13581                          * digest has failed.
13582                          */
13583                         goto out;
13584                 }
13585         }
13586 #endif
13587
13588         /*
13589          * Put TCP length in extended header, and then checksum extended
13590          * header and data.
13591          */
13592         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
13593 #ifdef INET6
13594         if (isipv6) {
13595                 /*
13596                  * ip6_plen is not need to be filled now, and will be filled
13597                  * in ip6_output.
13598                  */
13599 #ifdef NETFLIX_TCPOUDP
13600                 if (tp->t_port) {
13601                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
13602                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13603                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
13604                         th->th_sum = htons(0);
13605                         UDPSTAT_INC(udps_opackets);
13606                 } else {
13607 #endif
13608                         csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
13609                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13610                         th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
13611                             optlen + len, IPPROTO_TCP, 0);
13612 #ifdef NETFLIX_TCPOUDP
13613                 }
13614 #endif
13615         }
13616 #endif
13617 #if defined(INET6) && defined(INET)
13618         else
13619 #endif
13620 #ifdef INET
13621         {
13622 #ifdef NETFLIX_TCPOUDP
13623                 if (tp->t_port) {
13624                         m->m_pkthdr.csum_flags = CSUM_UDP;
13625                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
13626                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
13627                             ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
13628                         th->th_sum = htons(0);
13629                         UDPSTAT_INC(udps_opackets);
13630                 } else {
13631 #endif
13632                         csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP;
13633                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
13634                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
13635                             ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
13636                             IPPROTO_TCP + len + optlen));
13637 #ifdef NETFLIX_TCPOUDP
13638                 }
13639 #endif
13640                 /* IP version must be set here for ipv4/ipv6 checking later */
13641                 KASSERT(ip->ip_v == IPVERSION,
13642                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
13643         }
13644 #endif
13645
13646         /*
13647          * Enable TSO and specify the size of the segments. The TCP pseudo
13648          * header checksum is always provided. XXX: Fixme: This is currently
13649          * not the case for IPv6.
13650          */
13651         if (tso) {
13652                 KASSERT(len > maxseg,
13653                     ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg));
13654                 m->m_pkthdr.csum_flags |= CSUM_TSO;
13655                 csum_flags |= CSUM_TSO;
13656                 m->m_pkthdr.tso_segsz = maxseg;
13657         }
13658         KASSERT(len + hdrlen == m_length(m, NULL),
13659             ("%s: mbuf chain different than expected: %d + %u != %u",
13660             __func__, len, hdrlen, m_length(m, NULL)));
13661
13662 #ifdef TCP_HHOOK
13663         /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */
13664         hhook_run_tcp_est_out(tp, th, &to, len, tso);
13665 #endif
13666 #ifdef TCPDEBUG
13667         /*
13668          * Trace.
13669          */
13670         if (so->so_options & SO_DEBUG) {
13671                 u_short save = 0;
13672
13673 #ifdef INET6
13674                 if (!isipv6)
13675 #endif
13676                 {
13677                         save = ipov->ih_len;
13678                         ipov->ih_len = htons(m->m_pkthdr.len    /* - hdrlen +
13679                               * (th->th_off << 2) */ );
13680                 }
13681                 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
13682 #ifdef INET6
13683                 if (!isipv6)
13684 #endif
13685                         ipov->ih_len = save;
13686         }
13687 #endif                          /* TCPDEBUG */
13688
13689         /* Log to the black box */
13690         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13691                 union tcp_log_stackspecific log;
13692
13693                 bbr_fill_in_logging_data(bbr, &log.u_bbr, cts);
13694                 /* Record info on type of transmission */
13695                 log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay;
13696                 log.u_bbr.flex2 = (bbr->r_recovery_bw << 3);
13697                 log.u_bbr.flex3 = maxseg;
13698                 log.u_bbr.flex4 = delay_calc;
13699                 /* Encode filled_all into the upper flex5 bit */
13700                 log.u_bbr.flex5 = bbr->rc_past_init_win;
13701                 log.u_bbr.flex5 <<= 1;
13702                 log.u_bbr.flex5 |= bbr->rc_no_pacing;
13703                 log.u_bbr.flex5 <<= 29;
13704                 if (filled_all)
13705                         log.u_bbr.flex5 |= 0x80000000;
13706                 log.u_bbr.flex5 |= tp->t_maxseg;
13707                 log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs;
13708                 log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr);
13709                 /* lets poke in the low and the high here for debugging */
13710                 log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg;
13711                 if (rsm || sack_rxmit) {
13712                         if (doing_tlp)
13713                                 log.u_bbr.flex8 = 2;
13714                         else
13715                                 log.u_bbr.flex8 = 1;
13716                 } else {
13717                         log.u_bbr.flex8 = 0;
13718                 }
13719                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
13720                     len, &log, false, NULL, NULL, 0, tv);
13721         } else {
13722                 lgb = NULL;
13723         }
13724         /*
13725          * Fill in IP length and desired time to live and send to IP level.
13726          * There should be a better way to handle ttl and tos; we could keep
13727          * them in the template, but need a way to checksum without them.
13728          */
13729         /*
13730          * m->m_pkthdr.len should have been set before cksum calcuration,
13731          * because in6_cksum() need it.
13732          */
13733 #ifdef INET6
13734         if (isipv6) {
13735                 /*
13736                  * we separately set hoplimit for every segment, since the
13737                  * user might want to change the value via setsockopt. Also,
13738                  * desired default hop limit might be changed via Neighbor
13739                  * Discovery.
13740                  */
13741                 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
13742
13743                 /*
13744                  * Set the packet size here for the benefit of DTrace
13745                  * probes. ip6_output() will set it properly; it's supposed
13746                  * to include the option header lengths as well.
13747                  */
13748                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
13749
13750                 if (V_path_mtu_discovery && maxseg > V_tcp_minmss)
13751                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13752                 else
13753                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13754
13755                 if (tp->t_state == TCPS_SYN_SENT)
13756                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
13757
13758                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
13759                 /* TODO: IPv6 IP6TOS_ECT bit on */
13760                 error = ip6_output(m, inp->in6p_outputopts,
13761                     &inp->inp_route6,
13762                     ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
13763                     NULL, NULL, inp);
13764
13765                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
13766                         mtu = inp->inp_route6.ro_nh->nh_mtu;
13767         }
13768 #endif                          /* INET6 */
13769 #if defined(INET) && defined(INET6)
13770         else
13771 #endif
13772 #ifdef INET
13773         {
13774                 ip->ip_len = htons(m->m_pkthdr.len);
13775 #ifdef INET6
13776                 if (isipv6)
13777                         ip->ip_ttl = in6_selecthlim(inp, NULL);
13778 #endif                          /* INET6 */
13779                 /*
13780                  * If we do path MTU discovery, then we set DF on every
13781                  * packet. This might not be the best thing to do according
13782                  * to RFC3390 Section 2. However the tcp hostcache migitates
13783                  * the problem so it affects only the first tcp connection
13784                  * with a host.
13785                  *
13786                  * NB: Don't set DF on small MTU/MSS to have a safe
13787                  * fallback.
13788                  */
13789                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
13790                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
13791                         if (tp->t_port == 0 || len < V_tcp_minmss) {
13792                                 ip->ip_off |= htons(IP_DF);
13793                         }
13794                 } else {
13795                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
13796                 }
13797
13798                 if (tp->t_state == TCPS_SYN_SENT)
13799                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
13800
13801                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
13802
13803                 error = ip_output(m, inp->inp_options, &inp->inp_route,
13804                     ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
13805                     inp);
13806                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
13807                         mtu = inp->inp_route.ro_nh->nh_mtu;
13808         }
13809 #endif                          /* INET */
13810 out:
13811
13812         if (lgb) {
13813                 lgb->tlb_errno = error;
13814                 lgb = NULL;
13815         }
13816         /*
13817          * In transmit state, time the transmission and arrange for the
13818          * retransmit.  In persist state, just set snd_max.
13819          */
13820         if (error == 0) {
13821                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
13822                     (tp->t_flags & TF_SACK_PERMIT) &&
13823                     tp->rcv_numsacks > 0)
13824                         tcp_clean_dsack_blocks(tp);
13825                 /* We sent an ack clear the bbr_segs_rcvd count */
13826                 bbr->output_error_seen = 0;
13827                 bbr->oerror_cnt = 0;
13828                 bbr->bbr_segs_rcvd = 0;
13829                 if (len == 0)
13830                         counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1);
13831                 /* Do accounting for new sends */
13832                 if ((len > 0) && (rsm == NULL)) {
13833                         int idx;
13834                         if (tp->snd_una == tp->snd_max) {
13835                                 /*
13836                                  * Special case to match google, when
13837                                  * nothing is in flight the delivered
13838                                  * time does get updated to the current
13839                                  * time (see tcp_rate_bsd.c).
13840                                  */
13841                                 bbr->r_ctl.rc_del_time = cts;
13842                         }
13843                         if (len >= maxseg) {
13844                                 idx = (len / maxseg) + 3;
13845                                 if (idx >= TCP_MSS_ACCT_ATIMER)
13846                                         counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1);
13847                                 else
13848                                         counter_u64_add(bbr_out_size[idx], 1);
13849                         } else {
13850                                 /* smaller than a MSS */
13851                                 idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options);
13852                                 if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV)
13853                                         idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1);
13854                                 counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1);
13855                         }
13856                 }
13857         }
13858         abandon = 0;
13859         /*
13860          * We must do the send accounting before we log the output,
13861          * otherwise the state of the rsm could change and we account to the
13862          * wrong bucket.
13863          */
13864         if (len > 0) {
13865                 bbr_do_send_accounting(tp, bbr, rsm, len, error);
13866                 if (error == 0) {
13867                         if (tp->snd_una == tp->snd_max)
13868                                 bbr->r_ctl.rc_tlp_rxt_last_time = cts;
13869                 }
13870         }
13871         bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error,
13872             cts, mb, &abandon, rsm, 0, sb);
13873         if (abandon) {
13874                 /*
13875                  * If bbr_log_output destroys the TCB or sees a TH_RST being
13876                  * sent we should hit this condition.
13877                  */
13878                 return (0);
13879         }
13880         if (bbr->rc_in_persist == 0) {
13881                 /*
13882                  * Advance snd_nxt over sequence space of this segment.
13883                  */
13884                 if (error)
13885                         /* We don't log or do anything with errors */
13886                         goto skip_upd;
13887
13888                 if (tp->snd_una == tp->snd_max &&
13889                     (len || (flags & (TH_SYN | TH_FIN)))) {
13890                         /*
13891                          * Update the time we just added data since none was
13892                          * outstanding.
13893                          */
13894                         bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13895                         bbr->rc_tp->t_acktime  = ticks;
13896                 }
13897                 if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) {
13898                         if (flags & TH_SYN) {
13899                                 /*
13900                                  * Smack the snd_max to iss + 1
13901                                  * if its a FO we will add len below.
13902                                  */
13903                                 tp->snd_max = tp->iss + 1;
13904                         }
13905                         if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13906                                 tp->snd_max++;
13907                                 tp->t_flags |= TF_SENTFIN;
13908                         }
13909                 }
13910                 if (sack_rxmit == 0)
13911                         tp->snd_max += len;
13912 skip_upd:
13913                 if ((error == 0) && len)
13914                         tot_len += len;
13915         } else {
13916                 /* Persists case */
13917                 int32_t xlen = len;
13918
13919                 if (error)
13920                         goto nomore;
13921
13922                 if (flags & TH_SYN)
13923                         ++xlen;
13924                 if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) {
13925                         ++xlen;
13926                         tp->t_flags |= TF_SENTFIN;
13927                 }
13928                 if (xlen && (tp->snd_una == tp->snd_max)) {
13929                         /*
13930                          * Update the time we just added data since none was
13931                          * outstanding.
13932                          */
13933                         bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__);
13934                         bbr->rc_tp->t_acktime = ticks;
13935                 }
13936                 if (sack_rxmit == 0)
13937                         tp->snd_max += xlen;
13938                 tot_len += (len + optlen + ipoptlen);
13939         }
13940 nomore:
13941         if (error) {
13942                 /*
13943                  * Failures do not advance the seq counter above. For the
13944                  * case of ENOBUFS we will fall out and become ack-clocked.
13945                  * capping the cwnd at the current flight.
13946                  * Everything else will just have to retransmit with the timer
13947                  * (no pacer).
13948                  */
13949                 SOCKBUF_UNLOCK_ASSERT(sb);
13950                 BBR_STAT_INC(bbr_saw_oerr);
13951                 /* Clear all delay/early tracks */
13952                 bbr->r_ctl.rc_hptsi_agg_delay = 0;
13953                 bbr->r_ctl.rc_agg_early = 0;
13954                 bbr->r_agg_early_set = 0;
13955                 bbr->output_error_seen = 1;
13956                 if (bbr->oerror_cnt < 0xf)
13957                         bbr->oerror_cnt++;
13958                 if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) {
13959                         /* drop the session */
13960                         tcp_set_inp_to_drop(inp, ENETDOWN);
13961                 }
13962                 switch (error) {
13963                 case ENOBUFS:
13964                         /*
13965                          * Make this guy have to get ack's to send
13966                          * more but lets make sure we don't
13967                          * slam him below a T-O (1MSS).
13968                          */
13969                         if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) {
13970                                 tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
13971                                                                     bbr->r_ctl.rc_lost_bytes)) - maxseg;
13972                                 if (tp->snd_cwnd < maxseg)
13973                                         tp->snd_cwnd = maxseg;
13974                         }
13975                         slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt;
13976                         BBR_STAT_INC(bbr_saw_enobuf);
13977                         if (bbr->bbr_hdrw_pacing)
13978                                 counter_u64_add(bbr_hdwr_pacing_enobuf, 1);
13979                         else
13980                                 counter_u64_add(bbr_nohdwr_pacing_enobuf, 1);
13981                         /*
13982                          * Here even in the enobuf's case we want to do our
13983                          * state update. The reason being we may have been
13984                          * called by the input function. If so we have had
13985                          * things change.
13986                          */
13987                         error = 0;
13988                         goto enobufs;
13989                 case EMSGSIZE:
13990                         /*
13991                          * For some reason the interface we used initially
13992                          * to send segments changed to another or lowered
13993                          * its MTU. If TSO was active we either got an
13994                          * interface without TSO capabilits or TSO was
13995                          * turned off. If we obtained mtu from ip_output()
13996                          * then update it and try again.
13997                          */
13998                         /* Turn on tracing (or try to) */
13999                         {
14000                                 int old_maxseg;
14001
14002                                 old_maxseg = tp->t_maxseg;
14003                                 BBR_STAT_INC(bbr_saw_emsgsiz);
14004                                 bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts);
14005                                 if (mtu != 0)
14006                                         tcp_mss_update(tp, -1, mtu, NULL, NULL);
14007                                 if (old_maxseg <= tp->t_maxseg) {
14008                                         /* Huh it did not shrink? */
14009                                         tp->t_maxseg = old_maxseg - 40;
14010                                         bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts);
14011                                 }
14012                                 /*
14013                                  * Nuke all other things that can interfere
14014                                  * with slot
14015                                  */
14016                                 if ((tot_len + len) && (len >= tp->t_maxseg)) {
14017                                         slot = bbr_get_pacing_delay(bbr,
14018                                             bbr->r_ctl.rc_bbr_hptsi_gain,
14019                                             (tot_len + len), cts, 0);
14020                                         if (slot < bbr_error_base_paceout)
14021                                                 slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
14022                                 } else
14023                                         slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt;
14024                                 bbr->rc_output_starts_timer = 1;
14025                                 bbr_start_hpts_timer(bbr, tp, cts, 10, slot,
14026                                     tot_len);
14027                                 return (error);
14028                         }
14029                 case EPERM:
14030                         tp->t_softerror = error;
14031                         /* Fall through */
14032                 case EHOSTDOWN:
14033                 case EHOSTUNREACH:
14034                 case ENETDOWN:
14035                 case ENETUNREACH:
14036                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
14037                                 tp->t_softerror = error;
14038                         }
14039                         /* FALLTHROUGH */
14040                 default:
14041                         slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt;
14042                         bbr->rc_output_starts_timer = 1;
14043                         bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0);
14044                         return (error);
14045                 }
14046 #ifdef STATS
14047         } else if (((tp->t_flags & TF_GPUTINPROG) == 0) &&
14048                     len &&
14049                     (rsm == NULL) &&
14050             (bbr->rc_in_persist == 0)) {
14051                 tp->gput_seq = bbr_seq;
14052                 tp->gput_ack = bbr_seq +
14053                     min(sbavail(&so->so_snd) - sb_offset, sendwin);
14054                 tp->gput_ts = cts;
14055                 tp->t_flags |= TF_GPUTINPROG;
14056 #endif
14057         }
14058         KMOD_TCPSTAT_INC(tcps_sndtotal);
14059         if ((bbr->bbr_hdw_pace_ena) &&
14060             (bbr->bbr_attempt_hdwr_pace == 0) &&
14061             (bbr->rc_past_init_win) &&
14062             (bbr->rc_bbr_state != BBR_STATE_STARTUP) &&
14063             (get_filter_value(&bbr->r_ctl.rc_delrate)) &&
14064             (inp->inp_route.ro_nh &&
14065              inp->inp_route.ro_nh->nh_ifp)) {
14066                 /*
14067                  * We are past the initial window and
14068                  * have at least one measurement so we
14069                  * could use hardware pacing if its available.
14070                  * We have an interface and we have not attempted
14071                  * to setup hardware pacing, lets try to now.
14072                  */
14073                 uint64_t rate_wanted;
14074                 int err = 0;
14075
14076                 rate_wanted = bbr_get_hardware_rate(bbr);
14077                 bbr->bbr_attempt_hdwr_pace = 1;
14078                 bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp,
14079                                                       inp->inp_route.ro_nh->nh_ifp,
14080                                                       rate_wanted,
14081                                                       (RS_PACING_GEQ|RS_PACING_SUB_OK),
14082                                                       &err);
14083                 if (bbr->r_ctl.crte) {
14084                         bbr_type_log_hdwr_pacing(bbr,
14085                                                  bbr->r_ctl.crte->ptbl->rs_ifp,
14086                                                  rate_wanted,
14087                                                  bbr->r_ctl.crte->rate,
14088                                                  __LINE__, cts, err);
14089                         BBR_STAT_INC(bbr_hdwr_rl_add_ok);
14090                         counter_u64_add(bbr_flows_nohdwr_pacing, -1);
14091                         counter_u64_add(bbr_flows_whdwr_pacing, 1);
14092                         bbr->bbr_hdrw_pacing = 1;
14093                         /* Now what is our gain status? */
14094                         if (bbr->r_ctl.crte->rate < rate_wanted) {
14095                                 /* We have a problem */
14096                                 bbr_setup_less_of_rate(bbr, cts,
14097                                                        bbr->r_ctl.crte->rate, rate_wanted);
14098                         } else {
14099                                 /* We are good */
14100                                 bbr->gain_is_limited = 0;
14101                                 bbr->skip_gain = 0;
14102                         }
14103                         tcp_bbr_tso_size_check(bbr, cts);
14104                 } else {
14105                         bbr_type_log_hdwr_pacing(bbr,
14106                                                  inp->inp_route.ro_nh->nh_ifp,
14107                                                  rate_wanted,
14108                                                  0,
14109                                                  __LINE__, cts, err);
14110                         BBR_STAT_INC(bbr_hdwr_rl_add_fail);
14111                 }
14112         }
14113         if (bbr->bbr_hdrw_pacing) {
14114                 /*
14115                  * Worry about cases where the route
14116                  * changes or something happened that we
14117                  * lost our hardware pacing possibly during
14118                  * the last ip_output call.
14119                  */
14120                 if (inp->inp_snd_tag == NULL) {
14121                         /* A change during ip output disabled hw pacing? */
14122                         bbr->bbr_hdrw_pacing = 0;
14123                 } else if ((inp->inp_route.ro_nh == NULL) ||
14124                     (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) {
14125                         /*
14126                          * We had an interface or route change,
14127                          * detach from the current hdwr pacing
14128                          * and setup to re-attempt next go
14129                          * round.
14130                          */
14131                         bbr->bbr_hdrw_pacing = 0;
14132                         bbr->bbr_attempt_hdwr_pace = 0;
14133                         tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp);
14134                         tcp_bbr_tso_size_check(bbr, cts);
14135                 }
14136         }
14137         /*
14138          * Data sent (as far as we can tell). If this advertises a larger
14139          * window than any other segment, then remember the size of the
14140          * advertised window. Any pending ACK has now been sent.
14141          */
14142         if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
14143                 tp->rcv_adv = tp->rcv_nxt + recwin;
14144
14145         tp->last_ack_sent = tp->rcv_nxt;
14146         if ((error == 0) &&
14147             (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) &&
14148             (doing_tlp == 0) &&
14149             (tso == 0) &&
14150             (len > 0) &&
14151             ((flags & TH_RST) == 0) &&
14152             ((flags & TH_SYN) == 0) &&
14153             (IN_RECOVERY(tp->t_flags) == 0) &&
14154             (bbr->rc_in_persist == 0) &&
14155             (tot_len < bbr->r_ctl.rc_pace_max_segs)) {
14156                 /*
14157                  * For non-tso we need to goto again until we have sent out
14158                  * enough data to match what we are hptsi out every hptsi
14159                  * interval.
14160                  */
14161                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14162                         /* Make sure snd_nxt is drug up */
14163                         tp->snd_nxt = tp->snd_max;
14164                 }
14165                 if (rsm != NULL) {
14166                         rsm = NULL;
14167                         goto skip_again;
14168                 }
14169                 rsm = NULL;
14170                 sack_rxmit = 0;
14171                 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14172                 goto again;
14173         }
14174 skip_again:
14175         if ((error == 0) && (flags & TH_FIN))
14176                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
14177         if ((error == 0) && (flags & TH_RST))
14178                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
14179         if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) {
14180                 /*
14181                  * Calculate/Re-Calculate the hptsi slot in usecs based on
14182                  * what we have sent so far
14183                  */
14184                 slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0);
14185                 if (bbr->rc_no_pacing)
14186                         slot = 0;
14187         }
14188         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
14189 enobufs:
14190         if (bbr->rc_use_google == 0)
14191                 bbr_check_bbr_for_state(bbr, cts, __LINE__, 0);
14192         bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked +
14193                                                         bbr->r_ctl.rc_lost_bytes)));
14194         bbr->rc_output_starts_timer = 1;
14195         if (bbr->bbr_use_rack_cheat &&
14196             (more_to_rxt ||
14197              ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) {
14198                 /* Rack cheats and shotguns out all rxt's 1ms apart */
14199                 if (slot > 1000)
14200                         slot = 1000;
14201         }
14202         if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) {
14203                 /*
14204                  * We don't change the tso size until some number of sends
14205                  * to give the hardware commands time to get down
14206                  * to the interface.
14207                  */
14208                 bbr->r_ctl.bbr_hdwr_cnt_noset_snt++;
14209                 if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) {
14210                         bbr->hw_pacing_set = 1;
14211                         tcp_bbr_tso_size_check(bbr, cts);
14212                 }
14213         }
14214         bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len);
14215         if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
14216                 /* Make sure snd_nxt is drug up */
14217                 tp->snd_nxt = tp->snd_max;
14218         }
14219         return (error);
14220
14221 }
14222
14223 /*
14224  * See bbr_output_wtime() for return values.
14225  */
14226 static int
14227 bbr_output(struct tcpcb *tp)
14228 {
14229         int32_t ret;
14230         struct timeval tv;
14231         struct tcp_bbr *bbr;
14232
14233         NET_EPOCH_ASSERT();
14234
14235         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14236         INP_WLOCK_ASSERT(tp->t_inpcb);
14237         (void)tcp_get_usecs(&tv);
14238         ret = bbr_output_wtime(tp, &tv);
14239         return (ret);
14240 }
14241
14242 static void
14243 bbr_mtu_chg(struct tcpcb *tp)
14244 {
14245         struct tcp_bbr *bbr;
14246         struct bbr_sendmap *rsm, *frsm = NULL;
14247         uint32_t maxseg;
14248
14249         /*
14250          * The MTU has changed. a) Clear the sack filter. b) Mark everything
14251          * over the current size as SACK_PASS so a retransmit will occur.
14252          */
14253
14254         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14255         maxseg = tp->t_maxseg - bbr->rc_last_options;
14256         sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una);
14257         TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) {
14258                 /* Don't mess with ones acked (by sack?) */
14259                 if (rsm->r_flags & BBR_ACKED)
14260                         continue;
14261                 if ((rsm->r_end - rsm->r_start) > maxseg) {
14262                         /*
14263                          * We mark sack-passed on all the previous large
14264                          * sends we did. This will force them to retransmit.
14265                          */
14266                         rsm->r_flags |= BBR_SACK_PASSED;
14267                         if (((rsm->r_flags & BBR_MARKED_LOST) == 0) &&
14268                             bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) {
14269                                 bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start;
14270                                 bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start;
14271                                 rsm->r_flags |= BBR_MARKED_LOST;
14272                         }
14273                         if (frsm == NULL)
14274                                 frsm = rsm;
14275                 }
14276         }
14277         if (frsm) {
14278                 bbr->r_ctl.rc_resend = frsm;
14279         }
14280 }
14281
14282 /*
14283  * bbr_ctloutput() must drop the inpcb lock before performing copyin on
14284  * socket option arguments.  When it re-acquires the lock after the copy, it
14285  * has to revalidate that the connection is still valid for the socket
14286  * option.
14287  */
14288 static int
14289 bbr_set_sockopt(struct socket *so, struct sockopt *sopt,
14290                 struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14291 {
14292         struct epoch_tracker et;
14293         int32_t error = 0, optval;
14294
14295         switch (sopt->sopt_name) {
14296         case TCP_RACK_PACE_MAX_SEG:
14297         case TCP_RACK_MIN_TO:
14298         case TCP_RACK_REORD_THRESH:
14299         case TCP_RACK_REORD_FADE:
14300         case TCP_RACK_TLP_THRESH:
14301         case TCP_RACK_PKT_DELAY:
14302         case TCP_BBR_ALGORITHM:
14303         case TCP_BBR_TSLIMITS:
14304         case TCP_BBR_IWINTSO:
14305         case TCP_BBR_RECFORCE:
14306         case TCP_BBR_STARTUP_PG:
14307         case TCP_BBR_DRAIN_PG:
14308         case TCP_BBR_RWND_IS_APP:
14309         case TCP_BBR_PROBE_RTT_INT:
14310         case TCP_BBR_PROBE_RTT_GAIN:
14311         case TCP_BBR_PROBE_RTT_LEN:
14312         case TCP_BBR_STARTUP_LOSS_EXIT:
14313         case TCP_BBR_USEDEL_RATE:
14314         case TCP_BBR_MIN_RTO:
14315         case TCP_BBR_MAX_RTO:
14316         case TCP_BBR_PACE_PER_SEC:
14317         case TCP_DELACK:
14318         case TCP_BBR_PACE_DEL_TAR:
14319         case TCP_BBR_SEND_IWND_IN_TSO:
14320         case TCP_BBR_EXTRA_STATE:
14321         case TCP_BBR_UTTER_MAX_TSO:
14322         case TCP_BBR_MIN_TOPACEOUT:
14323         case TCP_BBR_FLOOR_MIN_TSO:
14324         case TCP_BBR_TSTMP_RAISES:
14325         case TCP_BBR_POLICER_DETECT:
14326         case TCP_BBR_USE_RACK_CHEAT:
14327         case TCP_DATA_AFTER_CLOSE:
14328         case TCP_BBR_HDWR_PACE:
14329         case TCP_BBR_PACE_SEG_MAX:
14330         case TCP_BBR_PACE_SEG_MIN:
14331         case TCP_BBR_PACE_CROSS:
14332         case TCP_BBR_PACE_OH:
14333 #ifdef NETFLIX_PEAKRATE
14334         case TCP_MAXPEAKRATE:
14335 #endif
14336         case TCP_BBR_TMR_PACE_OH:
14337         case TCP_BBR_RACK_RTT_USE:
14338         case TCP_BBR_RETRAN_WTSO:
14339                 break;
14340         default:
14341                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14342                 break;
14343         }
14344         INP_WUNLOCK(inp);
14345         error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
14346         if (error)
14347                 return (error);
14348         INP_WLOCK(inp);
14349         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
14350                 INP_WUNLOCK(inp);
14351                 return (ECONNRESET);
14352         }
14353         tp = intotcpcb(inp);
14354         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14355         switch (sopt->sopt_name) {
14356         case TCP_BBR_PACE_PER_SEC:
14357                 BBR_OPTS_INC(tcp_bbr_pace_per_sec);
14358                 bbr->r_ctl.bbr_hptsi_per_second = optval;
14359                 break;
14360         case TCP_BBR_PACE_DEL_TAR:
14361                 BBR_OPTS_INC(tcp_bbr_pace_del_tar);
14362                 bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval;
14363                 break;
14364         case TCP_BBR_PACE_SEG_MAX:
14365                 BBR_OPTS_INC(tcp_bbr_pace_seg_max);
14366                 bbr->r_ctl.bbr_hptsi_segments_max = optval;
14367                 break;
14368         case TCP_BBR_PACE_SEG_MIN:
14369                 BBR_OPTS_INC(tcp_bbr_pace_seg_min);
14370                 bbr->r_ctl.bbr_hptsi_bytes_min = optval;
14371                 break;
14372         case TCP_BBR_PACE_CROSS:
14373                 BBR_OPTS_INC(tcp_bbr_pace_cross);
14374                 bbr->r_ctl.bbr_cross_over = optval;
14375                 break;
14376         case TCP_BBR_ALGORITHM:
14377                 BBR_OPTS_INC(tcp_bbr_algorithm);
14378                 if (optval && (bbr->rc_use_google == 0)) {
14379                         /* Turn on the google mode */
14380                         bbr_google_mode_on(bbr);
14381                         if ((optval > 3) && (optval < 500)) {
14382                                 /*
14383                                  * Must be at least greater than .3%
14384                                  * and must be less than 50.0%.
14385                                  */
14386                                 bbr->r_ctl.bbr_google_discount = optval;
14387                         }
14388                 } else if ((optval == 0) && (bbr->rc_use_google == 1)) {
14389                         /* Turn off the google mode */
14390                         bbr_google_mode_off(bbr);
14391                 }
14392                 break;
14393         case TCP_BBR_TSLIMITS:
14394                 BBR_OPTS_INC(tcp_bbr_tslimits);
14395                 if (optval == 1)
14396                         bbr->rc_use_ts_limit = 1;
14397                 else if (optval == 0)
14398                         bbr->rc_use_ts_limit = 0;
14399                 else
14400                         error = EINVAL;
14401                 break;
14402
14403         case TCP_BBR_IWINTSO:
14404                 BBR_OPTS_INC(tcp_bbr_iwintso);
14405                 if ((optval >= 0) && (optval < 128)) {
14406                         uint32_t twin;
14407
14408                         bbr->rc_init_win = optval;
14409                         twin = bbr_initial_cwnd(bbr, tp);
14410                         if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd))
14411                                 tp->snd_cwnd = twin;
14412                         else
14413                                 error = EBUSY;
14414                 } else
14415                         error = EINVAL;
14416                 break;
14417         case TCP_BBR_STARTUP_PG:
14418                 BBR_OPTS_INC(tcp_bbr_startup_pg);
14419                 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) {
14420                         bbr->r_ctl.rc_startup_pg = optval;
14421                         if (bbr->rc_bbr_state == BBR_STATE_STARTUP) {
14422                                 bbr->r_ctl.rc_bbr_hptsi_gain = optval;
14423                         }
14424                 } else
14425                         error = EINVAL;
14426                 break;
14427         case TCP_BBR_DRAIN_PG:
14428                 BBR_OPTS_INC(tcp_bbr_drain_pg);
14429                 if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE))
14430                         bbr->r_ctl.rc_drain_pg = optval;
14431                 else
14432                         error = EINVAL;
14433                 break;
14434         case TCP_BBR_PROBE_RTT_LEN:
14435                 BBR_OPTS_INC(tcp_bbr_probertt_len);
14436                 if (optval <= 1)
14437                         reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND));
14438                 else
14439                         error = EINVAL;
14440                 break;
14441         case TCP_BBR_PROBE_RTT_GAIN:
14442                 BBR_OPTS_INC(tcp_bbr_probertt_gain);
14443                 if (optval <= BBR_UNIT)
14444                         bbr->r_ctl.bbr_rttprobe_gain_val = optval;
14445                 else
14446                         error = EINVAL;
14447                 break;
14448         case TCP_BBR_PROBE_RTT_INT:
14449                 BBR_OPTS_INC(tcp_bbr_probe_rtt_int);
14450                 if (optval > 1000)
14451                         bbr->r_ctl.rc_probertt_int = optval;
14452                 else
14453                         error = EINVAL;
14454                 break;
14455         case TCP_BBR_MIN_TOPACEOUT:
14456                 BBR_OPTS_INC(tcp_bbr_topaceout);
14457                 if (optval == 0) {
14458                         bbr->no_pacing_until = 0;
14459                         bbr->rc_no_pacing = 0;
14460                 } else if (optval <= 0x00ff) {
14461                         bbr->no_pacing_until = optval;
14462                         if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) &&
14463                             (bbr->rc_bbr_state == BBR_STATE_STARTUP)){
14464                                 /* Turn on no pacing */
14465                                 bbr->rc_no_pacing = 1;
14466                         }
14467                 } else
14468                         error = EINVAL;
14469                 break;
14470         case TCP_BBR_STARTUP_LOSS_EXIT:
14471                 BBR_OPTS_INC(tcp_bbr_startup_loss_exit);
14472                 bbr->rc_loss_exit = optval;
14473                 break;
14474         case TCP_BBR_USEDEL_RATE:
14475                 error = EINVAL;
14476                 break;
14477         case TCP_BBR_MIN_RTO:
14478                 BBR_OPTS_INC(tcp_bbr_min_rto);
14479                 bbr->r_ctl.rc_min_rto_ms = optval;
14480                 break;
14481         case TCP_BBR_MAX_RTO:
14482                 BBR_OPTS_INC(tcp_bbr_max_rto);
14483                 bbr->rc_max_rto_sec = optval;
14484                 break;
14485         case TCP_RACK_MIN_TO:
14486                 /* Minimum time between rack t-o's in ms */
14487                 BBR_OPTS_INC(tcp_rack_min_to);
14488                 bbr->r_ctl.rc_min_to = optval;
14489                 break;
14490         case TCP_RACK_REORD_THRESH:
14491                 /* RACK reorder threshold (shift amount) */
14492                 BBR_OPTS_INC(tcp_rack_reord_thresh);
14493                 if ((optval > 0) && (optval < 31))
14494                         bbr->r_ctl.rc_reorder_shift = optval;
14495                 else
14496                         error = EINVAL;
14497                 break;
14498         case TCP_RACK_REORD_FADE:
14499                 /* Does reordering fade after ms time */
14500                 BBR_OPTS_INC(tcp_rack_reord_fade);
14501                 bbr->r_ctl.rc_reorder_fade = optval;
14502                 break;
14503         case TCP_RACK_TLP_THRESH:
14504                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
14505                 BBR_OPTS_INC(tcp_rack_tlp_thresh);
14506                 if (optval)
14507                         bbr->rc_tlp_threshold = optval;
14508                 else
14509                         error = EINVAL;
14510                 break;
14511         case TCP_BBR_USE_RACK_CHEAT:
14512                 BBR_OPTS_INC(tcp_use_rackcheat);
14513                 if (bbr->rc_use_google) {
14514                         error = EINVAL;
14515                         break;
14516                 }
14517                 BBR_OPTS_INC(tcp_rack_cheat);
14518                 if (optval)
14519                         bbr->bbr_use_rack_cheat = 1;
14520                 else
14521                         bbr->bbr_use_rack_cheat = 0;
14522                 break;
14523         case TCP_BBR_FLOOR_MIN_TSO:
14524                 BBR_OPTS_INC(tcp_utter_max_tso);
14525                 if ((optval >= 0) && (optval < 40))
14526                         bbr->r_ctl.bbr_hptsi_segments_floor = optval;
14527                 else
14528                         error = EINVAL;
14529                 break;
14530         case TCP_BBR_UTTER_MAX_TSO:
14531                 BBR_OPTS_INC(tcp_utter_max_tso);
14532                 if ((optval >= 0) && (optval < 0xffff))
14533                         bbr->r_ctl.bbr_utter_max = optval;
14534                 else
14535                         error = EINVAL;
14536                 break;
14537
14538         case TCP_BBR_EXTRA_STATE:
14539                 BBR_OPTS_INC(tcp_extra_state);
14540                 if (optval)
14541                         bbr->rc_use_idle_restart = 1;
14542                 else
14543                         bbr->rc_use_idle_restart = 0;
14544                 break;
14545         case TCP_BBR_SEND_IWND_IN_TSO:
14546                 BBR_OPTS_INC(tcp_iwnd_tso);
14547                 if (optval) {
14548                         bbr->bbr_init_win_cheat = 1;
14549                         if (bbr->rc_past_init_win == 0) {
14550                                 uint32_t cts;
14551                                 cts = tcp_get_usecs(&bbr->rc_tv);
14552                                 tcp_bbr_tso_size_check(bbr, cts);
14553                         }
14554                 } else
14555                         bbr->bbr_init_win_cheat = 0;
14556                 break;
14557         case TCP_BBR_HDWR_PACE:
14558                 BBR_OPTS_INC(tcp_hdwr_pacing);
14559                 if (optval){
14560                         bbr->bbr_hdw_pace_ena = 1;
14561                         bbr->bbr_attempt_hdwr_pace = 0;
14562                 } else {
14563                         bbr->bbr_hdw_pace_ena = 0;
14564 #ifdef RATELIMIT
14565                         if (bbr->bbr_hdrw_pacing) {
14566                                 bbr->bbr_hdrw_pacing = 0;
14567                                 in_pcbdetach_txrtlmt(bbr->rc_inp);
14568                         }
14569 #endif
14570                 }
14571                 break;
14572
14573         case TCP_DELACK:
14574                 BBR_OPTS_INC(tcp_delack);
14575                 if (optval < 100) {
14576                         if (optval == 0) /* off */
14577                                 tp->t_delayed_ack = 0;
14578                         else if (optval == 1) /* on which is 2 */
14579                                 tp->t_delayed_ack = 2;
14580                         else /* higher than 2 and less than 100 */
14581                                 tp->t_delayed_ack = optval;
14582                         if (tp->t_flags & TF_DELACK) {
14583                                 tp->t_flags &= ~TF_DELACK;
14584                                 tp->t_flags |= TF_ACKNOW;
14585                                 NET_EPOCH_ENTER(et);
14586                                 bbr_output(tp);
14587                                 NET_EPOCH_EXIT(et);
14588                         }
14589                 } else
14590                         error = EINVAL;
14591                 break;
14592         case TCP_RACK_PKT_DELAY:
14593                 /* RACK added ms i.e. rack-rtt + reord + N */
14594                 BBR_OPTS_INC(tcp_rack_pkt_delay);
14595                 bbr->r_ctl.rc_pkt_delay = optval;
14596                 break;
14597 #ifdef NETFLIX_PEAKRATE
14598         case TCP_MAXPEAKRATE:
14599                 BBR_OPTS_INC(tcp_maxpeak);
14600                 error = tcp_set_maxpeakrate(tp, optval);
14601                 if (!error)
14602                         tp->t_peakrate_thr = tp->t_maxpeakrate;
14603                 break;
14604 #endif
14605         case TCP_BBR_RETRAN_WTSO:
14606                 BBR_OPTS_INC(tcp_retran_wtso);
14607                 if (optval)
14608                         bbr->rc_resends_use_tso = 1;
14609                 else
14610                         bbr->rc_resends_use_tso = 0;
14611                 break;
14612         case TCP_DATA_AFTER_CLOSE:
14613                 BBR_OPTS_INC(tcp_data_ac);
14614                 if (optval)
14615                         bbr->rc_allow_data_af_clo = 1;
14616                 else
14617                         bbr->rc_allow_data_af_clo = 0;
14618                 break;
14619         case TCP_BBR_POLICER_DETECT:
14620                 BBR_OPTS_INC(tcp_policer_det);
14621                 if (bbr->rc_use_google == 0)
14622                         error = EINVAL;
14623                 else if (optval)
14624                         bbr->r_use_policer = 1;
14625                 else
14626                         bbr->r_use_policer = 0;
14627                 break;
14628
14629         case TCP_BBR_TSTMP_RAISES:
14630                 BBR_OPTS_INC(tcp_ts_raises);
14631                 if (optval)
14632                         bbr->ts_can_raise = 1;
14633                 else
14634                         bbr->ts_can_raise = 0;
14635                 break;
14636         case TCP_BBR_TMR_PACE_OH:
14637                 BBR_OPTS_INC(tcp_pacing_oh_tmr);
14638                 if (bbr->rc_use_google) {
14639                         error = EINVAL;
14640                 } else {
14641                         if (optval)
14642                                 bbr->r_ctl.rc_incr_tmrs = 1;
14643                         else
14644                                 bbr->r_ctl.rc_incr_tmrs = 0;
14645                 }
14646                 break;
14647         case TCP_BBR_PACE_OH:
14648                 BBR_OPTS_INC(tcp_pacing_oh);
14649                 if (bbr->rc_use_google) {
14650                         error = EINVAL;
14651                 } else {
14652                         if (optval > (BBR_INCL_TCP_OH|
14653                                       BBR_INCL_IP_OH|
14654                                       BBR_INCL_ENET_OH)) {
14655                                 error = EINVAL;
14656                                 break;
14657                         }
14658                         if (optval & BBR_INCL_TCP_OH)
14659                                 bbr->r_ctl.rc_inc_tcp_oh = 1;
14660                         else
14661                                 bbr->r_ctl.rc_inc_tcp_oh = 0;
14662                         if (optval & BBR_INCL_IP_OH)
14663                                 bbr->r_ctl.rc_inc_ip_oh = 1;
14664                         else
14665                                 bbr->r_ctl.rc_inc_ip_oh = 0;
14666                         if (optval & BBR_INCL_ENET_OH)
14667                                 bbr->r_ctl.rc_inc_enet_oh = 1;
14668                         else
14669                                 bbr->r_ctl.rc_inc_enet_oh = 0;
14670                 }
14671                 break;
14672         default:
14673                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14674                 break;
14675         }
14676 #ifdef NETFLIX_STATS
14677         tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
14678 #endif
14679         INP_WUNLOCK(inp);
14680         return (error);
14681 }
14682
14683 /*
14684  * return 0 on success, error-num on failure
14685  */
14686 static int
14687 bbr_get_sockopt(struct socket *so, struct sockopt *sopt,
14688     struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr)
14689 {
14690         int32_t error, optval;
14691
14692         /*
14693          * Because all our options are either boolean or an int, we can just
14694          * pull everything into optval and then unlock and copy. If we ever
14695          * add a option that is not a int, then this will have quite an
14696          * impact to this routine.
14697          */
14698         switch (sopt->sopt_name) {
14699         case TCP_BBR_PACE_PER_SEC:
14700                 optval = bbr->r_ctl.bbr_hptsi_per_second;
14701                 break;
14702         case TCP_BBR_PACE_DEL_TAR:
14703                 optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar;
14704                 break;
14705         case TCP_BBR_PACE_SEG_MAX:
14706                 optval = bbr->r_ctl.bbr_hptsi_segments_max;
14707                 break;
14708         case TCP_BBR_MIN_TOPACEOUT:
14709                 optval = bbr->no_pacing_until;
14710                 break;
14711         case TCP_BBR_PACE_SEG_MIN:
14712                 optval = bbr->r_ctl.bbr_hptsi_bytes_min;
14713                 break;
14714         case TCP_BBR_PACE_CROSS:
14715                 optval = bbr->r_ctl.bbr_cross_over;
14716                 break;
14717         case TCP_BBR_ALGORITHM:
14718                 optval = bbr->rc_use_google;
14719                 break;
14720         case TCP_BBR_TSLIMITS:
14721                 optval = bbr->rc_use_ts_limit;
14722                 break;
14723         case TCP_BBR_IWINTSO:
14724                 optval = bbr->rc_init_win;
14725                 break;
14726         case TCP_BBR_STARTUP_PG:
14727                 optval = bbr->r_ctl.rc_startup_pg;
14728                 break;
14729         case TCP_BBR_DRAIN_PG:
14730                 optval = bbr->r_ctl.rc_drain_pg;
14731                 break;
14732         case TCP_BBR_PROBE_RTT_INT:
14733                 optval = bbr->r_ctl.rc_probertt_int;
14734                 break;
14735         case TCP_BBR_PROBE_RTT_LEN:
14736                 optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND);
14737                 break;
14738         case TCP_BBR_PROBE_RTT_GAIN:
14739                 optval = bbr->r_ctl.bbr_rttprobe_gain_val;
14740                 break;
14741         case TCP_BBR_STARTUP_LOSS_EXIT:
14742                 optval = bbr->rc_loss_exit;
14743                 break;
14744         case TCP_BBR_USEDEL_RATE:
14745                 error = EINVAL;
14746                 break;
14747         case TCP_BBR_MIN_RTO:
14748                 optval = bbr->r_ctl.rc_min_rto_ms;
14749                 break;
14750         case TCP_BBR_MAX_RTO:
14751                 optval = bbr->rc_max_rto_sec;
14752                 break;
14753         case TCP_RACK_PACE_MAX_SEG:
14754                 /* Max segments in a pace */
14755                 optval = bbr->r_ctl.rc_pace_max_segs;
14756                 break;
14757         case TCP_RACK_MIN_TO:
14758                 /* Minimum time between rack t-o's in ms */
14759                 optval = bbr->r_ctl.rc_min_to;
14760                 break;
14761         case TCP_RACK_REORD_THRESH:
14762                 /* RACK reorder threshold (shift amount) */
14763                 optval = bbr->r_ctl.rc_reorder_shift;
14764                 break;
14765         case TCP_RACK_REORD_FADE:
14766                 /* Does reordering fade after ms time */
14767                 optval = bbr->r_ctl.rc_reorder_fade;
14768                 break;
14769         case TCP_BBR_USE_RACK_CHEAT:
14770                 /* Do we use the rack cheat for rxt */
14771                 optval = bbr->bbr_use_rack_cheat;
14772                 break;
14773         case TCP_BBR_FLOOR_MIN_TSO:
14774                 optval = bbr->r_ctl.bbr_hptsi_segments_floor;
14775                 break;
14776         case TCP_BBR_UTTER_MAX_TSO:
14777                 optval = bbr->r_ctl.bbr_utter_max;
14778                 break;
14779         case TCP_BBR_SEND_IWND_IN_TSO:
14780                 /* Do we send TSO size segments initially */
14781                 optval = bbr->bbr_init_win_cheat;
14782                 break;
14783         case TCP_BBR_EXTRA_STATE:
14784                 optval = bbr->rc_use_idle_restart;
14785                 break;
14786         case TCP_RACK_TLP_THRESH:
14787                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
14788                 optval = bbr->rc_tlp_threshold;
14789                 break;
14790         case TCP_RACK_PKT_DELAY:
14791                 /* RACK added ms i.e. rack-rtt + reord + N */
14792                 optval = bbr->r_ctl.rc_pkt_delay;
14793                 break;
14794         case TCP_BBR_RETRAN_WTSO:
14795                 optval = bbr->rc_resends_use_tso;
14796                 break;
14797         case TCP_DATA_AFTER_CLOSE:
14798                 optval = bbr->rc_allow_data_af_clo;
14799                 break;
14800         case TCP_DELACK:
14801                 optval = tp->t_delayed_ack;
14802                 break;
14803         case TCP_BBR_HDWR_PACE:
14804                 optval = bbr->bbr_hdw_pace_ena;
14805                 break;
14806         case TCP_BBR_POLICER_DETECT:
14807                 optval = bbr->r_use_policer;
14808                 break;
14809         case TCP_BBR_TSTMP_RAISES:
14810                 optval = bbr->ts_can_raise;
14811                 break;
14812         case TCP_BBR_TMR_PACE_OH:
14813                 optval = bbr->r_ctl.rc_incr_tmrs;
14814                 break;
14815         case TCP_BBR_PACE_OH:
14816                 optval = 0;
14817                 if (bbr->r_ctl.rc_inc_tcp_oh)
14818                         optval |= BBR_INCL_TCP_OH;
14819                 if (bbr->r_ctl.rc_inc_ip_oh)
14820                         optval |= BBR_INCL_IP_OH;
14821                 if (bbr->r_ctl.rc_inc_enet_oh)
14822                         optval |= BBR_INCL_ENET_OH;
14823                 break;
14824         default:
14825                 return (tcp_default_ctloutput(so, sopt, inp, tp));
14826                 break;
14827         }
14828         INP_WUNLOCK(inp);
14829         error = sooptcopyout(sopt, &optval, sizeof optval);
14830         return (error);
14831 }
14832
14833 /*
14834  * return 0 on success, error-num on failure
14835  */
14836 static int
14837 bbr_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
14838 {
14839         int32_t error = EINVAL;
14840         struct tcp_bbr *bbr;
14841
14842         bbr = (struct tcp_bbr *)tp->t_fb_ptr;
14843         if (bbr == NULL) {
14844                 /* Huh? */
14845                 goto out;
14846         }
14847         if (sopt->sopt_dir == SOPT_SET) {
14848                 return (bbr_set_sockopt(so, sopt, inp, tp, bbr));
14849         } else if (sopt->sopt_dir == SOPT_GET) {
14850                 return (bbr_get_sockopt(so, sopt, inp, tp, bbr));
14851         }
14852 out:
14853         INP_WUNLOCK(inp);
14854         return (error);
14855 }
14856
14857 static int
14858 bbr_pru_options(struct tcpcb *tp, int flags)
14859 {
14860         if (flags & PRUS_OOB)
14861                 return (EOPNOTSUPP);
14862         return (0);
14863 }
14864
14865 struct tcp_function_block __tcp_bbr = {
14866         .tfb_tcp_block_name = __XSTRING(STACKNAME),
14867         .tfb_tcp_output = bbr_output,
14868         .tfb_do_queued_segments = ctf_do_queued_segments,
14869         .tfb_do_segment_nounlock = bbr_do_segment_nounlock,
14870         .tfb_tcp_do_segment = bbr_do_segment,
14871         .tfb_tcp_ctloutput = bbr_ctloutput,
14872         .tfb_tcp_fb_init = bbr_init,
14873         .tfb_tcp_fb_fini = bbr_fini,
14874         .tfb_tcp_timer_stop_all = bbr_stopall,
14875         .tfb_tcp_timer_activate = bbr_timer_activate,
14876         .tfb_tcp_timer_active = bbr_timer_active,
14877         .tfb_tcp_timer_stop = bbr_timer_stop,
14878         .tfb_tcp_rexmit_tmr = bbr_remxt_tmr,
14879         .tfb_tcp_handoff_ok = bbr_handoff_ok,
14880         .tfb_tcp_mtu_chg = bbr_mtu_chg,
14881         .tfb_pru_options = bbr_pru_options,
14882 };
14883
14884 static const char *bbr_stack_names[] = {
14885         __XSTRING(STACKNAME),
14886 #ifdef STACKALIAS
14887         __XSTRING(STACKALIAS),
14888 #endif
14889 };
14890
14891 static bool bbr_mod_inited = false;
14892
14893 static int
14894 tcp_addbbr(module_t mod, int32_t type, void *data)
14895 {
14896         int32_t err = 0;
14897         int num_stacks;
14898
14899         switch (type) {
14900         case MOD_LOAD:
14901                 printf("Attempting to load " __XSTRING(MODNAME) "\n");
14902                 bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
14903                     sizeof(struct bbr_sendmap),
14904                     NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
14905                 bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
14906                     sizeof(struct tcp_bbr),
14907                     NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
14908                 sysctl_ctx_init(&bbr_sysctl_ctx);
14909                 bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx,
14910                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
14911                     OID_AUTO,
14912 #ifdef STACKALIAS
14913                     __XSTRING(STACKALIAS),
14914 #else
14915                     __XSTRING(STACKNAME),
14916 #endif
14917                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
14918                     "");
14919                 if (bbr_sysctl_root == NULL) {
14920                         printf("Failed to add sysctl node\n");
14921                         err = EFAULT;
14922                         goto free_uma;
14923                 }
14924                 bbr_init_sysctls();
14925                 num_stacks = nitems(bbr_stack_names);
14926                 err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK,
14927                     bbr_stack_names, &num_stacks);
14928                 if (err) {
14929                         printf("Failed to register %s stack name for "
14930                             "%s module\n", bbr_stack_names[num_stacks],
14931                             __XSTRING(MODNAME));
14932                         sysctl_ctx_free(&bbr_sysctl_ctx);
14933         free_uma:
14934                         uma_zdestroy(bbr_zone);
14935                         uma_zdestroy(bbr_pcb_zone);
14936                         bbr_counter_destroy();
14937                         printf("Failed to register " __XSTRING(MODNAME)
14938                             " module err:%d\n", err);
14939                         return (err);
14940                 }
14941                 tcp_lro_reg_mbufq();
14942                 bbr_mod_inited = true;
14943                 printf(__XSTRING(MODNAME) " is now available\n");
14944                 break;
14945         case MOD_QUIESCE:
14946                 err = deregister_tcp_functions(&__tcp_bbr, true, false);
14947                 break;
14948         case MOD_UNLOAD:
14949                 err = deregister_tcp_functions(&__tcp_bbr, false, true);
14950                 if (err == EBUSY)
14951                         break;
14952                 if (bbr_mod_inited) {
14953                         uma_zdestroy(bbr_zone);
14954                         uma_zdestroy(bbr_pcb_zone);
14955                         sysctl_ctx_free(&bbr_sysctl_ctx);
14956                         bbr_counter_destroy();
14957                         printf(__XSTRING(MODNAME)
14958                             " is now no longer available\n");
14959                         bbr_mod_inited = false;
14960                 }
14961                 tcp_lro_dereg_mbufq();
14962                 err = 0;
14963                 break;
14964         default:
14965                 return (EOPNOTSUPP);
14966         }
14967         return (err);
14968 }
14969
14970 static moduledata_t tcp_bbr = {
14971         .name = __XSTRING(MODNAME),
14972             .evhand = tcp_addbbr,
14973             .priv = 0
14974 };
14975
14976 MODULE_VERSION(MODNAME, 1);
14977 DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
14978 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);