sys/netinet/tcp_stacks/rack.c

   1 /*-
   2  * Copyright (c) 2016-2020 Netflix, Inc.
   3  *
   4  * Redistribution and use in source and binary forms, with or without
   5  * modification, are permitted provided that the following conditions
   6  * are met:
   7  * 1. Redistributions of source code must retain the above copyright
   8  *    notice, this list of conditions and the following disclaimer.
   9  * 2. Redistributions in binary form must reproduce the above copyright
  10  *    notice, this list of conditions and the following disclaimer in the
  11  *    documentation and/or other materials provided with the distribution.
  12  *
  13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  23  * SUCH DAMAGE.
  24  *
  25  */
  26
  27 #include <sys/cdefs.h>
  28 __FBSDID("$FreeBSD$");
  29
  30 #include "opt_inet.h"
  31 #include "opt_inet6.h"
  32 #include "opt_ipsec.h"
  33 #include "opt_ratelimit.h"
  34 #include "opt_kern_tls.h"
  35 #if defined(INET) || defined(INET6)
  36 #include <sys/param.h>
  37 #include <sys/arb.h>
  38 #include <sys/module.h>
  39 #include <sys/kernel.h>
  40 #ifdef TCP_HHOOK
  41 #include <sys/hhook.h>
  42 #endif
  43 #include <sys/lock.h>
  44 #include <sys/malloc.h>
  45 #include <sys/lock.h>
  46 #include <sys/mutex.h>
  47 #include <sys/mbuf.h>
  48 #include <sys/proc.h>           /* for proc0 declaration */
  49 #include <sys/socket.h>
  50 #include <sys/socketvar.h>
  51 #include <sys/sysctl.h>
  52 #include <sys/systm.h>
  53 #ifdef STATS
  54 #include <sys/qmath.h>
  55 #include <sys/tree.h>
  56 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
  57 #else
  58 #include <sys/tree.h>
  59 #endif
  60 #include <sys/refcount.h>
  61 #include <sys/queue.h>
  62 #include <sys/tim_filter.h>
  63 #include <sys/smp.h>
  64 #include <sys/kthread.h>
  65 #include <sys/kern_prefetch.h>
  66 #include <sys/protosw.h>
  67 #ifdef TCP_ACCOUNTING
  68 #include <sys/sched.h>
  69 #include <machine/cpu.h>
  70 #endif
  71 #include <vm/uma.h>
  72
  73 #include <net/route.h>
  74 #include <net/route/nhop.h>
  75 #include <net/vnet.h>
  76
  77 #define TCPSTATES               /* for logging */
  78
  79 #include <netinet/in.h>
  80 #include <netinet/in_kdtrace.h>
  81 #include <netinet/in_pcb.h>
  82 #include <netinet/ip.h>
  83 #include <netinet/ip_icmp.h>    /* required for icmp_var.h */
  84 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
  85 #include <netinet/ip_var.h>
  86 #include <netinet/ip6.h>
  87 #include <netinet6/in6_pcb.h>
  88 #include <netinet6/ip6_var.h>
  89 #include <netinet/tcp.h>
  90 #define TCPOUTFLAGS
  91 #include <netinet/tcp_fsm.h>
  92 #include <netinet/tcp_log_buf.h>
  93 #include <netinet/tcp_seq.h>
  94 #include <netinet/tcp_timer.h>
  95 #include <netinet/tcp_var.h>
  96 #include <netinet/tcp_syncache.h>
  97 #include <netinet/tcp_hpts.h>
  98 #include <netinet/tcp_ratelimit.h>
  99 #include <netinet/tcp_accounting.h>
 100 #include <netinet/tcpip.h>
 101 #include <netinet/cc/cc.h>
 102 #include <netinet/cc/cc_newreno.h>
 103 #include <netinet/tcp_fastopen.h>
 104 #include <netinet/tcp_lro.h>
 105 #ifdef NETFLIX_SHARED_CWND
 106 #include <netinet/tcp_shared_cwnd.h>
 107 #endif
 108 #ifdef TCP_OFFLOAD
 109 #include <netinet/tcp_offload.h>
 110 #endif
 111 #ifdef INET6
 112 #include <netinet6/tcp6_var.h>
 113 #endif
 114 #include <netinet/tcp_ecn.h>
 115
 116 #include <netipsec/ipsec_support.h>
 117
 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 119 #include <netipsec/ipsec.h>
 120 #include <netipsec/ipsec6.h>
 121 #endif                          /* IPSEC */
 122
 123 #include <netinet/udp.h>
 124 #include <netinet/udp_var.h>
 125 #include <machine/in_cksum.h>
 126
 127 #ifdef MAC
 128 #include <security/mac/mac_framework.h>
 129 #endif
 130 #include "sack_filter.h"
 131 #include "tcp_rack.h"
 132 #include "rack_bbr_common.h"
 133
 134 uma_zone_t rack_zone;
 135 uma_zone_t rack_pcb_zone;
 136
 137 #ifndef TICKS2SBT
 138 #define TICKS2SBT(__t)  (tick_sbt * ((sbintime_t)(__t)))
 139 #endif
 140
 141 VNET_DECLARE(uint32_t, newreno_beta);
 142 VNET_DECLARE(uint32_t, newreno_beta_ecn);
 143 #define V_newreno_beta VNET(newreno_beta)
 144 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
 145
 146
 147 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
 148 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
 149
 150 struct sysctl_ctx_list rack_sysctl_ctx;
 151 struct sysctl_oid *rack_sysctl_root;
 152
 153 #define CUM_ACKED 1
 154 #define SACKED 2
 155
 156 /*
 157  * The RACK module incorporates a number of
 158  * TCP ideas that have been put out into the IETF
 159  * over the last few years:
 160  * - Matt Mathis's Rate Halving which slowly drops
 161  *    the congestion window so that the ack clock can
 162  *    be maintained during a recovery.
 163  * - Yuchung Cheng's RACK TCP (for which its named) that
 164  *    will stop us using the number of dup acks and instead
 165  *    use time as the gage of when we retransmit.
 166  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 167  *    of Dukkipati et.al.
 168  * RACK depends on SACK, so if an endpoint arrives that
 169  * cannot do SACK the state machine below will shuttle the
 170  * connection back to using the "default" TCP stack that is
 171  * in FreeBSD.
 172  *
 173  * To implement RACK the original TCP stack was first decomposed
 174  * into a functional state machine with individual states
 175  * for each of the possible TCP connection states. The do_segment
 176  * functions role in life is to mandate the connection supports SACK
 177  * initially and then assure that the RACK state matches the conenction
 178  * state before calling the states do_segment function. Each
 179  * state is simplified due to the fact that the original do_segment
 180  * has been decomposed and we *know* what state we are in (no
 181  * switches on the state) and all tests for SACK are gone. This
 182  * greatly simplifies what each state does.
 183  *
 184  * TCP output is also over-written with a new version since it
 185  * must maintain the new rack scoreboard.
 186  *
 187  */
 188 static int32_t rack_tlp_thresh = 1;
 189 static int32_t rack_tlp_limit = 2;      /* No more than 2 TLPs w-out new data */
 190 static int32_t rack_tlp_use_greater = 1;
 191 static int32_t rack_reorder_thresh = 2;
 192 static int32_t rack_reorder_fade = 60000000;    /* 0 - never fade, def 60,000,000
 193                                                  * - 60 seconds */
 194 static uint8_t rack_req_measurements = 1;
 195 /* Attack threshold detections */
 196 static uint32_t rack_highest_sack_thresh_seen = 0;
 197 static uint32_t rack_highest_move_thresh_seen = 0;
 198 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
 199 static int32_t rack_hw_pace_extra_slots = 2;    /* 2 extra MSS time betweens */
 200 static int32_t rack_hw_rate_caps = 1; /* 1; */
 201 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
 202 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
 203 static int32_t rack_hw_up_only = 1;
 204 static int32_t rack_stats_gets_ms_rtt = 1;
 205 static int32_t rack_prr_addbackmax = 2;
 206 static int32_t rack_do_hystart = 0;
 207 static int32_t rack_apply_rtt_with_reduced_conf = 0;
 208
 209 static int32_t rack_pkt_delay = 1000;
 210 static int32_t rack_send_a_lot_in_prr = 1;
 211 static int32_t rack_min_to = 1000;      /* Number of microsecond  min timeout */
 212 static int32_t rack_verbose_logging = 0;
 213 static int32_t rack_ignore_data_after_close = 1;
 214 static int32_t rack_enable_shared_cwnd = 1;
 215 static int32_t rack_use_cmp_acks = 1;
 216 static int32_t rack_use_fsb = 1;
 217 static int32_t rack_use_rfo = 1;
 218 static int32_t rack_use_rsm_rfo = 1;
 219 static int32_t rack_max_abc_post_recovery = 2;
 220 static int32_t rack_client_low_buf = 0;
 221 static int32_t rack_dsack_std_based = 0x3;      /* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */
 222 #ifdef TCP_ACCOUNTING
 223 static int32_t rack_tcp_accounting = 0;
 224 #endif
 225 static int32_t rack_limits_scwnd = 1;
 226 static int32_t rack_enable_mqueue_for_nonpaced = 0;
 227 static int32_t rack_disable_prr = 0;
 228 static int32_t use_rack_rr = 1;
 229 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
 230 static int32_t rack_persist_min = 250000;       /* 250usec */
 231 static int32_t rack_persist_max = 2000000;      /* 2 Second in usec's */
 232 static int32_t rack_sack_not_required = 1;      /* set to one to allow non-sack to use rack */
 233 static int32_t rack_default_init_window = 0;    /* Use system default */
 234 static int32_t rack_limit_time_with_srtt = 0;
 235 static int32_t rack_autosndbuf_inc = 20;        /* In percentage form */
 236 static int32_t rack_enobuf_hw_boost_mult = 2;   /* How many times the hw rate we boost slot using time_between */
 237 static int32_t rack_enobuf_hw_max = 12000;      /* 12 ms in usecs */
 238 static int32_t rack_enobuf_hw_min = 10000;      /* 10 ms in usecs */
 239 static int32_t rack_hw_rwnd_factor = 2;         /* How many max_segs the rwnd must be before we hold off sending */
 240
 241 /*
 242  * Currently regular tcp has a rto_min of 30ms
 243  * the backoff goes 12 times so that ends up
 244  * being a total of 122.850 seconds before a
 245  * connection is killed.
 246  */
 247 static uint32_t rack_def_data_window = 20;
 248 static uint32_t rack_goal_bdp = 2;
 249 static uint32_t rack_min_srtts = 1;
 250 static uint32_t rack_min_measure_usec = 0;
 251 static int32_t rack_tlp_min = 10000;    /* 10ms */
 252 static int32_t rack_rto_min = 30000;    /* 30,000 usec same as main freebsd */
 253 static int32_t rack_rto_max = 4000000;  /* 4 seconds in usec's */
 254 static const int32_t rack_free_cache = 2;
 255 static int32_t rack_hptsi_segments = 40;
 256 static int32_t rack_rate_sample_method = USE_RTT_LOW;
 257 static int32_t rack_pace_every_seg = 0;
 258 static int32_t rack_delayed_ack_time = 40000;   /* 40ms in usecs */
 259 static int32_t rack_slot_reduction = 4;
 260 static int32_t rack_wma_divisor = 8;            /* For WMA calculation */
 261 static int32_t rack_cwnd_block_ends_measure = 0;
 262 static int32_t rack_rwnd_block_ends_measure = 0;
 263 static int32_t rack_def_profile = 0;
 264
 265 static int32_t rack_lower_cwnd_at_tlp = 0;
 266 static int32_t rack_limited_retran = 0;
 267 static int32_t rack_always_send_oldest = 0;
 268 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
 269
 270 static uint16_t rack_per_of_gp_ss = 250;        /* 250 % slow-start */
 271 static uint16_t rack_per_of_gp_ca = 200;        /* 200 % congestion-avoidance */
 272 static uint16_t rack_per_of_gp_rec = 200;       /* 200 % of bw */
 273
 274 /* Probertt */
 275 static uint16_t rack_per_of_gp_probertt = 60;   /* 60% of bw */
 276 static uint16_t rack_per_of_gp_lowthresh = 40;  /* 40% is bottom */
 277 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
 278 static uint16_t rack_atexit_prtt_hbp = 130;     /* Clamp to 130% on exit prtt if highly buffered path */
 279 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
 280
 281 static uint32_t rack_max_drain_wait = 2;        /* How man gp srtt's before we give up draining */
 282 static uint32_t rack_must_drain = 1;            /* How many GP srtt's we *must* wait */
 283 static uint32_t rack_probertt_use_min_rtt_entry = 1;    /* Use the min to calculate the goal else gp_srtt */
 284 static uint32_t rack_probertt_use_min_rtt_exit = 0;
 285 static uint32_t rack_probe_rtt_sets_cwnd = 0;
 286 static uint32_t rack_probe_rtt_safety_val = 2000000;    /* No more than 2 sec in probe-rtt */
 287 static uint32_t rack_time_between_probertt = 9600000;   /* 9.6 sec in usecs */
 288 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;       /* How many srtt periods does probe-rtt last top fraction */
 289 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;       /* How many srtt periods does probe-rtt last bottom fraction */
 290 static uint32_t rack_min_probertt_hold = 40000;         /* Equal to delayed ack time */
 291 static uint32_t rack_probertt_filter_life = 10000000;
 292 static uint32_t rack_probertt_lower_within = 10;
 293 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds)  to count as a lowering */
 294 static int32_t rack_pace_one_seg = 0;           /* Shall we pace for less than 1.4Meg 1MSS at a time */
 295 static int32_t rack_probertt_clear_is = 1;
 296 static int32_t rack_max_drain_hbp = 1;          /* Extra drain times gpsrtt for highly buffered paths */
 297 static int32_t rack_hbp_thresh = 3;             /* what is the divisor max_rtt/min_rtt to decided a hbp */
 298
 299 /* Part of pacing */
 300 static int32_t rack_max_per_above = 30;         /* When we go to increment stop if above 100+this% */
 301
 302 /* Timely information */
 303 /* Combine these two gives the range of 'no change' to bw */
 304 /* ie the up/down provide the upper and lower bound */
 305 static int32_t rack_gp_per_bw_mul_up = 2;       /* 2% */
 306 static int32_t rack_gp_per_bw_mul_down = 4;     /* 4% */
 307 static int32_t rack_gp_rtt_maxmul = 3;          /* 3 x maxmin */
 308 static int32_t rack_gp_rtt_minmul = 1;          /* minrtt + (minrtt/mindiv) is lower rtt */
 309 static int32_t rack_gp_rtt_mindiv = 4;          /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
 310 static int32_t rack_gp_decrease_per = 20;       /* 20% decrease in multiplier */
 311 static int32_t rack_gp_increase_per = 2;        /* 2% increase in multiplier */
 312 static int32_t rack_per_lower_bound = 50;       /* Don't allow to drop below this multiplier */
 313 static int32_t rack_per_upper_bound_ss = 0;     /* Don't allow SS to grow above this */
 314 static int32_t rack_per_upper_bound_ca = 0;     /* Don't allow CA to grow above this */
 315 static int32_t rack_do_dyn_mul = 0;             /* Are the rack gp multipliers dynamic */
 316 static int32_t rack_gp_no_rec_chg = 1;          /* Prohibit recovery from reducing it's multiplier */
 317 static int32_t rack_timely_dec_clear = 6;       /* Do we clear decrement count at a value (6)? */
 318 static int32_t rack_timely_max_push_rise = 3;   /* One round of pushing */
 319 static int32_t rack_timely_max_push_drop = 3;   /* Three round of pushing */
 320 static int32_t rack_timely_min_segs = 4;        /* 4 segment minimum */
 321 static int32_t rack_use_max_for_nobackoff = 0;
 322 static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */
 323 static int32_t rack_timely_no_stopping = 0;
 324 static int32_t rack_down_raise_thresh = 100;
 325 static int32_t rack_req_segs = 1;
 326 static uint64_t rack_bw_rate_cap = 0;
 327 static uint32_t rack_trace_point_config = 0;
 328 static uint32_t rack_trace_point_bb_mode = 4;
 329 static int32_t rack_trace_point_count = 0;
 330
 331
 332 /* Weird delayed ack mode */
 333 static int32_t rack_use_imac_dack = 0;
 334 /* Rack specific counters */
 335 counter_u64_t rack_saw_enobuf;
 336 counter_u64_t rack_saw_enobuf_hw;
 337 counter_u64_t rack_saw_enetunreach;
 338 counter_u64_t rack_persists_sends;
 339 counter_u64_t rack_persists_acks;
 340 counter_u64_t rack_persists_loss;
 341 counter_u64_t rack_persists_lost_ends;
 342 #ifdef INVARIANTS
 343 counter_u64_t rack_adjust_map_bw;
 344 #endif
 345 /* Tail loss probe counters */
 346 counter_u64_t rack_tlp_tot;
 347 counter_u64_t rack_tlp_newdata;
 348 counter_u64_t rack_tlp_retran;
 349 counter_u64_t rack_tlp_retran_bytes;
 350 counter_u64_t rack_to_tot;
 351 counter_u64_t rack_hot_alloc;
 352 counter_u64_t rack_to_alloc;
 353 counter_u64_t rack_to_alloc_hard;
 354 counter_u64_t rack_to_alloc_emerg;
 355 counter_u64_t rack_to_alloc_limited;
 356 counter_u64_t rack_alloc_limited_conns;
 357 counter_u64_t rack_split_limited;
 358
 359 counter_u64_t rack_multi_single_eq;
 360 counter_u64_t rack_proc_non_comp_ack;
 361
 362 counter_u64_t rack_fto_send;
 363 counter_u64_t rack_fto_rsm_send;
 364 counter_u64_t rack_nfto_resend;
 365 counter_u64_t rack_non_fto_send;
 366 counter_u64_t rack_extended_rfo;
 367
 368 counter_u64_t rack_sack_proc_all;
 369 counter_u64_t rack_sack_proc_short;
 370 counter_u64_t rack_sack_proc_restart;
 371 counter_u64_t rack_sack_attacks_detected;
 372 counter_u64_t rack_sack_attacks_reversed;
 373 counter_u64_t rack_sack_used_next_merge;
 374 counter_u64_t rack_sack_splits;
 375 counter_u64_t rack_sack_used_prev_merge;
 376 counter_u64_t rack_sack_skipped_acked;
 377 counter_u64_t rack_ack_total;
 378 counter_u64_t rack_express_sack;
 379 counter_u64_t rack_sack_total;
 380 counter_u64_t rack_move_none;
 381 counter_u64_t rack_move_some;
 382
 383 counter_u64_t rack_input_idle_reduces;
 384 counter_u64_t rack_collapsed_win;
 385 counter_u64_t rack_collapsed_win_seen;
 386 counter_u64_t rack_collapsed_win_rxt;
 387 counter_u64_t rack_collapsed_win_rxt_bytes;
 388 counter_u64_t rack_try_scwnd;
 389 counter_u64_t rack_hw_pace_init_fail;
 390 counter_u64_t rack_hw_pace_lost;
 391
 392 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
 393 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
 394
 395
 396 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
 397
 398 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {  \
 399         (tv) = (value) + slop;   \
 400         if ((u_long)(tv) < (u_long)(tvmin)) \
 401                 (tv) = (tvmin); \
 402         if ((u_long)(tv) > (u_long)(tvmax)) \
 403                 (tv) = (tvmax); \
 404 } while (0)
 405
 406 static void
 407 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
 408
 409 static int
 410 rack_process_ack(struct mbuf *m, struct tcphdr *th,
 411     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
 412     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
 413 static int
 414 rack_process_data(struct mbuf *m, struct tcphdr *th,
 415     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 416     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
 417 static void
 418 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
 419    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
 420 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
 421 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
 422     uint8_t limit_type);
 423 static struct rack_sendmap *
 424 rack_check_recovery_mode(struct tcpcb *tp,
 425     uint32_t tsused);
 426 static void
 427 rack_cong_signal(struct tcpcb *tp,
 428                  uint32_t type, uint32_t ack, int );
 429 static void rack_counter_destroy(void);
 430 static int
 431 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt);
 432 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
 433 static void
 434 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
 435 static void
 436 rack_do_segment(struct mbuf *m, struct tcphdr *th,
 437     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 438     uint8_t iptos);
 439 static void rack_dtor(void *mem, int32_t size, void *arg);
 440 static void
 441 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
 442     uint32_t flex1, uint32_t flex2,
 443     uint32_t flex3, uint32_t flex4,
 444     uint32_t flex5, uint32_t flex6,
 445     uint16_t flex7, uint8_t mod);
 446
 447 static void
 448 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
 449    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
 450    struct rack_sendmap *rsm, uint8_t quality);
 451 static struct rack_sendmap *
 452 rack_find_high_nonack(struct tcp_rack *rack,
 453     struct rack_sendmap *rsm);
 454 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
 455 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
 456 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
 457 static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt);
 458 static void
 459 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
 460                             tcp_seq th_ack, int line, uint8_t quality);
 461 static uint32_t
 462 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
 463 static int32_t rack_handoff_ok(struct tcpcb *tp);
 464 static int32_t rack_init(struct tcpcb *tp);
 465 static void rack_init_sysctls(void);
 466 static void
 467 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
 468     struct tcphdr *th, int entered_rec, int dup_ack_struck);
 469 static void
 470 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
 471     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
 472     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
 473
 474 static void
 475 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
 476     struct rack_sendmap *rsm);
 477 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
 478 static int32_t rack_output(struct tcpcb *tp);
 479
 480 static uint32_t
 481 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
 482     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
 483     uint32_t cts, int *moved_two);
 484 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
 485 static void rack_remxt_tmr(struct tcpcb *tp);
 486 static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt);
 487 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
 488 static int32_t rack_stopall(struct tcpcb *tp);
 489 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
 490 static uint32_t
 491 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
 492     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
 493 static void
 494 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
 495     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
 496 static int
 497 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
 498     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
 499 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
 500 static int
 501 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
 502     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 503     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 504 static int
 505 rack_do_closing(struct mbuf *m, struct tcphdr *th,
 506     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 507     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 508 static int
 509 rack_do_established(struct mbuf *m, struct tcphdr *th,
 510     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 511     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 512 static int
 513 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
 514     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 515     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
 516 static int
 517 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
 518     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 519     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 520 static int
 521 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
 522     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 523     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 524 static int
 525 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
 526     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 527     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 528 static int
 529 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
 530     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 531     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 532 static int
 533 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
 534     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 535     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 536 struct rack_sendmap *
 537 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
 538     uint32_t tsused);
 539 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
 540     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
 541 static void
 542      tcp_rack_partialack(struct tcpcb *tp);
 543 static int
 544 rack_set_profile(struct tcp_rack *rack, int prof);
 545 static void
 546 rack_apply_deferred_options(struct tcp_rack *rack);
 547
 548 int32_t rack_clear_counter=0;
 549
 550 static inline void
 551 rack_trace_point(struct tcp_rack *rack, int num)
 552 {
 553         if (((rack_trace_point_config == num)  ||
 554              (rack_trace_point_config = 0xffffffff)) &&
 555             (rack_trace_point_bb_mode != 0) &&
 556             (rack_trace_point_count > 0) &&
 557             (rack->rc_tp->t_logstate == 0)) {
 558                 int res;
 559                 res = atomic_fetchadd_int(&rack_trace_point_count, -1);
 560                 if (res > 0) {
 561                         rack->rc_tp->t_logstate = rack_trace_point_bb_mode;
 562                 } else {
 563                         /* Loss a race assure its zero now */
 564                         rack_trace_point_count = 0;
 565                 }
 566         }
 567 }
 568
 569 static void
 570 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
 571 {
 572         struct sockopt sopt;
 573         struct cc_newreno_opts opt;
 574         struct newreno old;
 575         struct tcpcb *tp;
 576         int error, failed = 0;
 577
 578         tp = rack->rc_tp;
 579         if (tp->t_cc == NULL) {
 580                 /* Tcb is leaving */
 581                 return;
 582         }
 583         rack->rc_pacing_cc_set = 1;
 584         if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
 585                 /* Not new-reno we can't play games with beta! */
 586                 failed = 1;
 587                 goto out;
 588
 589         }
 590         if (CC_ALGO(tp)->ctl_output == NULL)  {
 591                 /* Huh, not using new-reno so no swaps.? */
 592                 failed = 2;
 593                 goto out;
 594         }
 595         /* Get the current values out */
 596         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 597         sopt.sopt_dir = SOPT_GET;
 598         opt.name = CC_NEWRENO_BETA;
 599         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 600         if (error)  {
 601                 failed = 3;
 602                 goto out;
 603         }
 604         old.beta = opt.val;
 605         opt.name = CC_NEWRENO_BETA_ECN;
 606         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 607         if (error)  {
 608                 failed = 4;
 609                 goto out;
 610         }
 611         old.beta_ecn = opt.val;
 612
 613         /* Now lets set in the values we have stored */
 614         sopt.sopt_dir = SOPT_SET;
 615         opt.name = CC_NEWRENO_BETA;
 616         opt.val = rack->r_ctl.rc_saved_beta.beta;
 617         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 618         if (error)  {
 619                 failed = 5;
 620                 goto out;
 621         }
 622         opt.name = CC_NEWRENO_BETA_ECN;
 623         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 624         error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
 625         if (error) {
 626                 failed = 6;
 627                 goto out;
 628         }
 629         /* Save off the values for restoral */
 630         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 631 out:
 632         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 633                 union tcp_log_stackspecific log;
 634                 struct timeval tv;
 635                 struct newreno *ptr;
 636
 637                 ptr = ((struct newreno *)tp->t_ccv.cc_data);
 638                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 639                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 640                 log.u_bbr.flex1 = ptr->beta;
 641                 log.u_bbr.flex2 = ptr->beta_ecn;
 642                 log.u_bbr.flex3 = ptr->newreno_flags;
 643                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 644                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 645                 log.u_bbr.flex6 = failed;
 646                 log.u_bbr.flex7 = rack->gp_ready;
 647                 log.u_bbr.flex7 <<= 1;
 648                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 649                 log.u_bbr.flex7 <<= 1;
 650                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 651                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 652                 log.u_bbr.flex8 = flex8;
 653                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 654                                0, &log, false, NULL, NULL, 0, &tv);
 655         }
 656 }
 657
 658 static void
 659 rack_set_cc_pacing(struct tcp_rack *rack)
 660 {
 661         if (rack->rc_pacing_cc_set)
 662                 return;
 663         /*
 664          * Use the swap utility placing in 3 for flex8 to id a
 665          * set of a new set of values.
 666          */
 667         rack->rc_pacing_cc_set = 1;
 668         rack_swap_beta_values(rack, 3);
 669 }
 670
 671 static void
 672 rack_undo_cc_pacing(struct tcp_rack *rack)
 673 {
 674         if (rack->rc_pacing_cc_set == 0)
 675                 return;
 676         /*
 677          * Use the swap utility placing in 4 for flex8 to id a
 678          * restoral of the old values.
 679          */
 680         rack->rc_pacing_cc_set = 0;
 681         rack_swap_beta_values(rack, 4);
 682 }
 683
 684 #ifdef NETFLIX_PEAKRATE
 685 static inline void
 686 rack_update_peakrate_thr(struct tcpcb *tp)
 687 {
 688         /* Keep in mind that t_maxpeakrate is in B/s. */
 689         uint64_t peak;
 690         peak = uqmax((tp->t_maxseg * 2),
 691                      (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
 692         tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
 693 }
 694 #endif
 695
 696 static int
 697 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 698 {
 699         uint32_t stat;
 700         int32_t error;
 701
 702         error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
 703         if (error || req->newptr == NULL)
 704                 return error;
 705
 706         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
 707         if (error)
 708                 return (error);
 709         if (stat == 1) {
 710 #ifdef INVARIANTS
 711                 printf("Clearing RACK counters\n");
 712 #endif
 713                 counter_u64_zero(rack_tlp_tot);
 714                 counter_u64_zero(rack_tlp_newdata);
 715                 counter_u64_zero(rack_tlp_retran);
 716                 counter_u64_zero(rack_tlp_retran_bytes);
 717                 counter_u64_zero(rack_to_tot);
 718                 counter_u64_zero(rack_saw_enobuf);
 719                 counter_u64_zero(rack_saw_enobuf_hw);
 720                 counter_u64_zero(rack_saw_enetunreach);
 721                 counter_u64_zero(rack_persists_sends);
 722                 counter_u64_zero(rack_persists_acks);
 723                 counter_u64_zero(rack_persists_loss);
 724                 counter_u64_zero(rack_persists_lost_ends);
 725 #ifdef INVARIANTS
 726                 counter_u64_zero(rack_adjust_map_bw);
 727 #endif
 728                 counter_u64_zero(rack_to_alloc_hard);
 729                 counter_u64_zero(rack_to_alloc_emerg);
 730                 counter_u64_zero(rack_sack_proc_all);
 731                 counter_u64_zero(rack_fto_send);
 732                 counter_u64_zero(rack_fto_rsm_send);
 733                 counter_u64_zero(rack_extended_rfo);
 734                 counter_u64_zero(rack_hw_pace_init_fail);
 735                 counter_u64_zero(rack_hw_pace_lost);
 736                 counter_u64_zero(rack_non_fto_send);
 737                 counter_u64_zero(rack_nfto_resend);
 738                 counter_u64_zero(rack_sack_proc_short);
 739                 counter_u64_zero(rack_sack_proc_restart);
 740                 counter_u64_zero(rack_to_alloc);
 741                 counter_u64_zero(rack_to_alloc_limited);
 742                 counter_u64_zero(rack_alloc_limited_conns);
 743                 counter_u64_zero(rack_split_limited);
 744                 counter_u64_zero(rack_multi_single_eq);
 745                 counter_u64_zero(rack_proc_non_comp_ack);
 746                 counter_u64_zero(rack_sack_attacks_detected);
 747                 counter_u64_zero(rack_sack_attacks_reversed);
 748                 counter_u64_zero(rack_sack_used_next_merge);
 749                 counter_u64_zero(rack_sack_used_prev_merge);
 750                 counter_u64_zero(rack_sack_splits);
 751                 counter_u64_zero(rack_sack_skipped_acked);
 752                 counter_u64_zero(rack_ack_total);
 753                 counter_u64_zero(rack_express_sack);
 754                 counter_u64_zero(rack_sack_total);
 755                 counter_u64_zero(rack_move_none);
 756                 counter_u64_zero(rack_move_some);
 757                 counter_u64_zero(rack_try_scwnd);
 758                 counter_u64_zero(rack_collapsed_win);
 759                 counter_u64_zero(rack_collapsed_win_rxt);
 760                 counter_u64_zero(rack_collapsed_win_seen);
 761                 counter_u64_zero(rack_collapsed_win_rxt_bytes);
 762         }
 763         rack_clear_counter = 0;
 764         return (0);
 765 }
 766
 767 static void
 768 rack_init_sysctls(void)
 769 {
 770         struct sysctl_oid *rack_counters;
 771         struct sysctl_oid *rack_attack;
 772         struct sysctl_oid *rack_pacing;
 773         struct sysctl_oid *rack_timely;
 774         struct sysctl_oid *rack_timers;
 775         struct sysctl_oid *rack_tlp;
 776         struct sysctl_oid *rack_misc;
 777         struct sysctl_oid *rack_features;
 778         struct sysctl_oid *rack_measure;
 779         struct sysctl_oid *rack_probertt;
 780         struct sysctl_oid *rack_hw_pacing;
 781         struct sysctl_oid *rack_tracepoint;
 782
 783         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 784             SYSCTL_CHILDREN(rack_sysctl_root),
 785             OID_AUTO,
 786             "sack_attack",
 787             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 788             "Rack Sack Attack Counters and Controls");
 789         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 790             SYSCTL_CHILDREN(rack_sysctl_root),
 791             OID_AUTO,
 792             "stats",
 793             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 794             "Rack Counters");
 795         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 796             SYSCTL_CHILDREN(rack_sysctl_root),
 797             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 798             &rack_rate_sample_method , USE_RTT_LOW,
 799             "What method should we use for rate sampling 0=high, 1=low ");
 800         /* Probe rtt related controls */
 801         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 802             SYSCTL_CHILDREN(rack_sysctl_root),
 803             OID_AUTO,
 804             "probertt",
 805             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 806             "ProbeRTT related Controls");
 807         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 808             SYSCTL_CHILDREN(rack_probertt),
 809             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 810             &rack_atexit_prtt_hbp, 130,
 811             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 812         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 813             SYSCTL_CHILDREN(rack_probertt),
 814             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 815             &rack_atexit_prtt, 130,
 816             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 817         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 818             SYSCTL_CHILDREN(rack_probertt),
 819             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 820             &rack_per_of_gp_probertt, 60,
 821             "What percentage of goodput do we pace at in probertt");
 822         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 823             SYSCTL_CHILDREN(rack_probertt),
 824             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 825             &rack_per_of_gp_probertt_reduce, 10,
 826             "What percentage of goodput do we reduce every gp_srtt");
 827         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 828             SYSCTL_CHILDREN(rack_probertt),
 829             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 830             &rack_per_of_gp_lowthresh, 40,
 831             "What percentage of goodput do we allow the multiplier to fall to");
 832         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 833             SYSCTL_CHILDREN(rack_probertt),
 834             OID_AUTO, "time_between", CTLFLAG_RW,
 835             & rack_time_between_probertt, 96000000,
 836             "How many useconds between the lowest rtt falling must past before we enter probertt");
 837         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 838             SYSCTL_CHILDREN(rack_probertt),
 839             OID_AUTO, "safety", CTLFLAG_RW,
 840             &rack_probe_rtt_safety_val, 2000000,
 841             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 842         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 843             SYSCTL_CHILDREN(rack_probertt),
 844             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 845             &rack_probe_rtt_sets_cwnd, 0,
 846             "Do we set the cwnd too (if always_lower is on)");
 847         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 848             SYSCTL_CHILDREN(rack_probertt),
 849             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 850             &rack_max_drain_wait, 2,
 851             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 852         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 853             SYSCTL_CHILDREN(rack_probertt),
 854             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 855             &rack_must_drain, 1,
 856             "We must drain this many gp_srtt's waiting for flight to reach goal");
 857         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 858             SYSCTL_CHILDREN(rack_probertt),
 859             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 860             &rack_probertt_use_min_rtt_entry, 1,
 861             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 862         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 863             SYSCTL_CHILDREN(rack_probertt),
 864             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 865             &rack_probertt_use_min_rtt_exit, 0,
 866             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 867         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 868             SYSCTL_CHILDREN(rack_probertt),
 869             OID_AUTO, "length_div", CTLFLAG_RW,
 870             &rack_probertt_gpsrtt_cnt_div, 0,
 871             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 872         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 873             SYSCTL_CHILDREN(rack_probertt),
 874             OID_AUTO, "length_mul", CTLFLAG_RW,
 875             &rack_probertt_gpsrtt_cnt_mul, 0,
 876             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 877         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 878             SYSCTL_CHILDREN(rack_probertt),
 879             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 880             &rack_min_probertt_hold, 200000,
 881             "What is the minimum time we hold probertt at target");
 882         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 883             SYSCTL_CHILDREN(rack_probertt),
 884             OID_AUTO, "filter_life", CTLFLAG_RW,
 885             &rack_probertt_filter_life, 10000000,
 886             "What is the time for the filters life in useconds");
 887         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 888             SYSCTL_CHILDREN(rack_probertt),
 889             OID_AUTO, "lower_within", CTLFLAG_RW,
 890             &rack_probertt_lower_within, 10,
 891             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 892         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 893             SYSCTL_CHILDREN(rack_probertt),
 894             OID_AUTO, "must_move", CTLFLAG_RW,
 895             &rack_min_rtt_movement, 250,
 896             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 897         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 898             SYSCTL_CHILDREN(rack_probertt),
 899             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 900             &rack_probertt_clear_is, 1,
 901             "Do we clear I/S counts on exiting probe-rtt");
 902         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 903             SYSCTL_CHILDREN(rack_probertt),
 904             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 905             &rack_max_drain_hbp, 1,
 906             "How many extra drain gpsrtt's do we get in highly buffered paths");
 907         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 908             SYSCTL_CHILDREN(rack_probertt),
 909             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 910             &rack_hbp_thresh, 3,
 911             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 912
 913         rack_tracepoint = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 914             SYSCTL_CHILDREN(rack_sysctl_root),
 915             OID_AUTO,
 916             "tp",
 917             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 918             "Rack tracepoint facility");
 919         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 920             SYSCTL_CHILDREN(rack_tracepoint),
 921             OID_AUTO, "number", CTLFLAG_RW,
 922             &rack_trace_point_config, 0,
 923             "What is the trace point number to activate (0=none, 0xffffffff = all)?");
 924         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 925             SYSCTL_CHILDREN(rack_tracepoint),
 926             OID_AUTO, "bbmode", CTLFLAG_RW,
 927             &rack_trace_point_bb_mode, 4,
 928             "What is BB logging mode that is activated?");
 929         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 930             SYSCTL_CHILDREN(rack_tracepoint),
 931             OID_AUTO, "count", CTLFLAG_RW,
 932             &rack_trace_point_count, 0,
 933             "How many connections will have BB logging turned on that hit the tracepoint?");
 934         /* Pacing related sysctls */
 935         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 936             SYSCTL_CHILDREN(rack_sysctl_root),
 937             OID_AUTO,
 938             "pacing",
 939             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 940             "Pacing related Controls");
 941         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 942             SYSCTL_CHILDREN(rack_pacing),
 943             OID_AUTO, "max_pace_over", CTLFLAG_RW,
 944             &rack_max_per_above, 30,
 945             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
 946         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 947             SYSCTL_CHILDREN(rack_pacing),
 948             OID_AUTO, "pace_to_one", CTLFLAG_RW,
 949             &rack_pace_one_seg, 0,
 950             "Do we allow low b/w pacing of 1MSS instead of two");
 951         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 952             SYSCTL_CHILDREN(rack_pacing),
 953             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
 954             &rack_limit_time_with_srtt, 0,
 955             "Do we limit pacing time based on srtt");
 956         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 957             SYSCTL_CHILDREN(rack_pacing),
 958             OID_AUTO, "init_win", CTLFLAG_RW,
 959             &rack_default_init_window, 0,
 960             "Do we have a rack initial window 0 = system default");
 961         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 962             SYSCTL_CHILDREN(rack_pacing),
 963             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
 964             &rack_per_of_gp_ss, 250,
 965             "If non zero, what percentage of goodput to pace at in slow start");
 966         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 967             SYSCTL_CHILDREN(rack_pacing),
 968             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
 969             &rack_per_of_gp_ca, 150,
 970             "If non zero, what percentage of goodput to pace at in congestion avoidance");
 971         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 972             SYSCTL_CHILDREN(rack_pacing),
 973             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
 974             &rack_per_of_gp_rec, 200,
 975             "If non zero, what percentage of goodput to pace at in recovery");
 976         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 977             SYSCTL_CHILDREN(rack_pacing),
 978             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
 979             &rack_hptsi_segments, 40,
 980             "What size is the max for TSO segments in pacing and burst mitigation");
 981         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 982             SYSCTL_CHILDREN(rack_pacing),
 983             OID_AUTO, "burst_reduces", CTLFLAG_RW,
 984             &rack_slot_reduction, 4,
 985             "When doing only burst mitigation what is the reduce divisor");
 986         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 987             SYSCTL_CHILDREN(rack_sysctl_root),
 988             OID_AUTO, "use_pacing", CTLFLAG_RW,
 989             &rack_pace_every_seg, 0,
 990             "If set we use pacing, if clear we use only the original burst mitigation");
 991         SYSCTL_ADD_U64(&rack_sysctl_ctx,
 992             SYSCTL_CHILDREN(rack_pacing),
 993             OID_AUTO, "rate_cap", CTLFLAG_RW,
 994             &rack_bw_rate_cap, 0,
 995             "If set we apply this value to the absolute rate cap used by pacing");
 996         SYSCTL_ADD_U8(&rack_sysctl_ctx,
 997             SYSCTL_CHILDREN(rack_sysctl_root),
 998             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
 999             &rack_req_measurements, 1,
1000             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1001         /* Hardware pacing */
1002         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1003             SYSCTL_CHILDREN(rack_sysctl_root),
1004             OID_AUTO,
1005             "hdwr_pacing",
1006             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1007             "Pacing related Controls");
1008         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1009             SYSCTL_CHILDREN(rack_hw_pacing),
1010             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1011             &rack_hw_rwnd_factor, 2,
1012             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1013         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1014             SYSCTL_CHILDREN(rack_hw_pacing),
1015             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1016             &rack_enobuf_hw_boost_mult, 2,
1017             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1018         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1019             SYSCTL_CHILDREN(rack_hw_pacing),
1020             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1021             &rack_enobuf_hw_max, 2,
1022             "What is the max boost the pacing time if we see a ENOBUFS?");
1023         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1024             SYSCTL_CHILDREN(rack_hw_pacing),
1025             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1026             &rack_enobuf_hw_min, 2,
1027             "What is the min boost the pacing time if we see a ENOBUFS?");
1028         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1029             SYSCTL_CHILDREN(rack_hw_pacing),
1030             OID_AUTO, "enable", CTLFLAG_RW,
1031             &rack_enable_hw_pacing, 0,
1032             "Should RACK attempt to use hw pacing?");
1033         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1034             SYSCTL_CHILDREN(rack_hw_pacing),
1035             OID_AUTO, "rate_cap", CTLFLAG_RW,
1036             &rack_hw_rate_caps, 1,
1037             "Does the highest hardware pacing rate cap the rate we will send at??");
1038         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1039             SYSCTL_CHILDREN(rack_hw_pacing),
1040             OID_AUTO, "rate_min", CTLFLAG_RW,
1041             &rack_hw_rate_min, 0,
1042             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1043         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1044             SYSCTL_CHILDREN(rack_hw_pacing),
1045             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1046             &rack_hw_rate_to_low, 0,
1047             "If we fall below this rate, dis-engage hw pacing?");
1048         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1049             SYSCTL_CHILDREN(rack_hw_pacing),
1050             OID_AUTO, "up_only", CTLFLAG_RW,
1051             &rack_hw_up_only, 1,
1052             "Do we allow hw pacing to lower the rate selected?");
1053         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1054             SYSCTL_CHILDREN(rack_hw_pacing),
1055             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1056             &rack_hw_pace_extra_slots, 2,
1057             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1058         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1059             SYSCTL_CHILDREN(rack_sysctl_root),
1060             OID_AUTO,
1061             "timely",
1062             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1063             "Rack Timely RTT Controls");
1064         /* Timely based GP dynmics */
1065         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1066             SYSCTL_CHILDREN(rack_timely),
1067             OID_AUTO, "upper", CTLFLAG_RW,
1068             &rack_gp_per_bw_mul_up, 2,
1069             "Rack timely upper range for equal b/w (in percentage)");
1070         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1071             SYSCTL_CHILDREN(rack_timely),
1072             OID_AUTO, "lower", CTLFLAG_RW,
1073             &rack_gp_per_bw_mul_down, 4,
1074             "Rack timely lower range for equal b/w (in percentage)");
1075         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1076             SYSCTL_CHILDREN(rack_timely),
1077             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1078             &rack_gp_rtt_maxmul, 3,
1079             "Rack timely multiplier of lowest rtt for rtt_max");
1080         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1081             SYSCTL_CHILDREN(rack_timely),
1082             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1083             &rack_gp_rtt_mindiv, 4,
1084             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1085         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1086             SYSCTL_CHILDREN(rack_timely),
1087             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1088             &rack_gp_rtt_minmul, 1,
1089             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1090         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1091             SYSCTL_CHILDREN(rack_timely),
1092             OID_AUTO, "decrease", CTLFLAG_RW,
1093             &rack_gp_decrease_per, 20,
1094             "Rack timely decrease percentage of our GP multiplication factor");
1095         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1096             SYSCTL_CHILDREN(rack_timely),
1097             OID_AUTO, "increase", CTLFLAG_RW,
1098             &rack_gp_increase_per, 2,
1099             "Rack timely increase perentage of our GP multiplication factor");
1100         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1101             SYSCTL_CHILDREN(rack_timely),
1102             OID_AUTO, "lowerbound", CTLFLAG_RW,
1103             &rack_per_lower_bound, 50,
1104             "Rack timely lowest percentage we allow GP multiplier to fall to");
1105         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1106             SYSCTL_CHILDREN(rack_timely),
1107             OID_AUTO, "upperboundss", CTLFLAG_RW,
1108             &rack_per_upper_bound_ss, 0,
1109             "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1110         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1111             SYSCTL_CHILDREN(rack_timely),
1112             OID_AUTO, "upperboundca", CTLFLAG_RW,
1113             &rack_per_upper_bound_ca, 0,
1114             "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1115         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1116             SYSCTL_CHILDREN(rack_timely),
1117             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1118             &rack_do_dyn_mul, 0,
1119             "Rack timely do we enable dynmaic timely goodput by default");
1120         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1121             SYSCTL_CHILDREN(rack_timely),
1122             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1123             &rack_gp_no_rec_chg, 1,
1124             "Rack timely do we prohibit the recovery multiplier from being lowered");
1125         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1126             SYSCTL_CHILDREN(rack_timely),
1127             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1128             &rack_timely_dec_clear, 6,
1129             "Rack timely what threshold do we count to before another boost during b/w decent");
1130         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1131             SYSCTL_CHILDREN(rack_timely),
1132             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1133             &rack_timely_max_push_rise, 3,
1134             "Rack timely how many times do we push up with b/w increase");
1135         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1136             SYSCTL_CHILDREN(rack_timely),
1137             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1138             &rack_timely_max_push_drop, 3,
1139             "Rack timely how many times do we push back on b/w decent");
1140         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1141             SYSCTL_CHILDREN(rack_timely),
1142             OID_AUTO, "min_segs", CTLFLAG_RW,
1143             &rack_timely_min_segs, 4,
1144             "Rack timely when setting the cwnd what is the min num segments");
1145         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1146             SYSCTL_CHILDREN(rack_timely),
1147             OID_AUTO, "noback_max", CTLFLAG_RW,
1148             &rack_use_max_for_nobackoff, 0,
1149             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1150         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1151             SYSCTL_CHILDREN(rack_timely),
1152             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1153             &rack_timely_int_timely_only, 0,
1154             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1155         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1156             SYSCTL_CHILDREN(rack_timely),
1157             OID_AUTO, "nonstop", CTLFLAG_RW,
1158             &rack_timely_no_stopping, 0,
1159             "Rack timely don't stop increase");
1160         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1161             SYSCTL_CHILDREN(rack_timely),
1162             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1163             &rack_down_raise_thresh, 100,
1164             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1165         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1166             SYSCTL_CHILDREN(rack_timely),
1167             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1168             &rack_req_segs, 1,
1169             "Bottom dragging if not these many segments outstanding and room");
1170
1171         /* TLP and Rack related parameters */
1172         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1173             SYSCTL_CHILDREN(rack_sysctl_root),
1174             OID_AUTO,
1175             "tlp",
1176             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1177             "TLP and Rack related Controls");
1178         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1179             SYSCTL_CHILDREN(rack_tlp),
1180             OID_AUTO, "use_rrr", CTLFLAG_RW,
1181             &use_rack_rr, 1,
1182             "Do we use Rack Rapid Recovery");
1183         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1184             SYSCTL_CHILDREN(rack_tlp),
1185             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1186             &rack_max_abc_post_recovery, 2,
1187             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1188         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1189             SYSCTL_CHILDREN(rack_tlp),
1190             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1191             &rack_non_rxt_use_cr, 0,
1192             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1193         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1194             SYSCTL_CHILDREN(rack_tlp),
1195             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1196             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1197             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1198         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1199             SYSCTL_CHILDREN(rack_tlp),
1200             OID_AUTO, "limit", CTLFLAG_RW,
1201             &rack_tlp_limit, 2,
1202             "How many TLP's can be sent without sending new data");
1203         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1204             SYSCTL_CHILDREN(rack_tlp),
1205             OID_AUTO, "use_greater", CTLFLAG_RW,
1206             &rack_tlp_use_greater, 1,
1207             "Should we use the rack_rtt time if its greater than srtt");
1208         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1209             SYSCTL_CHILDREN(rack_tlp),
1210             OID_AUTO, "tlpminto", CTLFLAG_RW,
1211             &rack_tlp_min, 10000,
1212             "TLP minimum timeout per the specification (in microseconds)");
1213         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1214             SYSCTL_CHILDREN(rack_tlp),
1215             OID_AUTO, "send_oldest", CTLFLAG_RW,
1216             &rack_always_send_oldest, 0,
1217             "Should we always send the oldest TLP and RACK-TLP");
1218         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1219             SYSCTL_CHILDREN(rack_tlp),
1220             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1221             &rack_limited_retran, 0,
1222             "How many times can a rack timeout drive out sends");
1223         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1224             SYSCTL_CHILDREN(rack_tlp),
1225             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1226             &rack_lower_cwnd_at_tlp, 0,
1227             "When a TLP completes a retran should we enter recovery");
1228         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229             SYSCTL_CHILDREN(rack_tlp),
1230             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1231             &rack_reorder_thresh, 2,
1232             "What factor for rack will be added when seeing reordering (shift right)");
1233         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1234             SYSCTL_CHILDREN(rack_tlp),
1235             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1236             &rack_tlp_thresh, 1,
1237             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1238         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1239             SYSCTL_CHILDREN(rack_tlp),
1240             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1241             &rack_reorder_fade, 60000000,
1242             "Does reorder detection fade, if so how many microseconds (0 means never)");
1243         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1244             SYSCTL_CHILDREN(rack_tlp),
1245             OID_AUTO, "pktdelay", CTLFLAG_RW,
1246             &rack_pkt_delay, 1000,
1247             "Extra RACK time (in microseconds) besides reordering thresh");
1248
1249         /* Timer related controls */
1250         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1251             SYSCTL_CHILDREN(rack_sysctl_root),
1252             OID_AUTO,
1253             "timers",
1254             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1255             "Timer related controls");
1256         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1257             SYSCTL_CHILDREN(rack_timers),
1258             OID_AUTO, "persmin", CTLFLAG_RW,
1259             &rack_persist_min, 250000,
1260             "What is the minimum time in microseconds between persists");
1261         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1262             SYSCTL_CHILDREN(rack_timers),
1263             OID_AUTO, "persmax", CTLFLAG_RW,
1264             &rack_persist_max, 2000000,
1265             "What is the largest delay in microseconds between persists");
1266         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1267             SYSCTL_CHILDREN(rack_timers),
1268             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1269             &rack_delayed_ack_time, 40000,
1270             "Delayed ack time (40ms in microseconds)");
1271         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1272             SYSCTL_CHILDREN(rack_timers),
1273             OID_AUTO, "minrto", CTLFLAG_RW,
1274             &rack_rto_min, 30000,
1275             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1276         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1277             SYSCTL_CHILDREN(rack_timers),
1278             OID_AUTO, "maxrto", CTLFLAG_RW,
1279             &rack_rto_max, 4000000,
1280             "Maximum RTO in microseconds -- should be at least as large as min_rto");
1281         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1282             SYSCTL_CHILDREN(rack_timers),
1283             OID_AUTO, "minto", CTLFLAG_RW,
1284             &rack_min_to, 1000,
1285             "Minimum rack timeout in microseconds");
1286         /* Measure controls */
1287         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1288             SYSCTL_CHILDREN(rack_sysctl_root),
1289             OID_AUTO,
1290             "measure",
1291             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1292             "Measure related controls");
1293         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1294             SYSCTL_CHILDREN(rack_measure),
1295             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1296             &rack_wma_divisor, 8,
1297             "When doing b/w calculation what is the  divisor for the WMA");
1298         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1299             SYSCTL_CHILDREN(rack_measure),
1300             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1301             &rack_cwnd_block_ends_measure, 0,
1302             "Does a cwnd just-return end the measurement window (app limited)");
1303         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1304             SYSCTL_CHILDREN(rack_measure),
1305             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1306             &rack_rwnd_block_ends_measure, 0,
1307             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1308         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1309             SYSCTL_CHILDREN(rack_measure),
1310             OID_AUTO, "min_target", CTLFLAG_RW,
1311             &rack_def_data_window, 20,
1312             "What is the minimum target window (in mss) for a GP measurements");
1313         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1314             SYSCTL_CHILDREN(rack_measure),
1315             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1316             &rack_goal_bdp, 2,
1317             "What is the goal BDP to measure");
1318         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1319             SYSCTL_CHILDREN(rack_measure),
1320             OID_AUTO, "min_srtts", CTLFLAG_RW,
1321             &rack_min_srtts, 1,
1322             "What is the goal BDP to measure");
1323         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1324             SYSCTL_CHILDREN(rack_measure),
1325             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1326             &rack_min_measure_usec, 0,
1327             "What is the Minimum time time for a measurement if 0, this is off");
1328         /* Features */
1329         rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1330             SYSCTL_CHILDREN(rack_sysctl_root),
1331             OID_AUTO,
1332             "features",
1333             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1334             "Feature controls");
1335         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1336             SYSCTL_CHILDREN(rack_features),
1337             OID_AUTO, "cmpack", CTLFLAG_RW,
1338             &rack_use_cmp_acks, 1,
1339             "Should RACK have LRO send compressed acks");
1340         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1341             SYSCTL_CHILDREN(rack_features),
1342             OID_AUTO, "fsb", CTLFLAG_RW,
1343             &rack_use_fsb, 1,
1344             "Should RACK use the fast send block?");
1345         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1346             SYSCTL_CHILDREN(rack_features),
1347             OID_AUTO, "rfo", CTLFLAG_RW,
1348             &rack_use_rfo, 1,
1349             "Should RACK use rack_fast_output()?");
1350         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1351             SYSCTL_CHILDREN(rack_features),
1352             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1353             &rack_use_rsm_rfo, 1,
1354             "Should RACK use rack_fast_rsm_output()?");
1355         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1356             SYSCTL_CHILDREN(rack_features),
1357             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1358             &rack_enable_mqueue_for_nonpaced, 0,
1359             "Should RACK use mbuf queuing for non-paced connections");
1360         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1361             SYSCTL_CHILDREN(rack_features),
1362             OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1363             &rack_do_hystart, 0,
1364             "Should RACK enable HyStart++ on connections?");
1365         /* Misc rack controls */
1366         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1367             SYSCTL_CHILDREN(rack_sysctl_root),
1368             OID_AUTO,
1369             "misc",
1370             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1371             "Misc related controls");
1372 #ifdef TCP_ACCOUNTING
1373         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1374             SYSCTL_CHILDREN(rack_misc),
1375             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1376             &rack_tcp_accounting, 0,
1377             "Should we turn on TCP accounting for all rack sessions?");
1378 #endif
1379         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1380             SYSCTL_CHILDREN(rack_misc),
1381             OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1382             &rack_apply_rtt_with_reduced_conf, 0,
1383             "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1384         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1385             SYSCTL_CHILDREN(rack_misc),
1386             OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1387             &rack_dsack_std_based, 3,
1388             "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1389         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1390             SYSCTL_CHILDREN(rack_misc),
1391             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1392             &rack_prr_addbackmax, 2,
1393             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1394         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1395             SYSCTL_CHILDREN(rack_misc),
1396             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1397             &rack_stats_gets_ms_rtt, 1,
1398             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1399         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1400             SYSCTL_CHILDREN(rack_misc),
1401             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1402             &rack_client_low_buf, 0,
1403             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1404         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1405             SYSCTL_CHILDREN(rack_misc),
1406             OID_AUTO, "defprofile", CTLFLAG_RW,
1407             &rack_def_profile, 0,
1408             "Should RACK use a default profile (0=no, num == profile num)?");
1409         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1410             SYSCTL_CHILDREN(rack_misc),
1411             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1412             &rack_enable_shared_cwnd, 1,
1413             "Should RACK try to use the shared cwnd on connections where allowed");
1414         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1415             SYSCTL_CHILDREN(rack_misc),
1416             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1417             &rack_limits_scwnd, 1,
1418             "Should RACK place low end time limits on the shared cwnd feature");
1419         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1420             SYSCTL_CHILDREN(rack_misc),
1421             OID_AUTO, "iMac_dack", CTLFLAG_RW,
1422             &rack_use_imac_dack, 0,
1423             "Should RACK try to emulate iMac delayed ack");
1424         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1425             SYSCTL_CHILDREN(rack_misc),
1426             OID_AUTO, "no_prr", CTLFLAG_RW,
1427             &rack_disable_prr, 0,
1428             "Should RACK not use prr and only pace (must have pacing on)");
1429         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1430             SYSCTL_CHILDREN(rack_misc),
1431             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1432             &rack_verbose_logging, 0,
1433             "Should RACK black box logging be verbose");
1434         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1435             SYSCTL_CHILDREN(rack_misc),
1436             OID_AUTO, "data_after_close", CTLFLAG_RW,
1437             &rack_ignore_data_after_close, 1,
1438             "Do we hold off sending a RST until all pending data is ack'd");
1439         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1440             SYSCTL_CHILDREN(rack_misc),
1441             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1442             &rack_sack_not_required, 1,
1443             "Do we allow rack to run on connections not supporting SACK");
1444         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1445             SYSCTL_CHILDREN(rack_misc),
1446             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1447             &rack_send_a_lot_in_prr, 1,
1448             "Send a lot in prr");
1449         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1450             SYSCTL_CHILDREN(rack_misc),
1451             OID_AUTO, "autoscale", CTLFLAG_RW,
1452             &rack_autosndbuf_inc, 20,
1453             "What percentage should rack scale up its snd buffer by?");
1454         /* Sack Attacker detection stuff */
1455         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1456             SYSCTL_CHILDREN(rack_attack),
1457             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1458             &rack_highest_sack_thresh_seen, 0,
1459             "Highest sack to ack ratio seen");
1460         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1461             SYSCTL_CHILDREN(rack_attack),
1462             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1463             &rack_highest_move_thresh_seen, 0,
1464             "Highest move to non-move ratio seen");
1465         rack_ack_total = counter_u64_alloc(M_WAITOK);
1466         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1467             SYSCTL_CHILDREN(rack_attack),
1468             OID_AUTO, "acktotal", CTLFLAG_RD,
1469             &rack_ack_total,
1470             "Total number of Ack's");
1471         rack_express_sack = counter_u64_alloc(M_WAITOK);
1472         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1473             SYSCTL_CHILDREN(rack_attack),
1474             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1475             &rack_express_sack,
1476             "Total expresss number of Sack's");
1477         rack_sack_total = counter_u64_alloc(M_WAITOK);
1478         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1479             SYSCTL_CHILDREN(rack_attack),
1480             OID_AUTO, "sacktotal", CTLFLAG_RD,
1481             &rack_sack_total,
1482             "Total number of SACKs");
1483         rack_move_none = counter_u64_alloc(M_WAITOK);
1484         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1485             SYSCTL_CHILDREN(rack_attack),
1486             OID_AUTO, "move_none", CTLFLAG_RD,
1487             &rack_move_none,
1488             "Total number of SACK index reuse of positions under threshold");
1489         rack_move_some = counter_u64_alloc(M_WAITOK);
1490         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1491             SYSCTL_CHILDREN(rack_attack),
1492             OID_AUTO, "move_some", CTLFLAG_RD,
1493             &rack_move_some,
1494             "Total number of SACK index reuse of positions over threshold");
1495         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1496         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1497             SYSCTL_CHILDREN(rack_attack),
1498             OID_AUTO, "attacks", CTLFLAG_RD,
1499             &rack_sack_attacks_detected,
1500             "Total number of SACK attackers that had sack disabled");
1501         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1502         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1503             SYSCTL_CHILDREN(rack_attack),
1504             OID_AUTO, "reversed", CTLFLAG_RD,
1505             &rack_sack_attacks_reversed,
1506             "Total number of SACK attackers that were later determined false positive");
1507         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1508         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1509             SYSCTL_CHILDREN(rack_attack),
1510             OID_AUTO, "nextmerge", CTLFLAG_RD,
1511             &rack_sack_used_next_merge,
1512             "Total number of times we used the next merge");
1513         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1514         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1515             SYSCTL_CHILDREN(rack_attack),
1516             OID_AUTO, "prevmerge", CTLFLAG_RD,
1517             &rack_sack_used_prev_merge,
1518             "Total number of times we used the prev merge");
1519         /* Counters */
1520         rack_fto_send = counter_u64_alloc(M_WAITOK);
1521         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1522             SYSCTL_CHILDREN(rack_counters),
1523             OID_AUTO, "fto_send", CTLFLAG_RD,
1524             &rack_fto_send, "Total number of rack_fast_output sends");
1525         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1526         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1527             SYSCTL_CHILDREN(rack_counters),
1528             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1529             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1530         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1531         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1532             SYSCTL_CHILDREN(rack_counters),
1533             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1534             &rack_nfto_resend, "Total number of rack_output retransmissions");
1535         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1536         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1537             SYSCTL_CHILDREN(rack_counters),
1538             OID_AUTO, "nfto_send", CTLFLAG_RD,
1539             &rack_non_fto_send, "Total number of rack_output first sends");
1540         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1541         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1542             SYSCTL_CHILDREN(rack_counters),
1543             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1544             &rack_extended_rfo, "Total number of times we extended rfo");
1545
1546         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1547         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1548             SYSCTL_CHILDREN(rack_counters),
1549             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1550             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1551         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1552
1553         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1554             SYSCTL_CHILDREN(rack_counters),
1555             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1556             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1557         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1558         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1559             SYSCTL_CHILDREN(rack_counters),
1560             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1561             &rack_tlp_tot,
1562             "Total number of tail loss probe expirations");
1563         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1564         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1565             SYSCTL_CHILDREN(rack_counters),
1566             OID_AUTO, "tlp_new", CTLFLAG_RD,
1567             &rack_tlp_newdata,
1568             "Total number of tail loss probe sending new data");
1569         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1570         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1571             SYSCTL_CHILDREN(rack_counters),
1572             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1573             &rack_tlp_retran,
1574             "Total number of tail loss probe sending retransmitted data");
1575         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1576         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1577             SYSCTL_CHILDREN(rack_counters),
1578             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1579             &rack_tlp_retran_bytes,
1580             "Total bytes of tail loss probe sending retransmitted data");
1581         rack_to_tot = counter_u64_alloc(M_WAITOK);
1582         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1583             SYSCTL_CHILDREN(rack_counters),
1584             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1585             &rack_to_tot,
1586             "Total number of times the rack to expired");
1587         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1588         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1589             SYSCTL_CHILDREN(rack_counters),
1590             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1591             &rack_saw_enobuf,
1592             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1593         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1594         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1595             SYSCTL_CHILDREN(rack_counters),
1596             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1597             &rack_saw_enobuf_hw,
1598             "Total number of times a send returned enobuf for hdwr paced connections");
1599         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1600         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1601             SYSCTL_CHILDREN(rack_counters),
1602             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1603             &rack_saw_enetunreach,
1604             "Total number of times a send received a enetunreachable");
1605         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1606         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1607             SYSCTL_CHILDREN(rack_counters),
1608             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1609             &rack_hot_alloc,
1610             "Total allocations from the top of our list");
1611         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1612         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1613             SYSCTL_CHILDREN(rack_counters),
1614             OID_AUTO, "allocs", CTLFLAG_RD,
1615             &rack_to_alloc,
1616             "Total allocations of tracking structures");
1617         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1618         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1619             SYSCTL_CHILDREN(rack_counters),
1620             OID_AUTO, "allochard", CTLFLAG_RD,
1621             &rack_to_alloc_hard,
1622             "Total allocations done with sleeping the hard way");
1623         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1624         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1625             SYSCTL_CHILDREN(rack_counters),
1626             OID_AUTO, "allocemerg", CTLFLAG_RD,
1627             &rack_to_alloc_emerg,
1628             "Total allocations done from emergency cache");
1629         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1630         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1631             SYSCTL_CHILDREN(rack_counters),
1632             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1633             &rack_to_alloc_limited,
1634             "Total allocations dropped due to limit");
1635         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1636         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1637             SYSCTL_CHILDREN(rack_counters),
1638             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1639             &rack_alloc_limited_conns,
1640             "Connections with allocations dropped due to limit");
1641         rack_split_limited = counter_u64_alloc(M_WAITOK);
1642         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1643             SYSCTL_CHILDREN(rack_counters),
1644             OID_AUTO, "split_limited", CTLFLAG_RD,
1645             &rack_split_limited,
1646             "Split allocations dropped due to limit");
1647         rack_persists_sends = counter_u64_alloc(M_WAITOK);
1648         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1649             SYSCTL_CHILDREN(rack_counters),
1650             OID_AUTO, "persist_sends", CTLFLAG_RD,
1651             &rack_persists_sends,
1652             "Number of times we sent a persist probe");
1653         rack_persists_acks = counter_u64_alloc(M_WAITOK);
1654         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1655             SYSCTL_CHILDREN(rack_counters),
1656             OID_AUTO, "persist_acks", CTLFLAG_RD,
1657             &rack_persists_acks,
1658             "Number of times a persist probe was acked");
1659         rack_persists_loss = counter_u64_alloc(M_WAITOK);
1660         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1661             SYSCTL_CHILDREN(rack_counters),
1662             OID_AUTO, "persist_loss", CTLFLAG_RD,
1663             &rack_persists_loss,
1664             "Number of times we detected a lost persist probe (no ack)");
1665         rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1666         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1667             SYSCTL_CHILDREN(rack_counters),
1668             OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1669             &rack_persists_lost_ends,
1670             "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1671 #ifdef INVARIANTS
1672         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1673         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674             SYSCTL_CHILDREN(rack_counters),
1675             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1676             &rack_adjust_map_bw,
1677             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1678 #endif
1679         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1680         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1681             SYSCTL_CHILDREN(rack_counters),
1682             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1683             &rack_multi_single_eq,
1684             "Number of compressed acks total represented");
1685         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1686         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1687             SYSCTL_CHILDREN(rack_counters),
1688             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1689             &rack_proc_non_comp_ack,
1690             "Number of non compresseds acks that we processed");
1691
1692
1693         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1694         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1695             SYSCTL_CHILDREN(rack_counters),
1696             OID_AUTO, "sack_long", CTLFLAG_RD,
1697             &rack_sack_proc_all,
1698             "Total times we had to walk whole list for sack processing");
1699         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1700         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1701             SYSCTL_CHILDREN(rack_counters),
1702             OID_AUTO, "sack_restart", CTLFLAG_RD,
1703             &rack_sack_proc_restart,
1704             "Total times we had to walk whole list due to a restart");
1705         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1706         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1707             SYSCTL_CHILDREN(rack_counters),
1708             OID_AUTO, "sack_short", CTLFLAG_RD,
1709             &rack_sack_proc_short,
1710             "Total times we took shortcut for sack processing");
1711         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1712         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1713             SYSCTL_CHILDREN(rack_attack),
1714             OID_AUTO, "skipacked", CTLFLAG_RD,
1715             &rack_sack_skipped_acked,
1716             "Total number of times we skipped previously sacked");
1717         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1718         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1719             SYSCTL_CHILDREN(rack_attack),
1720             OID_AUTO, "ofsplit", CTLFLAG_RD,
1721             &rack_sack_splits,
1722             "Total number of times we did the old fashion tree split");
1723         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1724         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1725             SYSCTL_CHILDREN(rack_counters),
1726             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1727             &rack_input_idle_reduces,
1728             "Total number of idle reductions on input");
1729         rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1730         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1731             SYSCTL_CHILDREN(rack_counters),
1732             OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1733             &rack_collapsed_win_seen,
1734             "Total number of collapsed window events seen (where our window shrinks)");
1735
1736         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1737         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1738             SYSCTL_CHILDREN(rack_counters),
1739             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1740             &rack_collapsed_win,
1741             "Total number of collapsed window events where we mark packets");
1742         rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1743         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744             SYSCTL_CHILDREN(rack_counters),
1745             OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1746             &rack_collapsed_win_rxt,
1747             "Total number of packets that were retransmitted");
1748         rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1749         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750             SYSCTL_CHILDREN(rack_counters),
1751             OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1752             &rack_collapsed_win_rxt_bytes,
1753             "Total number of bytes that were retransmitted");
1754         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1755         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756             SYSCTL_CHILDREN(rack_counters),
1757             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1758             &rack_try_scwnd,
1759             "Total number of scwnd attempts");
1760         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1761         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1762             OID_AUTO, "outsize", CTLFLAG_RD,
1763             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1764         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1765         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1766             OID_AUTO, "opts", CTLFLAG_RD,
1767             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1768         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1769             SYSCTL_CHILDREN(rack_sysctl_root),
1770             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1771             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1772 }
1773
1774 static __inline int
1775 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1776 {
1777         if (SEQ_GEQ(b->r_start, a->r_start) &&
1778             SEQ_LT(b->r_start, a->r_end)) {
1779                 /*
1780                  * The entry b is within the
1781                  * block a. i.e.:
1782                  * a --   |-------------|
1783                  * b --   |----|
1784                  * <or>
1785                  * b --       |------|
1786                  * <or>
1787                  * b --       |-----------|
1788                  */
1789                 return (0);
1790         } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1791                 /*
1792                  * b falls as either the next
1793                  * sequence block after a so a
1794                  * is said to be smaller than b.
1795                  * i.e:
1796                  * a --   |------|
1797                  * b --          |--------|
1798                  * or
1799                  * b --              |-----|
1800                  */
1801                 return (1);
1802         }
1803         /*
1804          * Whats left is where a is
1805          * larger than b. i.e:
1806          * a --         |-------|
1807          * b --  |---|
1808          * or even possibly
1809          * b --   |--------------|
1810          */
1811         return (-1);
1812 }
1813
1814 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1815 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1816
1817 static uint32_t
1818 rc_init_window(struct tcp_rack *rack)
1819 {
1820         uint32_t win;
1821
1822         if (rack->rc_init_win == 0) {
1823                 /*
1824                  * Nothing set by the user, use the system stack
1825                  * default.
1826                  */
1827                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1828         }
1829         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1830         return (win);
1831 }
1832
1833 static uint64_t
1834 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1835 {
1836         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1837                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1838         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1839                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1840         else
1841                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1842 }
1843
1844 static uint64_t
1845 rack_get_bw(struct tcp_rack *rack)
1846 {
1847         if (rack->use_fixed_rate) {
1848                 /* Return the fixed pacing rate */
1849                 return (rack_get_fixed_pacing_bw(rack));
1850         }
1851         if (rack->r_ctl.gp_bw == 0) {
1852                 /*
1853                  * We have yet no b/w measurement,
1854                  * if we have a user set initial bw
1855                  * return it. If we don't have that and
1856                  * we have an srtt, use the tcp IW (10) to
1857                  * calculate a fictional b/w over the SRTT
1858                  * which is more or less a guess. Note
1859                  * we don't use our IW from rack on purpose
1860                  * so if we have like IW=30, we are not
1861                  * calculating a "huge" b/w.
1862                  */
1863                 uint64_t bw, srtt;
1864                 if (rack->r_ctl.init_rate)
1865                         return (rack->r_ctl.init_rate);
1866
1867                 /* Has the user set a max peak rate? */
1868 #ifdef NETFLIX_PEAKRATE
1869                 if (rack->rc_tp->t_maxpeakrate)
1870                         return (rack->rc_tp->t_maxpeakrate);
1871 #endif
1872                 /* Ok lets come up with the IW guess, if we have a srtt */
1873                 if (rack->rc_tp->t_srtt == 0) {
1874                         /*
1875                          * Go with old pacing method
1876                          * i.e. burst mitigation only.
1877                          */
1878                         return (0);
1879                 }
1880                 /* Ok lets get the initial TCP win (not racks) */
1881                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1882                 srtt = (uint64_t)rack->rc_tp->t_srtt;
1883                 bw *= (uint64_t)USECS_IN_SECOND;
1884                 bw /= srtt;
1885                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1886                         bw = rack->r_ctl.bw_rate_cap;
1887                 return (bw);
1888         } else {
1889                 uint64_t bw;
1890
1891                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
1892                         /* Averaging is done, we can return the value */
1893                         bw = rack->r_ctl.gp_bw;
1894                 } else {
1895                         /* Still doing initial average must calculate */
1896                         bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
1897                 }
1898 #ifdef NETFLIX_PEAKRATE
1899                 if ((rack->rc_tp->t_maxpeakrate) &&
1900                     (bw > rack->rc_tp->t_maxpeakrate)) {
1901                         /* The user has set a peak rate to pace at
1902                          * don't allow us to pace faster than that.
1903                          */
1904                         return (rack->rc_tp->t_maxpeakrate);
1905                 }
1906 #endif
1907                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1908                         bw = rack->r_ctl.bw_rate_cap;
1909                 return (bw);
1910         }
1911 }
1912
1913 static uint16_t
1914 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1915 {
1916         if (rack->use_fixed_rate) {
1917                 return (100);
1918         } else if (rack->in_probe_rtt && (rsm == NULL))
1919                 return (rack->r_ctl.rack_per_of_gp_probertt);
1920         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
1921                   rack->r_ctl.rack_per_of_gp_rec)) {
1922                 if (rsm) {
1923                         /* a retransmission always use the recovery rate */
1924                         return (rack->r_ctl.rack_per_of_gp_rec);
1925                 } else if (rack->rack_rec_nonrxt_use_cr) {
1926                         /* Directed to use the configured rate */
1927                         goto configured_rate;
1928                 } else if (rack->rack_no_prr &&
1929                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1930                         /* No PRR, lets just use the b/w estimate only */
1931                         return (100);
1932                 } else {
1933                         /*
1934                          * Here we may have a non-retransmit but we
1935                          * have no overrides, so just use the recovery
1936                          * rate (prr is in effect).
1937                          */
1938                         return (rack->r_ctl.rack_per_of_gp_rec);
1939                 }
1940         }
1941 configured_rate:
1942         /* For the configured rate we look at our cwnd vs the ssthresh */
1943         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1944                 return (rack->r_ctl.rack_per_of_gp_ss);
1945         else
1946                 return (rack->r_ctl.rack_per_of_gp_ca);
1947 }
1948
1949 static void
1950 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
1951 {
1952         /*
1953          * Types of logs (mod value)
1954          * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
1955          * 2 = a dsack round begins, persist is reset to 16.
1956          * 3 = a dsack round ends
1957          * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
1958          * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
1959          * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
1960          */
1961         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1962                 union tcp_log_stackspecific log;
1963                 struct timeval tv;
1964
1965                 memset(&log, 0, sizeof(log));
1966                 log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
1967                 log.u_bbr.flex1 <<= 1;
1968                 log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
1969                 log.u_bbr.flex1 <<= 1;
1970                 log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
1971                 log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
1972                 log.u_bbr.flex3 = rack->r_ctl.num_dsack;
1973                 log.u_bbr.flex4 = flex4;
1974                 log.u_bbr.flex5 = flex5;
1975                 log.u_bbr.flex6 = flex6;
1976                 log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
1977                 log.u_bbr.flex8 = mod;
1978                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1979                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
1980                     &rack->rc_inp->inp_socket->so_rcv,
1981                     &rack->rc_inp->inp_socket->so_snd,
1982                     RACK_DSACK_HANDLING, 0,
1983                     0, &log, false, &tv);
1984         }
1985 }
1986
1987 static void
1988 rack_log_hdwr_pacing(struct tcp_rack *rack,
1989                      uint64_t rate, uint64_t hw_rate, int line,
1990                      int error, uint16_t mod)
1991 {
1992         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1993                 union tcp_log_stackspecific log;
1994                 struct timeval tv;
1995                 const struct ifnet *ifp;
1996
1997                 memset(&log, 0, sizeof(log));
1998                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
1999                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2000                 if (rack->r_ctl.crte) {
2001                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2002                 } else if (rack->rc_inp->inp_route.ro_nh &&
2003                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2004                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2005                 } else
2006                         ifp = NULL;
2007                 if (ifp) {
2008                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2009                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2010                 }
2011                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2012                 log.u_bbr.bw_inuse = rate;
2013                 log.u_bbr.flex5 = line;
2014                 log.u_bbr.flex6 = error;
2015                 log.u_bbr.flex7 = mod;
2016                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2017                 log.u_bbr.flex8 = rack->use_fixed_rate;
2018                 log.u_bbr.flex8 <<= 1;
2019                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2020                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2021                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2022                 if (rack->r_ctl.crte)
2023                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2024                 else
2025                         log.u_bbr.cur_del_rate = 0;
2026                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2027                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2028                     &rack->rc_inp->inp_socket->so_rcv,
2029                     &rack->rc_inp->inp_socket->so_snd,
2030                     BBR_LOG_HDWR_PACE, 0,
2031                     0, &log, false, &tv);
2032         }
2033 }
2034
2035 static uint64_t
2036 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2037 {
2038         /*
2039          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2040          */
2041         uint64_t bw_est, high_rate;
2042         uint64_t gain;
2043
2044         gain = (uint64_t)rack_get_output_gain(rack, rsm);
2045         bw_est = bw * gain;
2046         bw_est /= (uint64_t)100;
2047         /* Never fall below the minimum (def 64kbps) */
2048         if (bw_est < RACK_MIN_BW)
2049                 bw_est = RACK_MIN_BW;
2050         if (rack->r_rack_hw_rate_caps) {
2051                 /* Rate caps are in place */
2052                 if (rack->r_ctl.crte != NULL) {
2053                         /* We have a hdwr rate already */
2054                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2055                         if (bw_est >= high_rate) {
2056                                 /* We are capping bw at the highest rate table entry */
2057                                 rack_log_hdwr_pacing(rack,
2058                                                      bw_est, high_rate, __LINE__,
2059                                                      0, 3);
2060                                 bw_est = high_rate;
2061                                 if (capped)
2062                                         *capped = 1;
2063                         }
2064                 } else if ((rack->rack_hdrw_pacing == 0) &&
2065                            (rack->rack_hdw_pace_ena) &&
2066                            (rack->rack_attempt_hdwr_pace == 0) &&
2067                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2068                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2069                         /*
2070                          * Special case, we have not yet attempted hardware
2071                          * pacing, and yet we may, when we do, find out if we are
2072                          * above the highest rate. We need to know the maxbw for the interface
2073                          * in question (if it supports ratelimiting). We get back
2074                          * a 0, if the interface is not found in the RL lists.
2075                          */
2076                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2077                         if (high_rate) {
2078                                 /* Yep, we have a rate is it above this rate? */
2079                                 if (bw_est > high_rate) {
2080                                         bw_est = high_rate;
2081                                         if (capped)
2082                                                 *capped = 1;
2083                                 }
2084                         }
2085                 }
2086         }
2087         return (bw_est);
2088 }
2089
2090 static void
2091 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2092 {
2093         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2094                 union tcp_log_stackspecific log;
2095                 struct timeval tv;
2096
2097                 if ((mod != 1) && (rack_verbose_logging == 0)) {
2098                         /*
2099                          * We get 3 values currently for mod
2100                          * 1 - We are retransmitting and this tells the reason.
2101                          * 2 - We are clearing a dup-ack count.
2102                          * 3 - We are incrementing a dup-ack count.
2103                          *
2104                          * The clear/increment are only logged
2105                          * if you have BBverbose on.
2106                          */
2107                         return;
2108                 }
2109                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2110                 log.u_bbr.flex1 = tsused;
2111                 log.u_bbr.flex2 = thresh;
2112                 log.u_bbr.flex3 = rsm->r_flags;
2113                 log.u_bbr.flex4 = rsm->r_dupack;
2114                 log.u_bbr.flex5 = rsm->r_start;
2115                 log.u_bbr.flex6 = rsm->r_end;
2116                 log.u_bbr.flex8 = mod;
2117                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2118                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2119                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2120                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2121                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2122                 log.u_bbr.pacing_gain = rack->r_must_retran;
2123                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2124                     &rack->rc_inp->inp_socket->so_rcv,
2125                     &rack->rc_inp->inp_socket->so_snd,
2126                     BBR_LOG_SETTINGS_CHG, 0,
2127                     0, &log, false, &tv);
2128         }
2129 }
2130
2131 static void
2132 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2133 {
2134         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2135                 union tcp_log_stackspecific log;
2136                 struct timeval tv;
2137
2138                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2139                 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2140                 log.u_bbr.flex2 = to;
2141                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2142                 log.u_bbr.flex4 = slot;
2143                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2144                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2145                 log.u_bbr.flex7 = rack->rc_in_persist;
2146                 log.u_bbr.flex8 = which;
2147                 if (rack->rack_no_prr)
2148                         log.u_bbr.pkts_out = 0;
2149                 else
2150                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2151                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2152                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2153                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2154                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2155                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2156                 log.u_bbr.pacing_gain = rack->r_must_retran;
2157                 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2158                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2159                 log.u_bbr.lost = rack_rto_min;
2160                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2161                     &rack->rc_inp->inp_socket->so_rcv,
2162                     &rack->rc_inp->inp_socket->so_snd,
2163                     BBR_LOG_TIMERSTAR, 0,
2164                     0, &log, false, &tv);
2165         }
2166 }
2167
2168 static void
2169 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2170 {
2171         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2172                 union tcp_log_stackspecific log;
2173                 struct timeval tv;
2174
2175                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2176                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2177                 log.u_bbr.flex8 = to_num;
2178                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2179                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2180                 if (rsm == NULL)
2181                         log.u_bbr.flex3 = 0;
2182                 else
2183                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2184                 if (rack->rack_no_prr)
2185                         log.u_bbr.flex5 = 0;
2186                 else
2187                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2188                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2189                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2190                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2191                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2192                 log.u_bbr.pacing_gain = rack->r_must_retran;
2193                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2194                     &rack->rc_inp->inp_socket->so_rcv,
2195                     &rack->rc_inp->inp_socket->so_snd,
2196                     BBR_LOG_RTO, 0,
2197                     0, &log, false, &tv);
2198         }
2199 }
2200
2201 static void
2202 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2203                  struct rack_sendmap *prev,
2204                  struct rack_sendmap *rsm,
2205                  struct rack_sendmap *next,
2206                  int flag, uint32_t th_ack, int line)
2207 {
2208         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2209                 union tcp_log_stackspecific log;
2210                 struct timeval tv;
2211
2212                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2213                 log.u_bbr.flex8 = flag;
2214                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2215                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2216                 log.u_bbr.delRate = (uint64_t)rsm;
2217                 log.u_bbr.rttProp = (uint64_t)next;
2218                 log.u_bbr.flex7 = 0;
2219                 if (prev) {
2220                         log.u_bbr.flex1 = prev->r_start;
2221                         log.u_bbr.flex2 = prev->r_end;
2222                         log.u_bbr.flex7 |= 0x4;
2223                 }
2224                 if (rsm) {
2225                         log.u_bbr.flex3 = rsm->r_start;
2226                         log.u_bbr.flex4 = rsm->r_end;
2227                         log.u_bbr.flex7 |= 0x2;
2228                 }
2229                 if (next) {
2230                         log.u_bbr.flex5 = next->r_start;
2231                         log.u_bbr.flex6 = next->r_end;
2232                         log.u_bbr.flex7 |= 0x1;
2233                 }
2234                 log.u_bbr.applimited = line;
2235                 log.u_bbr.pkts_out = th_ack;
2236                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2237                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2238                 if (rack->rack_no_prr)
2239                         log.u_bbr.lost = 0;
2240                 else
2241                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2242                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2243                     &rack->rc_inp->inp_socket->so_rcv,
2244                     &rack->rc_inp->inp_socket->so_snd,
2245                     TCP_LOG_MAPCHG, 0,
2246                     0, &log, false, &tv);
2247         }
2248 }
2249
2250 static void
2251 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2252                  struct rack_sendmap *rsm, int conf)
2253 {
2254         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2255                 union tcp_log_stackspecific log;
2256                 struct timeval tv;
2257                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2258                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2259                 log.u_bbr.flex1 = t;
2260                 log.u_bbr.flex2 = len;
2261                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2262                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2263                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2264                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2265                 log.u_bbr.flex7 = conf;
2266                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2267                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2268                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2269                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2270                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2271                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2272                 if (rsm) {
2273                         log.u_bbr.pkt_epoch = rsm->r_start;
2274                         log.u_bbr.lost = rsm->r_end;
2275                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2276                         /* We loose any upper of the 24 bits */
2277                         log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2278                 } else {
2279                         /* Its a SYN */
2280                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2281                         log.u_bbr.lost = 0;
2282                         log.u_bbr.cwnd_gain = 0;
2283                         log.u_bbr.pacing_gain = 0;
2284                 }
2285                 /* Write out general bits of interest rrs here */
2286                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2287                 log.u_bbr.use_lt_bw <<= 1;
2288                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2289                 log.u_bbr.use_lt_bw <<= 1;
2290                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2291                 log.u_bbr.use_lt_bw <<= 1;
2292                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2293                 log.u_bbr.use_lt_bw <<= 1;
2294                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2295                 log.u_bbr.use_lt_bw <<= 1;
2296                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2297                 log.u_bbr.use_lt_bw <<= 1;
2298                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2299                 log.u_bbr.use_lt_bw <<= 1;
2300                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2301                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2302                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2303                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2304                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2305                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2306                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2307                 log.u_bbr.bw_inuse <<= 32;
2308                 if (rsm)
2309                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2310                 TCP_LOG_EVENTP(tp, NULL,
2311                     &rack->rc_inp->inp_socket->so_rcv,
2312                     &rack->rc_inp->inp_socket->so_snd,
2313                     BBR_LOG_BBRRTT, 0,
2314                     0, &log, false, &tv);
2315
2316
2317         }
2318 }
2319
2320 static void
2321 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2322 {
2323         /*
2324          * Log the rtt sample we are
2325          * applying to the srtt algorithm in
2326          * useconds.
2327          */
2328         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2329                 union tcp_log_stackspecific log;
2330                 struct timeval tv;
2331
2332                 /* Convert our ms to a microsecond */
2333                 memset(&log, 0, sizeof(log));
2334                 log.u_bbr.flex1 = rtt;
2335                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2336                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2337                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2338                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2339                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2340                 log.u_bbr.flex7 = 1;
2341                 log.u_bbr.flex8 = rack->sack_attack_disable;
2342                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2343                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2344                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2345                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2346                 log.u_bbr.pacing_gain = rack->r_must_retran;
2347                 /*
2348                  * We capture in delRate the upper 32 bits as
2349                  * the confidence level we had declared, and the
2350                  * lower 32 bits as the actual RTT using the arrival
2351                  * timestamp.
2352                  */
2353                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2354                 log.u_bbr.delRate <<= 32;
2355                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2356                 /* Lets capture all the things that make up t_rtxcur */
2357                 log.u_bbr.applimited = rack_rto_min;
2358                 log.u_bbr.epoch = rack_rto_max;
2359                 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2360                 log.u_bbr.lost = rack_rto_min;
2361                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2362                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2363                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2364                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2365                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2366                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2367                     &rack->rc_inp->inp_socket->so_rcv,
2368                     &rack->rc_inp->inp_socket->so_snd,
2369                     TCP_LOG_RTT, 0,
2370                     0, &log, false, &tv);
2371         }
2372 }
2373
2374 static void
2375 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2376 {
2377         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2378                 union tcp_log_stackspecific log;
2379                 struct timeval tv;
2380
2381                 /* Convert our ms to a microsecond */
2382                 memset(&log, 0, sizeof(log));
2383                 log.u_bbr.flex1 = rtt;
2384                 log.u_bbr.flex2 = send_time;
2385                 log.u_bbr.flex3 = ack_time;
2386                 log.u_bbr.flex4 = where;
2387                 log.u_bbr.flex7 = 2;
2388                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2389                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2390                     &rack->rc_inp->inp_socket->so_rcv,
2391                     &rack->rc_inp->inp_socket->so_snd,
2392                     TCP_LOG_RTT, 0,
2393                     0, &log, false, &tv);
2394         }
2395 }
2396
2397
2398
2399 static inline void
2400 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2401 {
2402         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2403                 union tcp_log_stackspecific log;
2404                 struct timeval tv;
2405
2406                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2407                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2408                 log.u_bbr.flex1 = line;
2409                 log.u_bbr.flex2 = tick;
2410                 log.u_bbr.flex3 = tp->t_maxunacktime;
2411                 log.u_bbr.flex4 = tp->t_acktime;
2412                 log.u_bbr.flex8 = event;
2413                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2414                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2415                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2416                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2417                 log.u_bbr.pacing_gain = rack->r_must_retran;
2418                 TCP_LOG_EVENTP(tp, NULL,
2419                     &rack->rc_inp->inp_socket->so_rcv,
2420                     &rack->rc_inp->inp_socket->so_snd,
2421                     BBR_LOG_PROGRESS, 0,
2422                     0, &log, false, &tv);
2423         }
2424 }
2425
2426 static void
2427 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2428 {
2429         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2430                 union tcp_log_stackspecific log;
2431
2432                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2433                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2434                 log.u_bbr.flex1 = slot;
2435                 if (rack->rack_no_prr)
2436                         log.u_bbr.flex2 = 0;
2437                 else
2438                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2439                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2440                 log.u_bbr.flex8 = rack->rc_in_persist;
2441                 log.u_bbr.timeStamp = cts;
2442                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2443                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2444                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2445                 log.u_bbr.pacing_gain = rack->r_must_retran;
2446                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2447                     &rack->rc_inp->inp_socket->so_rcv,
2448                     &rack->rc_inp->inp_socket->so_snd,
2449                     BBR_LOG_BBRSND, 0,
2450                     0, &log, false, tv);
2451         }
2452 }
2453
2454 static void
2455 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2456 {
2457         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2458                 union tcp_log_stackspecific log;
2459                 struct timeval tv;
2460
2461                 memset(&log, 0, sizeof(log));
2462                 log.u_bbr.flex1 = did_out;
2463                 log.u_bbr.flex2 = nxt_pkt;
2464                 log.u_bbr.flex3 = way_out;
2465                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2466                 if (rack->rack_no_prr)
2467                         log.u_bbr.flex5 = 0;
2468                 else
2469                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2470                 log.u_bbr.flex6 = nsegs;
2471                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2472                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2473                 log.u_bbr.flex7 <<= 1;
2474                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2475                 log.u_bbr.flex7 <<= 1;
2476                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2477                 log.u_bbr.flex8 = rack->rc_in_persist;
2478                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2479                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2480                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2481                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2482                 log.u_bbr.use_lt_bw <<= 1;
2483                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2484                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2485                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2486                 log.u_bbr.pacing_gain = rack->r_must_retran;
2487                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2488                     &rack->rc_inp->inp_socket->so_rcv,
2489                     &rack->rc_inp->inp_socket->so_snd,
2490                     BBR_LOG_DOSEG_DONE, 0,
2491                     0, &log, false, &tv);
2492         }
2493 }
2494
2495 static void
2496 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2497 {
2498         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2499                 union tcp_log_stackspecific log;
2500                 struct timeval tv;
2501
2502                 memset(&log, 0, sizeof(log));
2503                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2504                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2505                 log.u_bbr.flex4 = arg1;
2506                 log.u_bbr.flex5 = arg2;
2507                 log.u_bbr.flex6 = arg3;
2508                 log.u_bbr.flex8 = frm;
2509                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2510                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2511                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2512                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2513                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2514                 log.u_bbr.pacing_gain = rack->r_must_retran;
2515                 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
2516                     &tptosocket(tp)->so_snd,
2517                     TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
2518         }
2519 }
2520
2521 static void
2522 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2523                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2524 {
2525         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2526                 union tcp_log_stackspecific log;
2527                 struct timeval tv;
2528
2529                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2530                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2531                 log.u_bbr.flex1 = slot;
2532                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2533                 log.u_bbr.flex4 = reason;
2534                 if (rack->rack_no_prr)
2535                         log.u_bbr.flex5 = 0;
2536                 else
2537                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2538                 log.u_bbr.flex7 = hpts_calling;
2539                 log.u_bbr.flex8 = rack->rc_in_persist;
2540                 log.u_bbr.lt_epoch = cwnd_to_use;
2541                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2542                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2543                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2544                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2545                 log.u_bbr.pacing_gain = rack->r_must_retran;
2546                 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2547                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2548                     &rack->rc_inp->inp_socket->so_rcv,
2549                     &rack->rc_inp->inp_socket->so_snd,
2550                     BBR_LOG_JUSTRET, 0,
2551                     tlen, &log, false, &tv);
2552         }
2553 }
2554
2555 static void
2556 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2557                    struct timeval *tv, uint32_t flags_on_entry)
2558 {
2559         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2560                 union tcp_log_stackspecific log;
2561
2562                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2563                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2564                 log.u_bbr.flex1 = line;
2565                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2566                 log.u_bbr.flex3 = flags_on_entry;
2567                 log.u_bbr.flex4 = us_cts;
2568                 if (rack->rack_no_prr)
2569                         log.u_bbr.flex5 = 0;
2570                 else
2571                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2572                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2573                 log.u_bbr.flex7 = hpts_removed;
2574                 log.u_bbr.flex8 = 1;
2575                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2576                 log.u_bbr.timeStamp = us_cts;
2577                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2578                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2579                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2580                 log.u_bbr.pacing_gain = rack->r_must_retran;
2581                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2582                     &rack->rc_inp->inp_socket->so_rcv,
2583                     &rack->rc_inp->inp_socket->so_snd,
2584                     BBR_LOG_TIMERCANC, 0,
2585                     0, &log, false, tv);
2586         }
2587 }
2588
2589 static void
2590 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2591                           uint32_t flex1, uint32_t flex2,
2592                           uint32_t flex3, uint32_t flex4,
2593                           uint32_t flex5, uint32_t flex6,
2594                           uint16_t flex7, uint8_t mod)
2595 {
2596         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2597                 union tcp_log_stackspecific log;
2598                 struct timeval tv;
2599
2600                 if (mod == 1) {
2601                         /* No you can't use 1, its for the real to cancel */
2602                         return;
2603                 }
2604                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2605                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2606                 log.u_bbr.flex1 = flex1;
2607                 log.u_bbr.flex2 = flex2;
2608                 log.u_bbr.flex3 = flex3;
2609                 log.u_bbr.flex4 = flex4;
2610                 log.u_bbr.flex5 = flex5;
2611                 log.u_bbr.flex6 = flex6;
2612                 log.u_bbr.flex7 = flex7;
2613                 log.u_bbr.flex8 = mod;
2614                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2615                     &rack->rc_inp->inp_socket->so_rcv,
2616                     &rack->rc_inp->inp_socket->so_snd,
2617                     BBR_LOG_TIMERCANC, 0,
2618                     0, &log, false, &tv);
2619         }
2620 }
2621
2622 static void
2623 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2624 {
2625         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2626                 union tcp_log_stackspecific log;
2627                 struct timeval tv;
2628
2629                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2630                 log.u_bbr.flex1 = timers;
2631                 log.u_bbr.flex2 = ret;
2632                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2633                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2634                 log.u_bbr.flex5 = cts;
2635                 if (rack->rack_no_prr)
2636                         log.u_bbr.flex6 = 0;
2637                 else
2638                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2639                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2640                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2641                 log.u_bbr.pacing_gain = rack->r_must_retran;
2642                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2643                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2644                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2645                     &rack->rc_inp->inp_socket->so_rcv,
2646                     &rack->rc_inp->inp_socket->so_snd,
2647                     BBR_LOG_TO_PROCESS, 0,
2648                     0, &log, false, &tv);
2649         }
2650 }
2651
2652 static void
2653 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
2654 {
2655         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2656                 union tcp_log_stackspecific log;
2657                 struct timeval tv;
2658
2659                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2660                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2661                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2662                 if (rack->rack_no_prr)
2663                         log.u_bbr.flex3 = 0;
2664                 else
2665                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2666                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2667                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2668                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2669                 log.u_bbr.flex7 = line;
2670                 log.u_bbr.flex8 = frm;
2671                 log.u_bbr.pkts_out = orig_cwnd;
2672                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2673                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2674                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2675                 log.u_bbr.use_lt_bw <<= 1;
2676                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2677                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2678                     &rack->rc_inp->inp_socket->so_rcv,
2679                     &rack->rc_inp->inp_socket->so_snd,
2680                     BBR_LOG_BBRUPD, 0,
2681                     0, &log, false, &tv);
2682         }
2683 }
2684
2685 #ifdef NETFLIX_EXP_DETECTION
2686 static void
2687 rack_log_sad(struct tcp_rack *rack, int event)
2688 {
2689         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2690                 union tcp_log_stackspecific log;
2691                 struct timeval tv;
2692
2693                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2694                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2695                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2696                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2697                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2698                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2699                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2700                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2701                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2702                 log.u_bbr.lt_epoch |= rack->do_detection;
2703                 log.u_bbr.applimited = tcp_map_minimum;
2704                 log.u_bbr.flex7 = rack->sack_attack_disable;
2705                 log.u_bbr.flex8 = event;
2706                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2707                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2708                 log.u_bbr.delivered = tcp_sad_decay_val;
2709                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2710                     &rack->rc_inp->inp_socket->so_rcv,
2711                     &rack->rc_inp->inp_socket->so_snd,
2712                     TCP_SAD_DETECTION, 0,
2713                     0, &log, false, &tv);
2714         }
2715 }
2716 #endif
2717
2718 static void
2719 rack_counter_destroy(void)
2720 {
2721         counter_u64_free(rack_fto_send);
2722         counter_u64_free(rack_fto_rsm_send);
2723         counter_u64_free(rack_nfto_resend);
2724         counter_u64_free(rack_hw_pace_init_fail);
2725         counter_u64_free(rack_hw_pace_lost);
2726         counter_u64_free(rack_non_fto_send);
2727         counter_u64_free(rack_extended_rfo);
2728         counter_u64_free(rack_ack_total);
2729         counter_u64_free(rack_express_sack);
2730         counter_u64_free(rack_sack_total);
2731         counter_u64_free(rack_move_none);
2732         counter_u64_free(rack_move_some);
2733         counter_u64_free(rack_sack_attacks_detected);
2734         counter_u64_free(rack_sack_attacks_reversed);
2735         counter_u64_free(rack_sack_used_next_merge);
2736         counter_u64_free(rack_sack_used_prev_merge);
2737         counter_u64_free(rack_tlp_tot);
2738         counter_u64_free(rack_tlp_newdata);
2739         counter_u64_free(rack_tlp_retran);
2740         counter_u64_free(rack_tlp_retran_bytes);
2741         counter_u64_free(rack_to_tot);
2742         counter_u64_free(rack_saw_enobuf);
2743         counter_u64_free(rack_saw_enobuf_hw);
2744         counter_u64_free(rack_saw_enetunreach);
2745         counter_u64_free(rack_hot_alloc);
2746         counter_u64_free(rack_to_alloc);
2747         counter_u64_free(rack_to_alloc_hard);
2748         counter_u64_free(rack_to_alloc_emerg);
2749         counter_u64_free(rack_to_alloc_limited);
2750         counter_u64_free(rack_alloc_limited_conns);
2751         counter_u64_free(rack_split_limited);
2752         counter_u64_free(rack_multi_single_eq);
2753         counter_u64_free(rack_proc_non_comp_ack);
2754         counter_u64_free(rack_sack_proc_all);
2755         counter_u64_free(rack_sack_proc_restart);
2756         counter_u64_free(rack_sack_proc_short);
2757         counter_u64_free(rack_sack_skipped_acked);
2758         counter_u64_free(rack_sack_splits);
2759         counter_u64_free(rack_input_idle_reduces);
2760         counter_u64_free(rack_collapsed_win);
2761         counter_u64_free(rack_collapsed_win_rxt);
2762         counter_u64_free(rack_collapsed_win_rxt_bytes);
2763         counter_u64_free(rack_collapsed_win_seen);
2764         counter_u64_free(rack_try_scwnd);
2765         counter_u64_free(rack_persists_sends);
2766         counter_u64_free(rack_persists_acks);
2767         counter_u64_free(rack_persists_loss);
2768         counter_u64_free(rack_persists_lost_ends);
2769 #ifdef INVARIANTS
2770         counter_u64_free(rack_adjust_map_bw);
2771 #endif
2772         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2773         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2774 }
2775
2776 static struct rack_sendmap *
2777 rack_alloc(struct tcp_rack *rack)
2778 {
2779         struct rack_sendmap *rsm;
2780
2781         /*
2782          * First get the top of the list it in
2783          * theory is the "hottest" rsm we have,
2784          * possibly just freed by ack processing.
2785          */
2786         if (rack->rc_free_cnt > rack_free_cache) {
2787                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2788                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2789                 counter_u64_add(rack_hot_alloc, 1);
2790                 rack->rc_free_cnt--;
2791                 return (rsm);
2792         }
2793         /*
2794          * Once we get under our free cache we probably
2795          * no longer have a "hot" one available. Lets
2796          * get one from UMA.
2797          */
2798         rsm = uma_zalloc(rack_zone, M_NOWAIT);
2799         if (rsm) {
2800                 rack->r_ctl.rc_num_maps_alloced++;
2801                 counter_u64_add(rack_to_alloc, 1);
2802                 return (rsm);
2803         }
2804         /*
2805          * Dig in to our aux rsm's (the last two) since
2806          * UMA failed to get us one.
2807          */
2808         if (rack->rc_free_cnt) {
2809                 counter_u64_add(rack_to_alloc_emerg, 1);
2810                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2811                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2812                 rack->rc_free_cnt--;
2813                 return (rsm);
2814         }
2815         return (NULL);
2816 }
2817
2818 static struct rack_sendmap *
2819 rack_alloc_full_limit(struct tcp_rack *rack)
2820 {
2821         if ((V_tcp_map_entries_limit > 0) &&
2822             (rack->do_detection == 0) &&
2823             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2824                 counter_u64_add(rack_to_alloc_limited, 1);
2825                 if (!rack->alloc_limit_reported) {
2826                         rack->alloc_limit_reported = 1;
2827                         counter_u64_add(rack_alloc_limited_conns, 1);
2828                 }
2829                 return (NULL);
2830         }
2831         return (rack_alloc(rack));
2832 }
2833
2834 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2835 static struct rack_sendmap *
2836 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2837 {
2838         struct rack_sendmap *rsm;
2839
2840         if (limit_type) {
2841                 /* currently there is only one limit type */
2842                 if (V_tcp_map_split_limit > 0 &&
2843                     (rack->do_detection == 0) &&
2844                     rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2845                         counter_u64_add(rack_split_limited, 1);
2846                         if (!rack->alloc_limit_reported) {
2847                                 rack->alloc_limit_reported = 1;
2848                                 counter_u64_add(rack_alloc_limited_conns, 1);
2849                         }
2850                         return (NULL);
2851                 }
2852         }
2853
2854         /* allocate and mark in the limit type, if set */
2855         rsm = rack_alloc(rack);
2856         if (rsm != NULL && limit_type) {
2857                 rsm->r_limit_type = limit_type;
2858                 rack->r_ctl.rc_num_split_allocs++;
2859         }
2860         return (rsm);
2861 }
2862
2863 static void
2864 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2865 {
2866         if (rsm->r_flags & RACK_APP_LIMITED) {
2867                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2868                         rack->r_ctl.rc_app_limited_cnt--;
2869                 }
2870         }
2871         if (rsm->r_limit_type) {
2872                 /* currently there is only one limit type */
2873                 rack->r_ctl.rc_num_split_allocs--;
2874         }
2875         if (rsm == rack->r_ctl.rc_first_appl) {
2876                 if (rack->r_ctl.rc_app_limited_cnt == 0)
2877                         rack->r_ctl.rc_first_appl = NULL;
2878                 else {
2879                         /* Follow the next one out */
2880                         struct rack_sendmap fe;
2881
2882                         fe.r_start = rsm->r_nseq_appl;
2883                         rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2884                 }
2885         }
2886         if (rsm == rack->r_ctl.rc_resend)
2887                 rack->r_ctl.rc_resend = NULL;
2888         if (rsm == rack->r_ctl.rc_end_appl)
2889                 rack->r_ctl.rc_end_appl = NULL;
2890         if (rack->r_ctl.rc_tlpsend == rsm)
2891                 rack->r_ctl.rc_tlpsend = NULL;
2892         if (rack->r_ctl.rc_sacklast == rsm)
2893                 rack->r_ctl.rc_sacklast = NULL;
2894         memset(rsm, 0, sizeof(struct rack_sendmap));
2895         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
2896         rack->rc_free_cnt++;
2897 }
2898
2899 static void
2900 rack_free_trim(struct tcp_rack *rack)
2901 {
2902         struct rack_sendmap *rsm;
2903
2904         /*
2905          * Free up all the tail entries until
2906          * we get our list down to the limit.
2907          */
2908         while (rack->rc_free_cnt > rack_free_cache) {
2909                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
2910                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2911                 rack->rc_free_cnt--;
2912                 uma_zfree(rack_zone, rsm);
2913         }
2914 }
2915
2916
2917 static uint32_t
2918 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2919 {
2920         uint64_t srtt, bw, len, tim;
2921         uint32_t segsiz, def_len, minl;
2922
2923         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2924         def_len = rack_def_data_window * segsiz;
2925         if (rack->rc_gp_filled == 0) {
2926                 /*
2927                  * We have no measurement (IW is in flight?) so
2928                  * we can only guess using our data_window sysctl
2929                  * value (usually 20MSS).
2930                  */
2931                 return (def_len);
2932         }
2933         /*
2934          * Now we have a number of factors to consider.
2935          *
2936          * 1) We have a desired BDP which is usually
2937          *    at least 2.
2938          * 2) We have a minimum number of rtt's usually 1 SRTT
2939          *    but we allow it too to be more.
2940          * 3) We want to make sure a measurement last N useconds (if
2941          *    we have set rack_min_measure_usec.
2942          *
2943          * We handle the first concern here by trying to create a data
2944          * window of max(rack_def_data_window, DesiredBDP). The
2945          * second concern we handle in not letting the measurement
2946          * window end normally until at least the required SRTT's
2947          * have gone by which is done further below in
2948          * rack_enough_for_measurement(). Finally the third concern
2949          * we also handle here by calculating how long that time
2950          * would take at the current BW and then return the
2951          * max of our first calculation and that length. Note
2952          * that if rack_min_measure_usec is 0, we don't deal
2953          * with concern 3. Also for both Concern 1 and 3 an
2954          * application limited period could end the measurement
2955          * earlier.
2956          *
2957          * So lets calculate the BDP with the "known" b/w using
2958          * the SRTT has our rtt and then multiply it by the
2959          * goal.
2960          */
2961         bw = rack_get_bw(rack);
2962         srtt = (uint64_t)tp->t_srtt;
2963         len = bw * srtt;
2964         len /= (uint64_t)HPTS_USEC_IN_SEC;
2965         len *= max(1, rack_goal_bdp);
2966         /* Now we need to round up to the nearest MSS */
2967         len = roundup(len, segsiz);
2968         if (rack_min_measure_usec) {
2969                 /* Now calculate our min length for this b/w */
2970                 tim = rack_min_measure_usec;
2971                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
2972                 if (minl == 0)
2973                         minl = 1;
2974                 minl = roundup(minl, segsiz);
2975                 if (len < minl)
2976                         len = minl;
2977         }
2978         /*
2979          * Now if we have a very small window we want
2980          * to attempt to get the window that is
2981          * as small as possible. This happens on
2982          * low b/w connections and we don't want to
2983          * span huge numbers of rtt's between measurements.
2984          *
2985          * We basically include 2 over our "MIN window" so
2986          * that the measurement can be shortened (possibly) by
2987          * an ack'ed packet.
2988          */
2989         if (len < def_len)
2990                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
2991         else
2992                 return (max((uint32_t)len, def_len));
2993
2994 }
2995
2996 static int
2997 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
2998 {
2999         uint32_t tim, srtts, segsiz;
3000
3001         /*
3002          * Has enough time passed for the GP measurement to be valid?
3003          */
3004         if ((tp->snd_max == tp->snd_una) ||
3005             (th_ack == tp->snd_max)){
3006                 /* All is acked */
3007                 *quality = RACK_QUALITY_ALLACKED;
3008                 return (1);
3009         }
3010         if (SEQ_LT(th_ack, tp->gput_seq)) {
3011                 /* Not enough bytes yet */
3012                 return (0);
3013         }
3014         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3015         if (SEQ_LT(th_ack, tp->gput_ack) &&
3016             ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3017                 /* Not enough bytes yet */
3018                 return (0);
3019         }
3020         if (rack->r_ctl.rc_first_appl &&
3021             (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3022                 /*
3023                  * We are up to the app limited send point
3024                  * we have to measure irrespective of the time..
3025                  */
3026                 *quality = RACK_QUALITY_APPLIMITED;
3027                 return (1);
3028         }
3029         /* Now what about time? */
3030         srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3031         tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3032         if (tim >= srtts) {
3033                 *quality = RACK_QUALITY_HIGH;
3034                 return (1);
3035         }
3036         /* Nope not even a full SRTT has passed */
3037         return (0);
3038 }
3039
3040 static void
3041 rack_log_timely(struct tcp_rack *rack,
3042                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3043                 uint64_t up_bnd, int line, uint8_t method)
3044 {
3045         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3046                 union tcp_log_stackspecific log;
3047                 struct timeval tv;
3048
3049                 memset(&log, 0, sizeof(log));
3050                 log.u_bbr.flex1 = logged;
3051                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3052                 log.u_bbr.flex2 <<= 4;
3053                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3054                 log.u_bbr.flex2 <<= 4;
3055                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3056                 log.u_bbr.flex2 <<= 4;
3057                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3058                 log.u_bbr.flex3 = rack->rc_gp_incr;
3059                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3060                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3061                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3062                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3063                 log.u_bbr.flex8 = method;
3064                 log.u_bbr.cur_del_rate = cur_bw;
3065                 log.u_bbr.delRate = low_bnd;
3066                 log.u_bbr.bw_inuse = up_bnd;
3067                 log.u_bbr.rttProp = rack_get_bw(rack);
3068                 log.u_bbr.pkt_epoch = line;
3069                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3070                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3071                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3072                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3073                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3074                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3075                 log.u_bbr.cwnd_gain <<= 1;
3076                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3077                 log.u_bbr.cwnd_gain <<= 1;
3078                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3079                 log.u_bbr.cwnd_gain <<= 1;
3080                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3081                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3082                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3083                     &rack->rc_inp->inp_socket->so_rcv,
3084                     &rack->rc_inp->inp_socket->so_snd,
3085                     TCP_TIMELY_WORK, 0,
3086                     0, &log, false, &tv);
3087         }
3088 }
3089
3090 static int
3091 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3092 {
3093         /*
3094          * Before we increase we need to know if
3095          * the estimate just made was less than
3096          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3097          *
3098          * If we already are pacing at a fast enough
3099          * rate to push us faster there is no sense of
3100          * increasing.
3101          *
3102          * We first caculate our actual pacing rate (ss or ca multiplier
3103          * times our cur_bw).
3104          *
3105          * Then we take the last measured rate and multipy by our
3106          * maximum pacing overage to give us a max allowable rate.
3107          *
3108          * If our act_rate is smaller than our max_allowable rate
3109          * then we should increase. Else we should hold steady.
3110          *
3111          */
3112         uint64_t act_rate, max_allow_rate;
3113
3114         if (rack_timely_no_stopping)
3115                 return (1);
3116
3117         if ((cur_bw == 0) || (last_bw_est == 0)) {
3118                 /*
3119                  * Initial startup case or
3120                  * everything is acked case.
3121                  */
3122                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3123                                 __LINE__, 9);
3124                 return (1);
3125         }
3126         if (mult <= 100) {
3127                 /*
3128                  * We can always pace at or slightly above our rate.
3129                  */
3130                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3131                                 __LINE__, 9);
3132                 return (1);
3133         }
3134         act_rate = cur_bw * (uint64_t)mult;
3135         act_rate /= 100;
3136         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3137         max_allow_rate /= 100;
3138         if (act_rate < max_allow_rate) {
3139                 /*
3140                  * Here the rate we are actually pacing at
3141                  * is smaller than 10% above our last measurement.
3142                  * This means we are pacing below what we would
3143                  * like to try to achieve (plus some wiggle room).
3144                  */
3145                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3146                                 __LINE__, 9);
3147                 return (1);
3148         } else {
3149                 /*
3150                  * Here we are already pacing at least rack_max_per_above(10%)
3151                  * what we are getting back. This indicates most likely
3152                  * that we are being limited (cwnd/rwnd/app) and can't
3153                  * get any more b/w. There is no sense of trying to
3154                  * raise up the pacing rate its not speeding us up
3155                  * and we already are pacing faster than we are getting.
3156                  */
3157                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3158                                 __LINE__, 8);
3159                 return (0);
3160         }
3161 }
3162
3163 static void
3164 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3165 {
3166         /*
3167          * When we drag bottom, we want to assure
3168          * that no multiplier is below 1.0, if so
3169          * we want to restore it to at least that.
3170          */
3171         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3172                 /* This is unlikely we usually do not touch recovery */
3173                 rack->r_ctl.rack_per_of_gp_rec = 100;
3174         }
3175         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3176                 rack->r_ctl.rack_per_of_gp_ca = 100;
3177         }
3178         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3179                 rack->r_ctl.rack_per_of_gp_ss = 100;
3180         }
3181 }
3182
3183 static void
3184 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3185 {
3186         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3187                 rack->r_ctl.rack_per_of_gp_ca = 100;
3188         }
3189         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3190                 rack->r_ctl.rack_per_of_gp_ss = 100;
3191         }
3192 }
3193
3194 static void
3195 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3196 {
3197         int32_t  calc, logged, plus;
3198
3199         logged = 0;
3200
3201         if (override) {
3202                 /*
3203                  * override is passed when we are
3204                  * loosing b/w and making one last
3205                  * gasp at trying to not loose out
3206                  * to a new-reno flow.
3207                  */
3208                 goto extra_boost;
3209         }
3210         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3211         if (rack->rc_gp_incr &&
3212             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3213                 /*
3214                  * Reset and get 5 strokes more before the boost. Note
3215                  * that the count is 0 based so we have to add one.
3216                  */
3217 extra_boost:
3218                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3219                 rack->rc_gp_timely_inc_cnt = 0;
3220         } else
3221                 plus = (uint32_t)rack_gp_increase_per;
3222         /* Must be at least 1% increase for true timely increases */
3223         if ((plus < 1) &&
3224             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3225                 plus = 1;
3226         if (rack->rc_gp_saw_rec &&
3227             (rack->rc_gp_no_rec_chg == 0) &&
3228             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3229                                   rack->r_ctl.rack_per_of_gp_rec)) {
3230                 /* We have been in recovery ding it too */
3231                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3232                 if (calc > 0xffff)
3233                         calc = 0xffff;
3234                 logged |= 1;
3235                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3236                 if (rack_per_upper_bound_ss &&
3237                     (rack->rc_dragged_bottom == 0) &&
3238                     (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3239                         rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3240         }
3241         if (rack->rc_gp_saw_ca &&
3242             (rack->rc_gp_saw_ss == 0) &&
3243             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3244                                   rack->r_ctl.rack_per_of_gp_ca)) {
3245                 /* In CA */
3246                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3247                 if (calc > 0xffff)
3248                         calc = 0xffff;
3249                 logged |= 2;
3250                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3251                 if (rack_per_upper_bound_ca &&
3252                     (rack->rc_dragged_bottom == 0) &&
3253                     (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3254                         rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3255         }
3256         if (rack->rc_gp_saw_ss &&
3257             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3258                                   rack->r_ctl.rack_per_of_gp_ss)) {
3259                 /* In SS */
3260                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3261                 if (calc > 0xffff)
3262                         calc = 0xffff;
3263                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3264                 if (rack_per_upper_bound_ss &&
3265                     (rack->rc_dragged_bottom == 0) &&
3266                     (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3267                         rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3268                 logged |= 4;
3269         }
3270         if (logged &&
3271             (rack->rc_gp_incr == 0)){
3272                 /* Go into increment mode */
3273                 rack->rc_gp_incr = 1;
3274                 rack->rc_gp_timely_inc_cnt = 0;
3275         }
3276         if (rack->rc_gp_incr &&
3277             logged &&
3278             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3279                 rack->rc_gp_timely_inc_cnt++;
3280         }
3281         rack_log_timely(rack,  logged, plus, 0, 0,
3282                         __LINE__, 1);
3283 }
3284
3285 static uint32_t
3286 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3287 {
3288         /*
3289          * norm_grad = rtt_diff / minrtt;
3290          * new_per = curper * (1 - B * norm_grad)
3291          *
3292          * B = rack_gp_decrease_per (default 10%)
3293          * rtt_dif = input var current rtt-diff
3294          * curper = input var current percentage
3295          * minrtt = from rack filter
3296          *
3297          */
3298         uint64_t perf;
3299
3300         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3301                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3302                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3303                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3304                      (uint64_t)1000000)) /
3305                 (uint64_t)1000000);
3306         if (perf > curper) {
3307                 /* TSNH */
3308                 perf = curper - 1;
3309         }
3310         return ((uint32_t)perf);
3311 }
3312
3313 static uint32_t
3314 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3315 {
3316         /*
3317          *                                   highrttthresh
3318          * result = curper * (1 - (B * ( 1 -  ------          ))
3319          *                                     gp_srtt
3320          *
3321          * B = rack_gp_decrease_per (default 10%)
3322          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3323          */
3324         uint64_t perf;
3325         uint32_t highrttthresh;
3326
3327         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3328
3329         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3330                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3331                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3332                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3333         return (perf);
3334 }
3335
3336 static void
3337 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3338 {
3339         uint64_t logvar, logvar2, logvar3;
3340         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3341
3342         if (rack->rc_gp_incr) {
3343                 /* Turn off increment counting */
3344                 rack->rc_gp_incr = 0;
3345                 rack->rc_gp_timely_inc_cnt = 0;
3346         }
3347         ss_red = ca_red = rec_red = 0;
3348         logged = 0;
3349         /* Calculate the reduction value */
3350         if (rtt_diff < 0) {
3351                 rtt_diff *= -1;
3352         }
3353         /* Must be at least 1% reduction */
3354         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3355                 /* We have been in recovery ding it too */
3356                 if (timely_says == 2) {
3357                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3358                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3359                         if (alt < new_per)
3360                                 val = alt;
3361                         else
3362                                 val = new_per;
3363                 } else
3364                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3365                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3366                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3367                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3368                 } else {
3369                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3370                         rec_red = 0;
3371                 }
3372                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3373                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3374                 logged |= 1;
3375         }
3376         if (rack->rc_gp_saw_ss) {
3377                 /* Sent in SS */
3378                 if (timely_says == 2) {
3379                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3380                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3381                         if (alt < new_per)
3382                                 val = alt;
3383                         else
3384                                 val = new_per;
3385                 } else
3386                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3387                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3388                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3389                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3390                 } else {
3391                         ss_red = new_per;
3392                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3393                         logvar = new_per;
3394                         logvar <<= 32;
3395                         logvar |= alt;
3396                         logvar2 = (uint32_t)rtt;
3397                         logvar2 <<= 32;
3398                         logvar2 |= (uint32_t)rtt_diff;
3399                         logvar3 = rack_gp_rtt_maxmul;
3400                         logvar3 <<= 32;
3401                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3402                         rack_log_timely(rack, timely_says,
3403                                         logvar2, logvar3,
3404                                         logvar, __LINE__, 10);
3405                 }
3406                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3407                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3408                 logged |= 4;
3409         } else if (rack->rc_gp_saw_ca) {
3410                 /* Sent in CA */
3411                 if (timely_says == 2) {
3412                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3413                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3414                         if (alt < new_per)
3415                                 val = alt;
3416                         else
3417                                 val = new_per;
3418                 } else
3419                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3420                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3421                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3422                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3423                 } else {
3424                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3425                         ca_red = 0;
3426                         logvar = new_per;
3427                         logvar <<= 32;
3428                         logvar |= alt;
3429                         logvar2 = (uint32_t)rtt;
3430                         logvar2 <<= 32;
3431                         logvar2 |= (uint32_t)rtt_diff;
3432                         logvar3 = rack_gp_rtt_maxmul;
3433                         logvar3 <<= 32;
3434                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3435                         rack_log_timely(rack, timely_says,
3436                                         logvar2, logvar3,
3437                                         logvar, __LINE__, 10);
3438                 }
3439                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3440                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3441                 logged |= 2;
3442         }
3443         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3444                 rack->rc_gp_timely_dec_cnt++;
3445                 if (rack_timely_dec_clear &&
3446                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3447                         rack->rc_gp_timely_dec_cnt = 0;
3448         }
3449         logvar = ss_red;
3450         logvar <<= 32;
3451         logvar |= ca_red;
3452         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3453                         __LINE__, 2);
3454 }
3455
3456 static void
3457 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3458                      uint32_t rtt, uint32_t line, uint8_t reas)
3459 {
3460         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3461                 union tcp_log_stackspecific log;
3462                 struct timeval tv;
3463
3464                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3465                 log.u_bbr.flex1 = line;
3466                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3467                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3468                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3469                 log.u_bbr.flex5 = rtt;
3470                 log.u_bbr.flex6 = rack->rc_highly_buffered;
3471                 log.u_bbr.flex6 <<= 1;
3472                 log.u_bbr.flex6 |= rack->forced_ack;
3473                 log.u_bbr.flex6 <<= 1;
3474                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3475                 log.u_bbr.flex6 <<= 1;
3476                 log.u_bbr.flex6 |= rack->in_probe_rtt;
3477                 log.u_bbr.flex6 <<= 1;
3478                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3479                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3480                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3481                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3482                 log.u_bbr.flex8 = reas;
3483                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3484                 log.u_bbr.delRate = rack_get_bw(rack);
3485                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3486                 log.u_bbr.cur_del_rate <<= 32;
3487                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3488                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3489                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3490                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3491                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3492                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3493                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3494                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3495                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3496                 log.u_bbr.rttProp = us_cts;
3497                 log.u_bbr.rttProp <<= 32;
3498                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3499                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3500                     &rack->rc_inp->inp_socket->so_rcv,
3501                     &rack->rc_inp->inp_socket->so_snd,
3502                     BBR_LOG_RTT_SHRINKS, 0,
3503                     0, &log, false, &rack->r_ctl.act_rcv_time);
3504         }
3505 }
3506
3507 static void
3508 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3509 {
3510         uint64_t bwdp;
3511
3512         bwdp = rack_get_bw(rack);
3513         bwdp *= (uint64_t)rtt;
3514         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3515         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3516         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3517                 /*
3518                  * A window protocol must be able to have 4 packets
3519                  * outstanding as the floor in order to function
3520                  * (especially considering delayed ack :D).
3521                  */
3522                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3523         }
3524 }
3525
3526 static void
3527 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3528 {
3529         /**
3530          * ProbeRTT is a bit different in rack_pacing than in
3531          * BBR. It is like BBR in that it uses the lowering of
3532          * the RTT as a signal that we saw something new and
3533          * counts from there for how long between. But it is
3534          * different in that its quite simple. It does not
3535          * play with the cwnd and wait until we get down
3536          * to N segments outstanding and hold that for
3537          * 200ms. Instead it just sets the pacing reduction
3538          * rate to a set percentage (70 by default) and hold
3539          * that for a number of recent GP Srtt's.
3540          */
3541         uint32_t segsiz;
3542
3543         if (rack->rc_gp_dyn_mul == 0)
3544                 return;
3545
3546         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3547                 /* We are idle */
3548                 return;
3549         }
3550         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3551             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3552                 /*
3553                  * Stop the goodput now, the idea here is
3554                  * that future measurements with in_probe_rtt
3555                  * won't register if they are not greater so
3556                  * we want to get what info (if any) is available
3557                  * now.
3558                  */
3559                 rack_do_goodput_measurement(rack->rc_tp, rack,
3560                                             rack->rc_tp->snd_una, __LINE__,
3561                                             RACK_QUALITY_PROBERTT);
3562         }
3563         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3564         rack->r_ctl.rc_time_probertt_entered = us_cts;
3565         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3566                      rack->r_ctl.rc_pace_min_segs);
3567         rack->in_probe_rtt = 1;
3568         rack->measure_saw_probe_rtt = 1;
3569         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3570         rack->r_ctl.rc_time_probertt_starts = 0;
3571         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3572         if (rack_probertt_use_min_rtt_entry)
3573                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3574         else
3575                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3576         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3577                              __LINE__, RACK_RTTS_ENTERPROBE);
3578 }
3579
3580 static void
3581 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3582 {
3583         struct rack_sendmap *rsm;
3584         uint32_t segsiz;
3585
3586         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3587                      rack->r_ctl.rc_pace_min_segs);
3588         rack->in_probe_rtt = 0;
3589         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3590             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3591                 /*
3592                  * Stop the goodput now, the idea here is
3593                  * that future measurements with in_probe_rtt
3594                  * won't register if they are not greater so
3595                  * we want to get what info (if any) is available
3596                  * now.
3597                  */
3598                 rack_do_goodput_measurement(rack->rc_tp, rack,
3599                                             rack->rc_tp->snd_una, __LINE__,
3600                                             RACK_QUALITY_PROBERTT);
3601         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3602                 /*
3603                  * We don't have enough data to make a measurement.
3604                  * So lets just stop and start here after exiting
3605                  * probe-rtt. We probably are not interested in
3606                  * the results anyway.
3607                  */
3608                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3609         }
3610         /*
3611          * Measurements through the current snd_max are going
3612          * to be limited by the slower pacing rate.
3613          *
3614          * We need to mark these as app-limited so we
3615          * don't collapse the b/w.
3616          */
3617         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3618         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3619                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3620                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3621                 else {
3622                         /*
3623                          * Go out to the end app limited and mark
3624                          * this new one as next and move the end_appl up
3625                          * to this guy.
3626                          */
3627                         if (rack->r_ctl.rc_end_appl)
3628                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3629                         rack->r_ctl.rc_end_appl = rsm;
3630                 }
3631                 rsm->r_flags |= RACK_APP_LIMITED;
3632                 rack->r_ctl.rc_app_limited_cnt++;
3633         }
3634         /*
3635          * Now, we need to examine our pacing rate multipliers.
3636          * If its under 100%, we need to kick it back up to
3637          * 100%. We also don't let it be over our "max" above
3638          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3639          * Note setting clamp_atexit_prtt to 0 has the effect
3640          * of setting CA/SS to 100% always at exit (which is
3641          * the default behavior).
3642          */
3643         if (rack_probertt_clear_is) {
3644                 rack->rc_gp_incr = 0;
3645                 rack->rc_gp_bwred = 0;
3646                 rack->rc_gp_timely_inc_cnt = 0;
3647                 rack->rc_gp_timely_dec_cnt = 0;
3648         }
3649         /* Do we do any clamping at exit? */
3650         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3651                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3652                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3653         }
3654         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3655                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3656                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3657         }
3658         /*
3659          * Lets set rtt_diff to 0, so that we will get a "boost"
3660          * after exiting.
3661          */
3662         rack->r_ctl.rc_rtt_diff = 0;
3663
3664         /* Clear all flags so we start fresh */
3665         rack->rc_tp->t_bytes_acked = 0;
3666         rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
3667         /*
3668          * If configured to, set the cwnd and ssthresh to
3669          * our targets.
3670          */
3671         if (rack_probe_rtt_sets_cwnd) {
3672                 uint64_t ebdp;
3673                 uint32_t setto;
3674
3675                 /* Set ssthresh so we get into CA once we hit our target */
3676                 if (rack_probertt_use_min_rtt_exit == 1) {
3677                         /* Set to min rtt */
3678                         rack_set_prtt_target(rack, segsiz,
3679                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3680                 } else if (rack_probertt_use_min_rtt_exit == 2) {
3681                         /* Set to current gp rtt */
3682                         rack_set_prtt_target(rack, segsiz,
3683                                              rack->r_ctl.rc_gp_srtt);
3684                 } else if (rack_probertt_use_min_rtt_exit == 3) {
3685                         /* Set to entry gp rtt */
3686                         rack_set_prtt_target(rack, segsiz,
3687                                              rack->r_ctl.rc_entry_gp_rtt);
3688                 } else {
3689                         uint64_t sum;
3690                         uint32_t setval;
3691
3692                         sum = rack->r_ctl.rc_entry_gp_rtt;
3693                         sum *= 10;
3694                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3695                         if (sum >= 20) {
3696                                 /*
3697                                  * A highly buffered path needs
3698                                  * cwnd space for timely to work.
3699                                  * Lets set things up as if
3700                                  * we are heading back here again.
3701                                  */
3702                                 setval = rack->r_ctl.rc_entry_gp_rtt;
3703                         } else if (sum >= 15) {
3704                                 /*
3705                                  * Lets take the smaller of the
3706                                  * two since we are just somewhat
3707                                  * buffered.
3708                                  */
3709                                 setval = rack->r_ctl.rc_gp_srtt;
3710                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3711                                         setval = rack->r_ctl.rc_entry_gp_rtt;
3712                         } else {
3713                                 /*
3714                                  * Here we are not highly buffered
3715                                  * and should pick the min we can to
3716                                  * keep from causing loss.
3717                                  */
3718                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3719                         }
3720                         rack_set_prtt_target(rack, segsiz,
3721                                              setval);
3722                 }
3723                 if (rack_probe_rtt_sets_cwnd > 1) {
3724                         /* There is a percentage here to boost */
3725                         ebdp = rack->r_ctl.rc_target_probertt_flight;
3726                         ebdp *= rack_probe_rtt_sets_cwnd;
3727                         ebdp /= 100;
3728                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3729                 } else
3730                         setto = rack->r_ctl.rc_target_probertt_flight;
3731                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3732                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3733                         /* Enforce a min */
3734                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3735                 }
3736                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3737                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3738         }
3739         rack_log_rtt_shrinks(rack,  us_cts,
3740                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3741                              __LINE__, RACK_RTTS_EXITPROBE);
3742         /* Clear times last so log has all the info */
3743         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3744         rack->r_ctl.rc_time_probertt_entered = us_cts;
3745         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3746         rack->r_ctl.rc_time_of_last_probertt = us_cts;
3747 }
3748
3749 static void
3750 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3751 {
3752         /* Check in on probe-rtt */
3753         if (rack->rc_gp_filled == 0) {
3754                 /* We do not do p-rtt unless we have gp measurements */
3755                 return;
3756         }
3757         if (rack->in_probe_rtt) {
3758                 uint64_t no_overflow;
3759                 uint32_t endtime, must_stay;
3760
3761                 if (rack->r_ctl.rc_went_idle_time &&
3762                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3763                         /*
3764                          * We went idle during prtt, just exit now.
3765                          */
3766                         rack_exit_probertt(rack, us_cts);
3767                 } else if (rack_probe_rtt_safety_val &&
3768                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3769                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3770                         /*
3771                          * Probe RTT safety value triggered!
3772                          */
3773                         rack_log_rtt_shrinks(rack,  us_cts,
3774                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3775                                              __LINE__, RACK_RTTS_SAFETY);
3776                         rack_exit_probertt(rack, us_cts);
3777                 }
3778                 /* Calculate the max we will wait */
3779                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3780                 if (rack->rc_highly_buffered)
3781                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3782                 /* Calculate the min we must wait */
3783                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3784                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3785                     TSTMP_LT(us_cts, endtime)) {
3786                         uint32_t calc;
3787                         /* Do we lower more? */
3788 no_exit:
3789                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3790                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3791                         else
3792                                 calc = 0;
3793                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3794                         if (calc) {
3795                                 /* Maybe */
3796                                 calc *= rack_per_of_gp_probertt_reduce;
3797                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3798                                 /* Limit it too */
3799                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3800                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3801                         }
3802                         /* We must reach target or the time set */
3803                         return;
3804                 }
3805                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3806                         if ((TSTMP_LT(us_cts, must_stay) &&
3807                              rack->rc_highly_buffered) ||
3808                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3809                               rack->r_ctl.rc_target_probertt_flight)) {
3810                                 /* We are not past the must_stay time */
3811                                 goto no_exit;
3812                         }
3813                         rack_log_rtt_shrinks(rack,  us_cts,
3814                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3815                                              __LINE__, RACK_RTTS_REACHTARGET);
3816                         rack->r_ctl.rc_time_probertt_starts = us_cts;
3817                         if (rack->r_ctl.rc_time_probertt_starts == 0)
3818                                 rack->r_ctl.rc_time_probertt_starts = 1;
3819                         /* Restore back to our rate we want to pace at in prtt */
3820                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3821                 }
3822                 /*
3823                  * Setup our end time, some number of gp_srtts plus 200ms.
3824                  */
3825                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3826                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3827                 if (rack_probertt_gpsrtt_cnt_div)
3828                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3829                 else
3830                         endtime = 0;
3831                 endtime += rack_min_probertt_hold;
3832                 endtime += rack->r_ctl.rc_time_probertt_starts;
3833                 if (TSTMP_GEQ(us_cts,  endtime)) {
3834                         /* yes, exit probertt */
3835                         rack_exit_probertt(rack, us_cts);
3836                 }
3837
3838         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3839                 /* Go into probertt, its been too long since we went lower */
3840                 rack_enter_probertt(rack, us_cts);
3841         }
3842 }
3843
3844 static void
3845 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3846                        uint32_t rtt, int32_t rtt_diff)
3847 {
3848         uint64_t cur_bw, up_bnd, low_bnd, subfr;
3849         uint32_t losses;
3850
3851         if ((rack->rc_gp_dyn_mul == 0) ||
3852             (rack->use_fixed_rate) ||
3853             (rack->in_probe_rtt) ||
3854             (rack->rc_always_pace == 0)) {
3855                 /* No dynamic GP multiplier in play */
3856                 return;
3857         }
3858         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3859         cur_bw = rack_get_bw(rack);
3860         /* Calculate our up and down range */
3861         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3862         up_bnd /= 100;
3863         up_bnd += rack->r_ctl.last_gp_comp_bw;
3864
3865         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3866         subfr /= 100;
3867         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3868         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3869                 /*
3870                  * This is the case where our RTT is above
3871                  * the max target and we have been configured
3872                  * to just do timely no bonus up stuff in that case.
3873                  *
3874                  * There are two configurations, set to 1, and we
3875                  * just do timely if we are over our max. If its
3876                  * set above 1 then we slam the multipliers down
3877                  * to 100 and then decrement per timely.
3878                  */
3879                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3880                                 __LINE__, 3);
3881                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3882                         rack_validate_multipliers_at_or_below_100(rack);
3883                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3884         } else if ((last_bw_est < low_bnd) && !losses) {
3885                 /*
3886                  * We are decreasing this is a bit complicated this
3887                  * means we are loosing ground. This could be
3888                  * because another flow entered and we are competing
3889                  * for b/w with it. This will push the RTT up which
3890                  * makes timely unusable unless we want to get shoved
3891                  * into a corner and just be backed off (the age
3892                  * old problem with delay based CC).
3893                  *
3894                  * On the other hand if it was a route change we
3895                  * would like to stay somewhat contained and not
3896                  * blow out the buffers.
3897                  */
3898                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3899                                 __LINE__, 3);
3900                 rack->r_ctl.last_gp_comp_bw = cur_bw;
3901                 if (rack->rc_gp_bwred == 0) {
3902                         /* Go into reduction counting */
3903                         rack->rc_gp_bwred = 1;
3904                         rack->rc_gp_timely_dec_cnt = 0;
3905                 }
3906                 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3907                     (timely_says == 0)) {
3908                         /*
3909                          * Push another time with a faster pacing
3910                          * to try to gain back (we include override to
3911                          * get a full raise factor).
3912                          */
3913                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3914                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3915                             (timely_says == 0) ||
3916                             (rack_down_raise_thresh == 0)) {
3917                                 /*
3918                                  * Do an override up in b/w if we were
3919                                  * below the threshold or if the threshold
3920                                  * is zero we always do the raise.
3921                                  */
3922                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3923                         } else {
3924                                 /* Log it stays the same */
3925                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3926                                                 __LINE__, 11);
3927                         }
3928                         rack->rc_gp_timely_dec_cnt++;
3929                         /* We are not incrementing really no-count */
3930                         rack->rc_gp_incr = 0;
3931                         rack->rc_gp_timely_inc_cnt = 0;
3932                 } else {
3933                         /*
3934                          * Lets just use the RTT
3935                          * information and give up
3936                          * pushing.
3937                          */
3938                         goto use_timely;
3939                 }
3940         } else if ((timely_says != 2) &&
3941                     !losses &&
3942                     (last_bw_est > up_bnd)) {
3943                 /*
3944                  * We are increasing b/w lets keep going, updating
3945                  * our b/w and ignoring any timely input, unless
3946                  * of course we are at our max raise (if there is one).
3947                  */
3948
3949                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3950                                 __LINE__, 3);
3951                 rack->r_ctl.last_gp_comp_bw = cur_bw;
3952                 if (rack->rc_gp_saw_ss &&
3953                     rack_per_upper_bound_ss &&
3954                      (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3955                             /*
3956                              * In cases where we can't go higher
3957                              * we should just use timely.
3958                              */
3959                             goto use_timely;
3960                 }
3961                 if (rack->rc_gp_saw_ca &&
3962                     rack_per_upper_bound_ca &&
3963                     (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
3964                             /*
3965                              * In cases where we can't go higher
3966                              * we should just use timely.
3967                              */
3968                             goto use_timely;
3969                 }
3970                 rack->rc_gp_bwred = 0;
3971                 rack->rc_gp_timely_dec_cnt = 0;
3972                 /* You get a set number of pushes if timely is trying to reduce */
3973                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
3974                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3975                 } else {
3976                         /* Log it stays the same */
3977                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
3978                             __LINE__, 12);
3979                 }
3980                 return;
3981         } else {
3982                 /*
3983                  * We are staying between the lower and upper range bounds
3984                  * so use timely to decide.
3985                  */
3986                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3987                                 __LINE__, 3);
3988 use_timely:
3989                 if (timely_says) {
3990                         rack->rc_gp_incr = 0;
3991                         rack->rc_gp_timely_inc_cnt = 0;
3992                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
3993                             !losses &&
3994                             (last_bw_est < low_bnd)) {
3995                                 /* We are loosing ground */
3996                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3997                                 rack->rc_gp_timely_dec_cnt++;
3998                                 /* We are not incrementing really no-count */
3999                                 rack->rc_gp_incr = 0;
4000                                 rack->rc_gp_timely_inc_cnt = 0;
4001                         } else
4002                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4003                 } else {
4004                         rack->rc_gp_bwred = 0;
4005                         rack->rc_gp_timely_dec_cnt = 0;
4006                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4007                 }
4008         }
4009 }
4010
4011 static int32_t
4012 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4013 {
4014         int32_t timely_says;
4015         uint64_t log_mult, log_rtt_a_diff;
4016
4017         log_rtt_a_diff = rtt;
4018         log_rtt_a_diff <<= 32;
4019         log_rtt_a_diff |= (uint32_t)rtt_diff;
4020         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4021                     rack_gp_rtt_maxmul)) {
4022                 /* Reduce the b/w multiplier */
4023                 timely_says = 2;
4024                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4025                 log_mult <<= 32;
4026                 log_mult |= prev_rtt;
4027                 rack_log_timely(rack,  timely_says, log_mult,
4028                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4029                                 log_rtt_a_diff, __LINE__, 4);
4030         } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4031                            ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4032                             max(rack_gp_rtt_mindiv , 1)))) {
4033                 /* Increase the b/w multiplier */
4034                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4035                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4036                          max(rack_gp_rtt_mindiv , 1));
4037                 log_mult <<= 32;
4038                 log_mult |= prev_rtt;
4039                 timely_says = 0;
4040                 rack_log_timely(rack,  timely_says, log_mult ,
4041                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4042                                 log_rtt_a_diff, __LINE__, 5);
4043         } else {
4044                 /*
4045                  * Use a gradient to find it the timely gradient
4046                  * is:
4047                  * grad = rc_rtt_diff / min_rtt;
4048                  *
4049                  * anything below or equal to 0 will be
4050                  * a increase indication. Anything above
4051                  * zero is a decrease. Note we take care
4052                  * of the actual gradient calculation
4053                  * in the reduction (its not needed for
4054                  * increase).
4055                  */
4056                 log_mult = prev_rtt;
4057                 if (rtt_diff <= 0) {
4058                         /*
4059                          * Rttdiff is less than zero, increase the
4060                          * b/w multiplier (its 0 or negative)
4061                          */
4062                         timely_says = 0;
4063                         rack_log_timely(rack,  timely_says, log_mult,
4064                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4065                 } else {
4066                         /* Reduce the b/w multiplier */
4067                         timely_says = 1;
4068                         rack_log_timely(rack,  timely_says, log_mult,
4069                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4070                 }
4071         }
4072         return (timely_says);
4073 }
4074
4075 static void
4076 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4077                             tcp_seq th_ack, int line, uint8_t quality)
4078 {
4079         uint64_t tim, bytes_ps, ltim, stim, utim;
4080         uint32_t segsiz, bytes, reqbytes, us_cts;
4081         int32_t gput, new_rtt_diff, timely_says;
4082         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4083         int did_add = 0;
4084
4085         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4086         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4087         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4088                 tim = us_cts - tp->gput_ts;
4089         else
4090                 tim = 0;
4091         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4092                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4093         else
4094                 stim = 0;
4095         /*
4096          * Use the larger of the send time or ack time. This prevents us
4097          * from being influenced by ack artifacts to come up with too
4098          * high of measurement. Note that since we are spanning over many more
4099          * bytes in most of our measurements hopefully that is less likely to
4100          * occur.
4101          */
4102         if (tim > stim)
4103                 utim = max(tim, 1);
4104         else
4105                 utim = max(stim, 1);
4106         /* Lets get a msec time ltim too for the old stuff */
4107         ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4108         gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4109         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4110         if ((tim == 0) && (stim == 0)) {
4111                 /*
4112                  * Invalid measurement time, maybe
4113                  * all on one ack/one send?
4114                  */
4115                 bytes = 0;
4116                 bytes_ps = 0;
4117                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4118                                            0, 0, 0, 10, __LINE__, NULL, quality);
4119                 goto skip_measurement;
4120         }
4121         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4122                 /* We never made a us_rtt measurement? */
4123                 bytes = 0;
4124                 bytes_ps = 0;
4125                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4126                                            0, 0, 0, 10, __LINE__, NULL, quality);
4127                 goto skip_measurement;
4128         }
4129         /*
4130          * Calculate the maximum possible b/w this connection
4131          * could have. We base our calculation on the lowest
4132          * rtt we have seen during the measurement and the
4133          * largest rwnd the client has given us in that time. This
4134          * forms a BDP that is the maximum that we could ever
4135          * get to the client. Anything larger is not valid.
4136          *
4137          * I originally had code here that rejected measurements
4138          * where the time was less than 1/2 the latest us_rtt.
4139          * But after thinking on that I realized its wrong since
4140          * say you had a 150Mbps or even 1Gbps link, and you
4141          * were a long way away.. example I am in Europe (100ms rtt)
4142          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4143          * bytes my time would be 1.2ms, and yet my rtt would say
4144          * the measurement was invalid the time was < 50ms. The
4145          * same thing is true for 150Mb (8ms of time).
4146          *
4147          * A better way I realized is to look at what the maximum
4148          * the connection could possibly do. This is gated on
4149          * the lowest RTT we have seen and the highest rwnd.
4150          * We should in theory never exceed that, if we are
4151          * then something on the path is storing up packets
4152          * and then feeding them all at once to our endpoint
4153          * messing up our measurement.
4154          */
4155         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4156         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4157         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4158         if (SEQ_LT(th_ack, tp->gput_seq)) {
4159                 /* No measurement can be made */
4160                 bytes = 0;
4161                 bytes_ps = 0;
4162                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4163                                            0, 0, 0, 10, __LINE__, NULL, quality);
4164                 goto skip_measurement;
4165         } else
4166                 bytes = (th_ack - tp->gput_seq);
4167         bytes_ps = (uint64_t)bytes;
4168         /*
4169          * Don't measure a b/w for pacing unless we have gotten at least
4170          * an initial windows worth of data in this measurement interval.
4171          *
4172          * Small numbers of bytes get badly influenced by delayed ack and
4173          * other artifacts. Note we take the initial window or our
4174          * defined minimum GP (defaulting to 10 which hopefully is the
4175          * IW).
4176          */
4177         if (rack->rc_gp_filled == 0) {
4178                 /*
4179                  * The initial estimate is special. We
4180                  * have blasted out an IW worth of packets
4181                  * without a real valid ack ts results. We
4182                  * then setup the app_limited_needs_set flag,
4183                  * this should get the first ack in (probably 2
4184                  * MSS worth) to be recorded as the timestamp.
4185                  * We thus allow a smaller number of bytes i.e.
4186                  * IW - 2MSS.
4187                  */
4188                 reqbytes -= (2 * segsiz);
4189                 /* Also lets fill previous for our first measurement to be neutral */
4190                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4191         }
4192         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4193                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4194                                            rack->r_ctl.rc_app_limited_cnt,
4195                                            0, 0, 10, __LINE__, NULL, quality);
4196                 goto skip_measurement;
4197         }
4198         /*
4199          * We now need to calculate the Timely like status so
4200          * we can update (possibly) the b/w multipliers.
4201          */
4202         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4203         if (rack->rc_gp_filled == 0) {
4204                 /* No previous reading */
4205                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4206         } else {
4207                 if (rack->measure_saw_probe_rtt == 0) {
4208                         /*
4209                          * We don't want a probertt to be counted
4210                          * since it will be negative incorrectly. We
4211                          * expect to be reducing the RTT when we
4212                          * pace at a slower rate.
4213                          */
4214                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4215                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4216                 }
4217         }
4218         timely_says = rack_make_timely_judgement(rack,
4219                 rack->r_ctl.rc_gp_srtt,
4220                 rack->r_ctl.rc_rtt_diff,
4221                 rack->r_ctl.rc_prev_gp_srtt
4222                 );
4223         bytes_ps *= HPTS_USEC_IN_SEC;
4224         bytes_ps /= utim;
4225         if (bytes_ps > rack->r_ctl.last_max_bw) {
4226                 /*
4227                  * Something is on path playing
4228                  * since this b/w is not possible based
4229                  * on our BDP (highest rwnd and lowest rtt
4230                  * we saw in the measurement window).
4231                  *
4232                  * Another option here would be to
4233                  * instead skip the measurement.
4234                  */
4235                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4236                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4237                                            11, __LINE__, NULL, quality);
4238                 bytes_ps = rack->r_ctl.last_max_bw;
4239         }
4240         /* We store gp for b/w in bytes per second */
4241         if (rack->rc_gp_filled == 0) {
4242                 /* Initial measurement */
4243                 if (bytes_ps) {
4244                         rack->r_ctl.gp_bw = bytes_ps;
4245                         rack->rc_gp_filled = 1;
4246                         rack->r_ctl.num_measurements = 1;
4247                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4248                 } else {
4249                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4250                                                    rack->r_ctl.rc_app_limited_cnt,
4251                                                    0, 0, 10, __LINE__, NULL, quality);
4252                 }
4253                 if (tcp_in_hpts(rack->rc_inp) &&
4254                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4255                         /*
4256                          * Ok we can't trust the pacer in this case
4257                          * where we transition from un-paced to paced.
4258                          * Or for that matter when the burst mitigation
4259                          * was making a wild guess and got it wrong.
4260                          * Stop the pacer and clear up all the aggregate
4261                          * delays etc.
4262                          */
4263                         tcp_hpts_remove(rack->rc_inp);
4264                         rack->r_ctl.rc_hpts_flags = 0;
4265                         rack->r_ctl.rc_last_output_to = 0;
4266                 }
4267                 did_add = 2;
4268         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4269                 /* Still a small number run an average */
4270                 rack->r_ctl.gp_bw += bytes_ps;
4271                 addpart = rack->r_ctl.num_measurements;
4272                 rack->r_ctl.num_measurements++;
4273                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4274                         /* We have collected enough to move forward */
4275                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4276                 }
4277                 did_add = 3;
4278         } else {
4279                 /*
4280                  * We want to take 1/wma of the goodput and add in to 7/8th
4281                  * of the old value weighted by the srtt. So if your measurement
4282                  * period is say 2 SRTT's long you would get 1/4 as the
4283                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4284                  *
4285                  * But we must be careful not to take too much i.e. if the
4286                  * srtt is say 20ms and the measurement is taken over
4287                  * 400ms our weight would be 400/20 i.e. 20. On the
4288                  * other hand if we get a measurement over 1ms with a
4289                  * 10ms rtt we only want to take a much smaller portion.
4290                  */
4291                 if (rack->r_ctl.num_measurements < 0xff) {
4292                         rack->r_ctl.num_measurements++;
4293                 }
4294                 srtt = (uint64_t)tp->t_srtt;
4295                 if (srtt == 0) {
4296                         /*
4297                          * Strange why did t_srtt go back to zero?
4298                          */
4299                         if (rack->r_ctl.rc_rack_min_rtt)
4300                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4301                         else
4302                                 srtt = HPTS_USEC_IN_MSEC;
4303                 }
4304                 /*
4305                  * XXXrrs: Note for reviewers, in playing with
4306                  * dynamic pacing I discovered this GP calculation
4307                  * as done originally leads to some undesired results.
4308                  * Basically you can get longer measurements contributing
4309                  * too much to the WMA. Thus I changed it if you are doing
4310                  * dynamic adjustments to only do the aportioned adjustment
4311                  * if we have a very small (time wise) measurement. Longer
4312                  * measurements just get there weight (defaulting to 1/8)
4313                  * add to the WMA. We may want to think about changing
4314                  * this to always do that for both sides i.e. dynamic
4315                  * and non-dynamic... but considering lots of folks
4316                  * were playing with this I did not want to change the
4317                  * calculation per.se. without your thoughts.. Lawerence?
4318                  * Peter??
4319                  */
4320                 if (rack->rc_gp_dyn_mul == 0) {
4321                         subpart = rack->r_ctl.gp_bw * utim;
4322                         subpart /= (srtt * 8);
4323                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4324                                 /*
4325                                  * The b/w update takes no more
4326                                  * away then 1/2 our running total
4327                                  * so factor it in.
4328                                  */
4329                                 addpart = bytes_ps * utim;
4330                                 addpart /= (srtt * 8);
4331                         } else {
4332                                 /*
4333                                  * Don't allow a single measurement
4334                                  * to account for more than 1/2 of the
4335                                  * WMA. This could happen on a retransmission
4336                                  * where utim becomes huge compared to
4337                                  * srtt (multiple retransmissions when using
4338                                  * the sending rate which factors in all the
4339                                  * transmissions from the first one).
4340                                  */
4341                                 subpart = rack->r_ctl.gp_bw / 2;
4342                                 addpart = bytes_ps / 2;
4343                         }
4344                         resid_bw = rack->r_ctl.gp_bw - subpart;
4345                         rack->r_ctl.gp_bw = resid_bw + addpart;
4346                         did_add = 1;
4347                 } else {
4348                         if ((utim / srtt) <= 1) {
4349                                 /*
4350                                  * The b/w update was over a small period
4351                                  * of time. The idea here is to prevent a small
4352                                  * measurement time period from counting
4353                                  * too much. So we scale it based on the
4354                                  * time so it attributes less than 1/rack_wma_divisor
4355                                  * of its measurement.
4356                                  */
4357                                 subpart = rack->r_ctl.gp_bw * utim;
4358                                 subpart /= (srtt * rack_wma_divisor);
4359                                 addpart = bytes_ps * utim;
4360                                 addpart /= (srtt * rack_wma_divisor);
4361                         } else {
4362                                 /*
4363                                  * The scaled measurement was long
4364                                  * enough so lets just add in the
4365                                  * portion of the measurement i.e. 1/rack_wma_divisor
4366                                  */
4367                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4368                                 addpart = bytes_ps / rack_wma_divisor;
4369                         }
4370                         if ((rack->measure_saw_probe_rtt == 0) ||
4371                             (bytes_ps > rack->r_ctl.gp_bw)) {
4372                                 /*
4373                                  * For probe-rtt we only add it in
4374                                  * if its larger, all others we just
4375                                  * add in.
4376                                  */
4377                                 did_add = 1;
4378                                 resid_bw = rack->r_ctl.gp_bw - subpart;
4379                                 rack->r_ctl.gp_bw = resid_bw + addpart;
4380                         }
4381                 }
4382         }
4383         if ((rack->gp_ready == 0) &&
4384             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4385                 /* We have enough measurements now */
4386                 rack->gp_ready = 1;
4387                 rack_set_cc_pacing(rack);
4388                 if (rack->defer_options)
4389                         rack_apply_deferred_options(rack);
4390         }
4391         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4392                                    rack_get_bw(rack), 22, did_add, NULL, quality);
4393         /* We do not update any multipliers if we are in or have seen a probe-rtt */
4394         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4395                 rack_update_multiplier(rack, timely_says, bytes_ps,
4396                                        rack->r_ctl.rc_gp_srtt,
4397                                        rack->r_ctl.rc_rtt_diff);
4398         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4399                                    rack_get_bw(rack), 3, line, NULL, quality);
4400         /* reset the gp srtt and setup the new prev */
4401         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4402         /* Record the lost count for the next measurement */
4403         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4404         /*
4405          * We restart our diffs based on the gpsrtt in the
4406          * measurement window.
4407          */
4408         rack->rc_gp_rtt_set = 0;
4409         rack->rc_gp_saw_rec = 0;
4410         rack->rc_gp_saw_ca = 0;
4411         rack->rc_gp_saw_ss = 0;
4412         rack->rc_dragged_bottom = 0;
4413 skip_measurement:
4414
4415 #ifdef STATS
4416         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4417                                  gput);
4418         /*
4419          * XXXLAS: This is a temporary hack, and should be
4420          * chained off VOI_TCP_GPUT when stats(9) grows an
4421          * API to deal with chained VOIs.
4422          */
4423         if (tp->t_stats_gput_prev > 0)
4424                 stats_voi_update_abs_s32(tp->t_stats,
4425                                          VOI_TCP_GPUT_ND,
4426                                          ((gput - tp->t_stats_gput_prev) * 100) /
4427                                          tp->t_stats_gput_prev);
4428 #endif
4429         tp->t_flags &= ~TF_GPUTINPROG;
4430         tp->t_stats_gput_prev = gput;
4431         /*
4432          * Now are we app limited now and there is space from where we
4433          * were to where we want to go?
4434          *
4435          * We don't do the other case i.e. non-applimited here since
4436          * the next send will trigger us picking up the missing data.
4437          */
4438         if (rack->r_ctl.rc_first_appl &&
4439             TCPS_HAVEESTABLISHED(tp->t_state) &&
4440             rack->r_ctl.rc_app_limited_cnt &&
4441             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4442             ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4443              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4444                 /*
4445                  * Yep there is enough outstanding to make a measurement here.
4446                  */
4447                 struct rack_sendmap *rsm, fe;
4448
4449                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4450                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4451                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4452                 rack->app_limited_needs_set = 0;
4453                 tp->gput_seq = th_ack;
4454                 if (rack->in_probe_rtt)
4455                         rack->measure_saw_probe_rtt = 1;
4456                 else if ((rack->measure_saw_probe_rtt) &&
4457                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4458                         rack->measure_saw_probe_rtt = 0;
4459                 if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4460                         /* There is a full window to gain info from */
4461                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4462                 } else {
4463                         /* We can only measure up to the applimited point */
4464                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4465                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4466                                 /*
4467                                  * We don't have enough to make a measurement.
4468                                  */
4469                                 tp->t_flags &= ~TF_GPUTINPROG;
4470                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4471                                                            0, 0, 0, 6, __LINE__, NULL, quality);
4472                                 return;
4473                         }
4474                 }
4475                 if (tp->t_state >= TCPS_FIN_WAIT_1) {
4476                         /*
4477                          * We will get no more data into the SB
4478                          * this means we need to have the data available
4479                          * before we start a measurement.
4480                          */
4481                         if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4482                                 /* Nope not enough data. */
4483                                 return;
4484                         }
4485                 }
4486                 tp->t_flags |= TF_GPUTINPROG;
4487                 /*
4488                  * Now we need to find the timestamp of the send at tp->gput_seq
4489                  * for the send based measurement.
4490                  */
4491                 fe.r_start = tp->gput_seq;
4492                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4493                 if (rsm) {
4494                         /* Ok send-based limit is set */
4495                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4496                                 /*
4497                                  * Move back to include the earlier part
4498                                  * so our ack time lines up right (this may
4499                                  * make an overlapping measurement but thats
4500                                  * ok).
4501                                  */
4502                                 tp->gput_seq = rsm->r_start;
4503                         }
4504                         if (rsm->r_flags & RACK_ACKED)
4505                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4506                         else
4507                                 rack->app_limited_needs_set = 1;
4508                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4509                 } else {
4510                         /*
4511                          * If we don't find the rsm due to some
4512                          * send-limit set the current time, which
4513                          * basically disables the send-limit.
4514                          */
4515                         struct timeval tv;
4516
4517                         microuptime(&tv);
4518                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4519                 }
4520                 rack_log_pacing_delay_calc(rack,
4521                                            tp->gput_seq,
4522                                            tp->gput_ack,
4523                                            (uint64_t)rsm,
4524                                            tp->gput_ts,
4525                                            rack->r_ctl.rc_app_limited_cnt,
4526                                            9,
4527                                            __LINE__, NULL, quality);
4528         }
4529 }
4530
4531 /*
4532  * CC wrapper hook functions
4533  */
4534 static void
4535 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4536     uint16_t type, int32_t recovery)
4537 {
4538         uint32_t prior_cwnd, acked;
4539         struct tcp_log_buffer *lgb = NULL;
4540         uint8_t labc_to_use, quality;
4541
4542         INP_WLOCK_ASSERT(tptoinpcb(tp));
4543         tp->t_ccv.nsegs = nsegs;
4544         acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
4545         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4546                 uint32_t max;
4547
4548                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4549                 if (tp->t_ccv.bytes_this_ack > max) {
4550                         tp->t_ccv.bytes_this_ack = max;
4551                 }
4552         }
4553 #ifdef STATS
4554         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4555             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4556 #endif
4557         quality = RACK_QUALITY_NONE;
4558         if ((tp->t_flags & TF_GPUTINPROG) &&
4559             rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4560                 /* Measure the Goodput */
4561                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4562 #ifdef NETFLIX_PEAKRATE
4563                 if ((type == CC_ACK) &&
4564                     (tp->t_maxpeakrate)) {
4565                         /*
4566                          * We update t_peakrate_thr. This gives us roughly
4567                          * one update per round trip time. Note
4568                          * it will only be used if pace_always is off i.e
4569                          * we don't do this for paced flows.
4570                          */
4571                         rack_update_peakrate_thr(tp);
4572                 }
4573 #endif
4574         }
4575         /* Which way our we limited, if not cwnd limited no advance in CA */
4576         if (tp->snd_cwnd <= tp->snd_wnd)
4577                 tp->t_ccv.flags |= CCF_CWND_LIMITED;
4578         else
4579                 tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
4580         if (tp->snd_cwnd > tp->snd_ssthresh) {
4581                 tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
4582                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4583                 /* For the setting of a window past use the actual scwnd we are using */
4584                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4585                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4586                         tp->t_ccv.flags |= CCF_ABC_SENTAWND;
4587                 }
4588         } else {
4589                 tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4590                 tp->t_bytes_acked = 0;
4591         }
4592         prior_cwnd = tp->snd_cwnd;
4593         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4594             (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4595                 labc_to_use = rack->rc_labc;
4596         else
4597                 labc_to_use = rack_max_abc_post_recovery;
4598         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4599                 union tcp_log_stackspecific log;
4600                 struct timeval tv;
4601
4602                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4603                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4604                 log.u_bbr.flex1 = th_ack;
4605                 log.u_bbr.flex2 = tp->t_ccv.flags;
4606                 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4607                 log.u_bbr.flex4 = tp->t_ccv.nsegs;
4608                 log.u_bbr.flex5 = labc_to_use;
4609                 log.u_bbr.flex6 = prior_cwnd;
4610                 log.u_bbr.flex7 = V_tcp_do_newsack;
4611                 log.u_bbr.flex8 = 1;
4612                 lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4613                                      0, &log, false, NULL, NULL, 0, &tv);
4614         }
4615         if (CC_ALGO(tp)->ack_received != NULL) {
4616                 /* XXXLAS: Find a way to live without this */
4617                 tp->t_ccv.curack = th_ack;
4618                 tp->t_ccv.labc = labc_to_use;
4619                 tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
4620                 CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
4621         }
4622         if (lgb) {
4623                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4624         }
4625         if (rack->r_must_retran) {
4626                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4627                         /*
4628                          * We now are beyond the rxt point so lets disable
4629                          * the flag.
4630                          */
4631                         rack->r_ctl.rc_out_at_rto = 0;
4632                         rack->r_must_retran = 0;
4633                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4634                         /*
4635                          * Only decrement the rc_out_at_rto if the cwnd advances
4636                          * at least a whole segment. Otherwise next time the peer
4637                          * acks, we won't be able to send this generaly happens
4638                          * when we are in Congestion Avoidance.
4639                          */
4640                         if (acked <= rack->r_ctl.rc_out_at_rto){
4641                                 rack->r_ctl.rc_out_at_rto -= acked;
4642                         } else {
4643                                 rack->r_ctl.rc_out_at_rto = 0;
4644                         }
4645                 }
4646         }
4647 #ifdef STATS
4648         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4649 #endif
4650         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4651                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4652         }
4653 #ifdef NETFLIX_PEAKRATE
4654         /* we enforce max peak rate if it is set and we are not pacing */
4655         if ((rack->rc_always_pace == 0) &&
4656             tp->t_peakrate_thr &&
4657             (tp->snd_cwnd > tp->t_peakrate_thr)) {
4658                 tp->snd_cwnd = tp->t_peakrate_thr;
4659         }
4660 #endif
4661 }
4662
4663 static void
4664 tcp_rack_partialack(struct tcpcb *tp)
4665 {
4666         struct tcp_rack *rack;
4667
4668         rack = (struct tcp_rack *)tp->t_fb_ptr;
4669         INP_WLOCK_ASSERT(tptoinpcb(tp));
4670         /*
4671          * If we are doing PRR and have enough
4672          * room to send <or> we are pacing and prr
4673          * is disabled we will want to see if we
4674          * can send data (by setting r_wanted_output to
4675          * true).
4676          */
4677         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4678             rack->rack_no_prr)
4679                 rack->r_wanted_output = 1;
4680 }
4681
4682 static void
4683 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4684 {
4685         struct tcp_rack *rack;
4686         uint32_t orig_cwnd;
4687
4688         orig_cwnd = tp->snd_cwnd;
4689         INP_WLOCK_ASSERT(tptoinpcb(tp));
4690         rack = (struct tcp_rack *)tp->t_fb_ptr;
4691         /* only alert CC if we alerted when we entered */
4692         if (CC_ALGO(tp)->post_recovery != NULL) {
4693                 tp->t_ccv.curack = th_ack;
4694                 CC_ALGO(tp)->post_recovery(&tp->t_ccv);
4695                 if (tp->snd_cwnd < tp->snd_ssthresh) {
4696                         /*
4697                          * Rack has burst control and pacing
4698                          * so lets not set this any lower than
4699                          * snd_ssthresh per RFC-6582 (option 2).
4700                          */
4701                         tp->snd_cwnd = tp->snd_ssthresh;
4702                 }
4703         }
4704         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4705                 union tcp_log_stackspecific log;
4706                 struct timeval tv;
4707
4708                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4709                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4710                 log.u_bbr.flex1 = th_ack;
4711                 log.u_bbr.flex2 = tp->t_ccv.flags;
4712                 log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4713                 log.u_bbr.flex4 = tp->t_ccv.nsegs;
4714                 log.u_bbr.flex5 = V_tcp_abc_l_var;
4715                 log.u_bbr.flex6 = orig_cwnd;
4716                 log.u_bbr.flex7 = V_tcp_do_newsack;
4717                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4718                 log.u_bbr.flex8 = 2;
4719                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4720                                0, &log, false, NULL, NULL, 0, &tv);
4721         }
4722         if ((rack->rack_no_prr == 0) &&
4723             (rack->no_prr_addback == 0) &&
4724             (rack->r_ctl.rc_prr_sndcnt > 0)) {
4725                 /*
4726                  * Suck the next prr cnt back into cwnd, but
4727                  * only do that if we are not application limited.
4728                  */
4729                 if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
4730                         /*
4731                          * We are allowed to add back to the cwnd the amount we did
4732                          * not get out if:
4733                          * a) no_prr_addback is off.
4734                          * b) we are not app limited
4735                          * c) we are doing prr
4736                          * <and>
4737                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4738                          */
4739                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4740                                             rack->r_ctl.rc_prr_sndcnt);
4741                 }
4742                 rack->r_ctl.rc_prr_sndcnt = 0;
4743                 rack_log_to_prr(rack, 1, 0, __LINE__);
4744         }
4745         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
4746         tp->snd_recover = tp->snd_una;
4747         if (rack->r_ctl.dsack_persist) {
4748                 rack->r_ctl.dsack_persist--;
4749                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4750                         rack->r_ctl.num_dsack = 0;
4751                 }
4752                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4753         }
4754         EXIT_RECOVERY(tp->t_flags);
4755 }
4756
4757 static void
4758 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
4759 {
4760         struct tcp_rack *rack;
4761         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4762
4763         INP_WLOCK_ASSERT(tptoinpcb(tp));
4764 #ifdef STATS
4765         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4766 #endif
4767         if (IN_RECOVERY(tp->t_flags) == 0) {
4768                 in_rec_at_entry = 0;
4769                 ssthresh_enter = tp->snd_ssthresh;
4770                 cwnd_enter = tp->snd_cwnd;
4771         } else
4772                 in_rec_at_entry = 1;
4773         rack = (struct tcp_rack *)tp->t_fb_ptr;
4774         switch (type) {
4775         case CC_NDUPACK:
4776                 tp->t_flags &= ~TF_WASFRECOVERY;
4777                 tp->t_flags &= ~TF_WASCRECOVERY;
4778                 if (!IN_FASTRECOVERY(tp->t_flags)) {
4779                         rack->r_ctl.rc_prr_delivered = 0;
4780                         rack->r_ctl.rc_prr_out = 0;
4781                         if (rack->rack_no_prr == 0) {
4782                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4783                                 rack_log_to_prr(rack, 2, in_rec_at_entry, line);
4784                         }
4785                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4786                         tp->snd_recover = tp->snd_max;
4787                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4788                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4789                 }
4790                 break;
4791         case CC_ECN:
4792                 if (!IN_CONGRECOVERY(tp->t_flags) ||
4793                     /*
4794                      * Allow ECN reaction on ACK to CWR, if
4795                      * that data segment was also CE marked.
4796                      */
4797                     SEQ_GEQ(ack, tp->snd_recover)) {
4798                         EXIT_CONGRECOVERY(tp->t_flags);
4799                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4800                         tp->snd_recover = tp->snd_max + 1;
4801                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4802                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4803                 }
4804                 break;
4805         case CC_RTO:
4806                 tp->t_dupacks = 0;
4807                 tp->t_bytes_acked = 0;
4808                 EXIT_RECOVERY(tp->t_flags);
4809                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4810                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4811                 orig_cwnd = tp->snd_cwnd;
4812                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4813                 rack_log_to_prr(rack, 16, orig_cwnd, line);
4814                 if (tp->t_flags2 & TF2_ECN_PERMIT)
4815                         tp->t_flags2 |= TF2_ECN_SND_CWR;
4816                 break;
4817         case CC_RTO_ERR:
4818                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4819                 /* RTO was unnecessary, so reset everything. */
4820                 tp->snd_cwnd = tp->snd_cwnd_prev;
4821                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4822                 tp->snd_recover = tp->snd_recover_prev;
4823                 if (tp->t_flags & TF_WASFRECOVERY) {
4824                         ENTER_FASTRECOVERY(tp->t_flags);
4825                         tp->t_flags &= ~TF_WASFRECOVERY;
4826                 }
4827                 if (tp->t_flags & TF_WASCRECOVERY) {
4828                         ENTER_CONGRECOVERY(tp->t_flags);
4829                         tp->t_flags &= ~TF_WASCRECOVERY;
4830                 }
4831                 tp->snd_nxt = tp->snd_max;
4832                 tp->t_badrxtwin = 0;
4833                 break;
4834         }
4835         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4836             (type != CC_RTO)){
4837                 tp->t_ccv.curack = ack;
4838                 CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
4839         }
4840         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4841                 rack_log_to_prr(rack, 15, cwnd_enter, line);
4842                 rack->r_ctl.dsack_byte_cnt = 0;
4843                 rack->r_ctl.retran_during_recovery = 0;
4844                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4845                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4846                 rack->r_ent_rec_ns = 1;
4847         }
4848 }
4849
4850 static inline void
4851 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4852 {
4853         uint32_t i_cwnd;
4854
4855         INP_WLOCK_ASSERT(tptoinpcb(tp));
4856
4857 #ifdef NETFLIX_STATS
4858         KMOD_TCPSTAT_INC(tcps_idle_restarts);
4859         if (tp->t_state == TCPS_ESTABLISHED)
4860                 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4861 #endif
4862         if (CC_ALGO(tp)->after_idle != NULL)
4863                 CC_ALGO(tp)->after_idle(&tp->t_ccv);
4864
4865         if (tp->snd_cwnd == 1)
4866                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
4867         else
4868                 i_cwnd = rc_init_window(rack);
4869
4870         /*
4871          * Being idle is no different than the initial window. If the cc
4872          * clamps it down below the initial window raise it to the initial
4873          * window.
4874          */
4875         if (tp->snd_cwnd < i_cwnd) {
4876                 tp->snd_cwnd = i_cwnd;
4877         }
4878 }
4879
4880 /*
4881  * Indicate whether this ack should be delayed.  We can delay the ack if
4882  * following conditions are met:
4883  *      - There is no delayed ack timer in progress.
4884  *      - Our last ack wasn't a 0-sized window. We never want to delay
4885  *        the ack that opens up a 0-sized window.
4886  *      - LRO wasn't used for this segment. We make sure by checking that the
4887  *        segment size is not larger than the MSS.
4888  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4889  *        connection.
4890  */
4891 #define DELAY_ACK(tp, tlen)                      \
4892         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4893         ((tp->t_flags & TF_DELACK) == 0) &&      \
4894         (tlen <= tp->t_maxseg) &&                \
4895         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4896
4897 static struct rack_sendmap *
4898 rack_find_lowest_rsm(struct tcp_rack *rack)
4899 {
4900         struct rack_sendmap *rsm;
4901
4902         /*
4903          * Walk the time-order transmitted list looking for an rsm that is
4904          * not acked. This will be the one that was sent the longest time
4905          * ago that is still outstanding.
4906          */
4907         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4908                 if (rsm->r_flags & RACK_ACKED) {
4909                         continue;
4910                 }
4911                 goto finish;
4912         }
4913 finish:
4914         return (rsm);
4915 }
4916
4917 static struct rack_sendmap *
4918 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4919 {
4920         struct rack_sendmap *prsm;
4921
4922         /*
4923          * Walk the sequence order list backward until we hit and arrive at
4924          * the highest seq not acked. In theory when this is called it
4925          * should be the last segment (which it was not).
4926          */
4927         prsm = rsm;
4928         RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4929                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4930                         continue;
4931                 }
4932                 return (prsm);
4933         }
4934         return (NULL);
4935 }
4936
4937 static uint32_t
4938 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4939 {
4940         int32_t lro;
4941         uint32_t thresh;
4942
4943         /*
4944          * lro is the flag we use to determine if we have seen reordering.
4945          * If it gets set we have seen reordering. The reorder logic either
4946          * works in one of two ways:
4947          *
4948          * If reorder-fade is configured, then we track the last time we saw
4949          * re-ordering occur. If we reach the point where enough time as
4950          * passed we no longer consider reordering has occuring.
4951          *
4952          * Or if reorder-face is 0, then once we see reordering we consider
4953          * the connection to alway be subject to reordering and just set lro
4954          * to 1.
4955          *
4956          * In the end if lro is non-zero we add the extra time for
4957          * reordering in.
4958          */
4959         if (srtt == 0)
4960                 srtt = 1;
4961         if (rack->r_ctl.rc_reorder_ts) {
4962                 if (rack->r_ctl.rc_reorder_fade) {
4963                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
4964                                 lro = cts - rack->r_ctl.rc_reorder_ts;
4965                                 if (lro == 0) {
4966                                         /*
4967                                          * No time as passed since the last
4968                                          * reorder, mark it as reordering.
4969                                          */
4970                                         lro = 1;
4971                                 }
4972                         } else {
4973                                 /* Negative time? */
4974                                 lro = 0;
4975                         }
4976                         if (lro > rack->r_ctl.rc_reorder_fade) {
4977                                 /* Turn off reordering seen too */
4978                                 rack->r_ctl.rc_reorder_ts = 0;
4979                                 lro = 0;
4980                         }
4981                 } else {
4982                         /* Reodering does not fade */
4983                         lro = 1;
4984                 }
4985         } else {
4986                 lro = 0;
4987         }
4988         if (rack->rc_rack_tmr_std_based == 0) {
4989                 thresh = srtt + rack->r_ctl.rc_pkt_delay;
4990         } else {
4991                 /* Standards based pkt-delay is 1/4 srtt */
4992                 thresh = srtt +  (srtt >> 2);
4993         }
4994         if (lro && (rack->rc_rack_tmr_std_based == 0)) {
4995                 /* It must be set, if not you get 1/4 rtt */
4996                 if (rack->r_ctl.rc_reorder_shift)
4997                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
4998                 else
4999                         thresh += (srtt >> 2);
5000         }
5001         if (rack->rc_rack_use_dsack &&
5002             lro &&
5003             (rack->r_ctl.num_dsack > 0)) {
5004                 /*
5005                  * We only increase the reordering window if we
5006                  * have seen reordering <and> we have a DSACK count.
5007                  */
5008                 thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5009                 rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
5010         }
5011         /* SRTT * 2 is the ceiling */
5012         if (thresh > (srtt * 2)) {
5013                 thresh = srtt * 2;
5014         }
5015         /* And we don't want it above the RTO max either */
5016         if (thresh > rack_rto_max) {
5017                 thresh = rack_rto_max;
5018         }
5019         rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
5020         return (thresh);
5021 }
5022
5023 static uint32_t
5024 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5025                      struct rack_sendmap *rsm, uint32_t srtt)
5026 {
5027         struct rack_sendmap *prsm;
5028         uint32_t thresh, len;
5029         int segsiz;
5030
5031         if (srtt == 0)
5032                 srtt = 1;
5033         if (rack->r_ctl.rc_tlp_threshold)
5034                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5035         else
5036                 thresh = (srtt * 2);
5037
5038         /* Get the previous sent packet, if any */
5039         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5040         len = rsm->r_end - rsm->r_start;
5041         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5042                 /* Exactly like the ID */
5043                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5044                         uint32_t alt_thresh;
5045                         /*
5046                          * Compensate for delayed-ack with the d-ack time.
5047                          */
5048                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5049                         if (alt_thresh > thresh)
5050                                 thresh = alt_thresh;
5051                 }
5052         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5053                 /* 2.1 behavior */
5054                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5055                 if (prsm && (len <= segsiz)) {
5056                         /*
5057                          * Two packets outstanding, thresh should be (2*srtt) +
5058                          * possible inter-packet delay (if any).
5059                          */
5060                         uint32_t inter_gap = 0;
5061                         int idx, nidx;
5062
5063                         idx = rsm->r_rtr_cnt - 1;
5064                         nidx = prsm->r_rtr_cnt - 1;
5065                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5066                                 /* Yes it was sent later (or at the same time) */
5067                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5068                         }
5069                         thresh += inter_gap;
5070                 } else if (len <= segsiz) {
5071                         /*
5072                          * Possibly compensate for delayed-ack.
5073                          */
5074                         uint32_t alt_thresh;
5075
5076                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5077                         if (alt_thresh > thresh)
5078                                 thresh = alt_thresh;
5079                 }
5080         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5081                 /* 2.2 behavior */
5082                 if (len <= segsiz) {
5083                         uint32_t alt_thresh;
5084                         /*
5085                          * Compensate for delayed-ack with the d-ack time.
5086                          */
5087                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5088                         if (alt_thresh > thresh)
5089                                 thresh = alt_thresh;
5090                 }
5091         }
5092         /* Not above an RTO */
5093         if (thresh > tp->t_rxtcur) {
5094                 thresh = tp->t_rxtcur;
5095         }
5096         /* Not above a RTO max */
5097         if (thresh > rack_rto_max) {
5098                 thresh = rack_rto_max;
5099         }
5100         /* Apply user supplied min TLP */
5101         if (thresh < rack_tlp_min) {
5102                 thresh = rack_tlp_min;
5103         }
5104         return (thresh);
5105 }
5106
5107 static uint32_t
5108 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5109 {
5110         /*
5111          * We want the rack_rtt which is the
5112          * last rtt we measured. However if that
5113          * does not exist we fallback to the srtt (which
5114          * we probably will never do) and then as a last
5115          * resort we use RACK_INITIAL_RTO if no srtt is
5116          * yet set.
5117          */
5118         if (rack->rc_rack_rtt)
5119                 return (rack->rc_rack_rtt);
5120         else if (tp->t_srtt == 0)
5121                 return (RACK_INITIAL_RTO);
5122         return (tp->t_srtt);
5123 }
5124
5125 static struct rack_sendmap *
5126 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5127 {
5128         /*
5129          * Check to see that we don't need to fall into recovery. We will
5130          * need to do so if our oldest transmit is past the time we should
5131          * have had an ack.
5132          */
5133         struct tcp_rack *rack;
5134         struct rack_sendmap *rsm;
5135         int32_t idx;
5136         uint32_t srtt, thresh;
5137
5138         rack = (struct tcp_rack *)tp->t_fb_ptr;
5139         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5140                 return (NULL);
5141         }
5142         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5143         if (rsm == NULL)
5144                 return (NULL);
5145
5146
5147         if (rsm->r_flags & RACK_ACKED) {
5148                 rsm = rack_find_lowest_rsm(rack);
5149                 if (rsm == NULL)
5150                         return (NULL);
5151         }
5152         idx = rsm->r_rtr_cnt - 1;
5153         srtt = rack_grab_rtt(tp, rack);
5154         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5155         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5156                 return (NULL);
5157         }
5158         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5159                 return (NULL);
5160         }
5161         /* Ok if we reach here we are over-due and this guy can be sent */
5162         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
5163         return (rsm);
5164 }
5165
5166 static uint32_t
5167 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5168 {
5169         int32_t t;
5170         int32_t tt;
5171         uint32_t ret_val;
5172
5173         t = (tp->t_srtt + (tp->t_rttvar << 2));
5174         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5175             rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5176         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5177         ret_val = (uint32_t)tt;
5178         return (ret_val);
5179 }
5180
5181 static uint32_t
5182 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5183 {
5184         /*
5185          * Start the FR timer, we do this based on getting the first one in
5186          * the rc_tmap. Note that if its NULL we must stop the timer. in all
5187          * events we need to stop the running timer (if its running) before
5188          * starting the new one.
5189          */
5190         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5191         uint32_t srtt_cur;
5192         int32_t idx;
5193         int32_t is_tlp_timer = 0;
5194         struct rack_sendmap *rsm;
5195
5196         if (rack->t_timers_stopped) {
5197                 /* All timers have been stopped none are to run */
5198                 return (0);
5199         }
5200         if (rack->rc_in_persist) {
5201                 /* We can't start any timer in persists */
5202                 return (rack_get_persists_timer_val(tp, rack));
5203         }
5204         rack->rc_on_min_to = 0;
5205         if ((tp->t_state < TCPS_ESTABLISHED) ||
5206             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5207                 goto activate_rxt;
5208         }
5209         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5210         if ((rsm == NULL) || sup_rack) {
5211                 /* Nothing on the send map or no rack */
5212 activate_rxt:
5213                 time_since_sent = 0;
5214                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5215                 if (rsm) {
5216                         /*
5217                          * Should we discount the RTX timer any?
5218                          *
5219                          * We want to discount it the smallest amount.
5220                          * If a timer (Rack/TLP or RXT) has gone off more
5221                          * recently thats the discount we want to use (now - timer time).
5222                          * If the retransmit of the oldest packet was more recent then
5223                          * we want to use that (now - oldest-packet-last_transmit_time).
5224                          *
5225                          */
5226                         idx = rsm->r_rtr_cnt - 1;
5227                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5228                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5229                         else
5230                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5231                         if (TSTMP_GT(cts, tstmp_touse))
5232                             time_since_sent = cts - tstmp_touse;
5233                 }
5234                 if (SEQ_LT(tp->snd_una, tp->snd_max) ||
5235                     sbavail(&tptosocket(tp)->so_snd)) {
5236                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5237                         to = tp->t_rxtcur;
5238                         if (to > time_since_sent)
5239                                 to -= time_since_sent;
5240                         else
5241                                 to = rack->r_ctl.rc_min_to;
5242                         if (to == 0)
5243                                 to = 1;
5244                         /* Special case for KEEPINIT */
5245                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5246                             (TP_KEEPINIT(tp) != 0) &&
5247                             rsm) {
5248                                 /*
5249                                  * We have to put a ceiling on the rxt timer
5250                                  * of the keep-init timeout.
5251                                  */
5252                                 uint32_t max_time, red;
5253
5254                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5255                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5256                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5257                                         if (red < max_time)
5258                                                 max_time -= red;
5259                                         else
5260                                                 max_time = 1;
5261                                 }
5262                                 /* Reduce timeout to the keep value if needed */
5263                                 if (max_time < to)
5264                                         to = max_time;
5265                         }
5266                         return (to);
5267                 }
5268                 return (0);
5269         }
5270         if (rsm->r_flags & RACK_ACKED) {
5271                 rsm = rack_find_lowest_rsm(rack);
5272                 if (rsm == NULL) {
5273                         /* No lowest? */
5274                         goto activate_rxt;
5275                 }
5276         }
5277         if (rack->sack_attack_disable) {
5278                 /*
5279                  * We don't want to do
5280                  * any TLP's if you are an attacker.
5281                  * Though if you are doing what
5282                  * is expected you may still have
5283                  * SACK-PASSED marks.
5284                  */
5285                 goto activate_rxt;
5286         }
5287         /* Convert from ms to usecs */
5288         if ((rsm->r_flags & RACK_SACK_PASSED) ||
5289             (rsm->r_flags & RACK_RWND_COLLAPSED) ||
5290             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5291                 if ((tp->t_flags & TF_SENTFIN) &&
5292                     ((tp->snd_max - tp->snd_una) == 1) &&
5293                     (rsm->r_flags & RACK_HAS_FIN)) {
5294                         /*
5295                          * We don't start a rack timer if all we have is a
5296                          * FIN outstanding.
5297                          */
5298                         goto activate_rxt;
5299                 }
5300                 if ((rack->use_rack_rr == 0) &&
5301                     (IN_FASTRECOVERY(tp->t_flags)) &&
5302                     (rack->rack_no_prr == 0) &&
5303                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5304                         /*
5305                          * We are not cheating, in recovery  and
5306                          * not enough ack's to yet get our next
5307                          * retransmission out.
5308                          *
5309                          * Note that classified attackers do not
5310                          * get to use the rack-cheat.
5311                          */
5312                         goto activate_tlp;
5313                 }
5314                 srtt = rack_grab_rtt(tp, rack);
5315                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5316                 idx = rsm->r_rtr_cnt - 1;
5317                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5318                 if (SEQ_GEQ(exp, cts)) {
5319                         to = exp - cts;
5320                         if (to < rack->r_ctl.rc_min_to) {
5321                                 to = rack->r_ctl.rc_min_to;
5322                                 if (rack->r_rr_config == 3)
5323                                         rack->rc_on_min_to = 1;
5324                         }
5325                 } else {
5326                         to = rack->r_ctl.rc_min_to;
5327                         if (rack->r_rr_config == 3)
5328                                 rack->rc_on_min_to = 1;
5329                 }
5330         } else {
5331                 /* Ok we need to do a TLP not RACK */
5332 activate_tlp:
5333                 if ((rack->rc_tlp_in_progress != 0) &&
5334                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5335                         /*
5336                          * The previous send was a TLP and we have sent
5337                          * N TLP's without sending new data.
5338                          */
5339                         goto activate_rxt;
5340                 }
5341                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5342                 if (rsm == NULL) {
5343                         /* We found no rsm to TLP with. */
5344                         goto activate_rxt;
5345                 }
5346                 if (rsm->r_flags & RACK_HAS_FIN) {
5347                         /* If its a FIN we dont do TLP */
5348                         rsm = NULL;
5349                         goto activate_rxt;
5350                 }
5351                 idx = rsm->r_rtr_cnt - 1;
5352                 time_since_sent = 0;
5353                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5354                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5355                 else
5356                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5357                 if (TSTMP_GT(cts, tstmp_touse))
5358                     time_since_sent = cts - tstmp_touse;
5359                 is_tlp_timer = 1;
5360                 if (tp->t_srtt) {
5361                         if ((rack->rc_srtt_measure_made == 0) &&
5362                             (tp->t_srtt == 1)) {
5363                                 /*
5364                                  * If another stack as run and set srtt to 1,
5365                                  * then the srtt was 0, so lets use the initial.
5366                                  */
5367                                 srtt = RACK_INITIAL_RTO;
5368                         } else {
5369                                 srtt_cur = tp->t_srtt;
5370                                 srtt = srtt_cur;
5371                         }
5372                 } else
5373                         srtt = RACK_INITIAL_RTO;
5374                 /*
5375                  * If the SRTT is not keeping up and the
5376                  * rack RTT has spiked we want to use
5377                  * the last RTT not the smoothed one.
5378                  */
5379                 if (rack_tlp_use_greater &&
5380                     tp->t_srtt &&
5381                     (srtt < rack_grab_rtt(tp, rack))) {
5382                         srtt = rack_grab_rtt(tp, rack);
5383                 }
5384                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5385                 if (thresh > time_since_sent) {
5386                         to = thresh - time_since_sent;
5387                 } else {
5388                         to = rack->r_ctl.rc_min_to;
5389                         rack_log_alt_to_to_cancel(rack,
5390                                                   thresh,               /* flex1 */
5391                                                   time_since_sent,      /* flex2 */
5392                                                   tstmp_touse,          /* flex3 */
5393                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5394                                                   (uint32_t)rsm->r_tim_lastsent[idx],
5395                                                   srtt,
5396                                                   idx, 99);
5397                 }
5398                 if (to < rack_tlp_min) {
5399                         to = rack_tlp_min;
5400                 }
5401                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5402                         /*
5403                          * If the TLP time works out to larger than the max
5404                          * RTO lets not do TLP.. just RTO.
5405                          */
5406                         goto activate_rxt;
5407                 }
5408         }
5409         if (is_tlp_timer == 0) {
5410                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5411         } else {
5412                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5413         }
5414         if (to == 0)
5415                 to = 1;
5416         return (to);
5417 }
5418
5419 static void
5420 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5421 {
5422         if (rack->rc_in_persist == 0) {
5423                 if (tp->t_flags & TF_GPUTINPROG) {
5424                         /*
5425                          * Stop the goodput now, the calling of the
5426                          * measurement function clears the flag.
5427                          */
5428                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5429                                                     RACK_QUALITY_PERSIST);
5430                 }
5431 #ifdef NETFLIX_SHARED_CWND
5432                 if (rack->r_ctl.rc_scw) {
5433                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5434                         rack->rack_scwnd_is_idle = 1;
5435                 }
5436 #endif
5437                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5438                 if (rack->r_ctl.rc_went_idle_time == 0)
5439                         rack->r_ctl.rc_went_idle_time = 1;
5440                 rack_timer_cancel(tp, rack, cts, __LINE__);
5441                 rack->r_ctl.persist_lost_ends = 0;
5442                 rack->probe_not_answered = 0;
5443                 rack->forced_ack = 0;
5444                 tp->t_rxtshift = 0;
5445                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5446                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5447                 rack->rc_in_persist = 1;
5448         }
5449 }
5450
5451 static void
5452 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5453 {
5454         if (tcp_in_hpts(rack->rc_inp)) {
5455                 tcp_hpts_remove(rack->rc_inp);
5456                 rack->r_ctl.rc_hpts_flags = 0;
5457         }
5458 #ifdef NETFLIX_SHARED_CWND
5459         if (rack->r_ctl.rc_scw) {
5460                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5461                 rack->rack_scwnd_is_idle = 0;
5462         }
5463 #endif
5464         if (rack->rc_gp_dyn_mul &&
5465             (rack->use_fixed_rate == 0) &&
5466             (rack->rc_always_pace)) {
5467                 /*
5468                  * Do we count this as if a probe-rtt just
5469                  * finished?
5470                  */
5471                 uint32_t time_idle, idle_min;
5472
5473                 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5474                 idle_min = rack_min_probertt_hold;
5475                 if (rack_probertt_gpsrtt_cnt_div) {
5476                         uint64_t extra;
5477                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5478                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5479                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5480                         idle_min += (uint32_t)extra;
5481                 }
5482                 if (time_idle >= idle_min) {
5483                         /* Yes, we count it as a probe-rtt. */
5484                         uint32_t us_cts;
5485
5486                         us_cts = tcp_get_usecs(NULL);
5487                         if (rack->in_probe_rtt == 0) {
5488                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5489                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5490                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5491                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5492                         } else {
5493                                 rack_exit_probertt(rack, us_cts);
5494                         }
5495                 }
5496         }
5497         rack->rc_in_persist = 0;
5498         rack->r_ctl.rc_went_idle_time = 0;
5499         tp->t_rxtshift = 0;
5500         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5501            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5502         rack->r_ctl.rc_agg_delayed = 0;
5503         rack->r_early = 0;
5504         rack->r_late = 0;
5505         rack->r_ctl.rc_agg_early = 0;
5506 }
5507
5508 static void
5509 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5510                    struct hpts_diag *diag, struct timeval *tv)
5511 {
5512         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5513                 union tcp_log_stackspecific log;
5514
5515                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5516                 log.u_bbr.flex1 = diag->p_nxt_slot;
5517                 log.u_bbr.flex2 = diag->p_cur_slot;
5518                 log.u_bbr.flex3 = diag->slot_req;
5519                 log.u_bbr.flex4 = diag->inp_hptsslot;
5520                 log.u_bbr.flex5 = diag->slot_remaining;
5521                 log.u_bbr.flex6 = diag->need_new_to;
5522                 log.u_bbr.flex7 = diag->p_hpts_active;
5523                 log.u_bbr.flex8 = diag->p_on_min_sleep;
5524                 /* Hijack other fields as needed */
5525                 log.u_bbr.epoch = diag->have_slept;
5526                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5527                 log.u_bbr.pkts_out = diag->co_ret;
5528                 log.u_bbr.applimited = diag->hpts_sleep_time;
5529                 log.u_bbr.delivered = diag->p_prev_slot;
5530                 log.u_bbr.inflight = diag->p_runningslot;
5531                 log.u_bbr.bw_inuse = diag->wheel_slot;
5532                 log.u_bbr.rttProp = diag->wheel_cts;
5533                 log.u_bbr.timeStamp = cts;
5534                 log.u_bbr.delRate = diag->maxslots;
5535                 log.u_bbr.cur_del_rate = diag->p_curtick;
5536                 log.u_bbr.cur_del_rate <<= 32;
5537                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5538                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5539                     &rack->rc_inp->inp_socket->so_rcv,
5540                     &rack->rc_inp->inp_socket->so_snd,
5541                     BBR_LOG_HPTSDIAG, 0,
5542                     0, &log, false, tv);
5543         }
5544
5545 }
5546
5547 static void
5548 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5549 {
5550         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5551                 union tcp_log_stackspecific log;
5552                 struct timeval tv;
5553
5554                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5555                 log.u_bbr.flex1 = sb->sb_flags;
5556                 log.u_bbr.flex2 = len;
5557                 log.u_bbr.flex3 = sb->sb_state;
5558                 log.u_bbr.flex8 = type;
5559                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5560                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5561                     &rack->rc_inp->inp_socket->so_rcv,
5562                     &rack->rc_inp->inp_socket->so_snd,
5563                     TCP_LOG_SB_WAKE, 0,
5564                     len, &log, false, &tv);
5565         }
5566 }
5567
5568 static void
5569 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5570       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5571 {
5572         struct hpts_diag diag;
5573         struct inpcb *inp = tptoinpcb(tp);
5574         struct timeval tv;
5575         uint32_t delayed_ack = 0;
5576         uint32_t hpts_timeout;
5577         uint32_t entry_slot = slot;
5578         uint8_t stopped;
5579         uint32_t left = 0;
5580         uint32_t us_cts;
5581
5582         if ((tp->t_state == TCPS_CLOSED) ||
5583             (tp->t_state == TCPS_LISTEN)) {
5584                 return;
5585         }
5586         if (tcp_in_hpts(inp)) {
5587                 /* Already on the pacer */
5588                 return;
5589         }
5590         stopped = rack->rc_tmr_stopped;
5591         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5592                 left = rack->r_ctl.rc_timer_exp - cts;
5593         }
5594         rack->r_ctl.rc_timer_exp = 0;
5595         rack->r_ctl.rc_hpts_flags = 0;
5596         us_cts = tcp_get_usecs(&tv);
5597         /* Now early/late accounting */
5598         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5599         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5600                 /*
5601                  * We have a early carry over set,
5602                  * we can always add more time so we
5603                  * can always make this compensation.
5604                  *
5605                  * Note if ack's are allowed to wake us do not
5606                  * penalize the next timer for being awoke
5607                  * by an ack aka the rc_agg_early (non-paced mode).
5608                  */
5609                 slot += rack->r_ctl.rc_agg_early;
5610                 rack->r_early = 0;
5611                 rack->r_ctl.rc_agg_early = 0;
5612         }
5613         if (rack->r_late) {
5614                 /*
5615                  * This is harder, we can
5616                  * compensate some but it
5617                  * really depends on what
5618                  * the current pacing time is.
5619                  */
5620                 if (rack->r_ctl.rc_agg_delayed >= slot) {
5621                         /*
5622                          * We can't compensate for it all.
5623                          * And we have to have some time
5624                          * on the clock. We always have a min
5625                          * 10 slots (10 x 10 i.e. 100 usecs).
5626                          */
5627                         if (slot <= HPTS_TICKS_PER_SLOT) {
5628                                 /* We gain delay */
5629                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5630                                 slot = HPTS_TICKS_PER_SLOT;
5631                         } else {
5632                                 /* We take off some */
5633                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5634                                 slot = HPTS_TICKS_PER_SLOT;
5635                         }
5636                 } else {
5637                         slot -= rack->r_ctl.rc_agg_delayed;
5638                         rack->r_ctl.rc_agg_delayed = 0;
5639                         /* Make sure we have 100 useconds at minimum */
5640                         if (slot < HPTS_TICKS_PER_SLOT) {
5641                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5642                                 slot = HPTS_TICKS_PER_SLOT;
5643                         }
5644                         if (rack->r_ctl.rc_agg_delayed == 0)
5645                                 rack->r_late = 0;
5646                 }
5647         }
5648         if (slot) {
5649                 /* We are pacing too */
5650                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5651         }
5652         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5653 #ifdef NETFLIX_EXP_DETECTION
5654         if (rack->sack_attack_disable &&
5655             (slot < tcp_sad_pacing_interval)) {
5656                 /*
5657                  * We have a potential attacker on
5658                  * the line. We have possibly some
5659                  * (or now) pacing time set. We want to
5660                  * slow down the processing of sacks by some
5661                  * amount (if it is an attacker). Set the default
5662                  * slot for attackers in place (unless the original
5663                  * interval is longer). Its stored in
5664                  * micro-seconds, so lets convert to msecs.
5665                  */
5666                 slot = tcp_sad_pacing_interval;
5667         }
5668 #endif
5669         if (tp->t_flags & TF_DELACK) {
5670                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5671                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5672         }
5673         if (delayed_ack && ((hpts_timeout == 0) ||
5674                             (delayed_ack < hpts_timeout)))
5675                 hpts_timeout = delayed_ack;
5676         else
5677                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5678         /*
5679          * If no timers are going to run and we will fall off the hptsi
5680          * wheel, we resort to a keep-alive timer if its configured.
5681          */
5682         if ((hpts_timeout == 0) &&
5683             (slot == 0)) {
5684                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5685                     (tp->t_state <= TCPS_CLOSING)) {
5686                         /*
5687                          * Ok we have no timer (persists, rack, tlp, rxt  or
5688                          * del-ack), we don't have segments being paced. So
5689                          * all that is left is the keepalive timer.
5690                          */
5691                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5692                                 /* Get the established keep-alive time */
5693                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5694                         } else {
5695                                 /*
5696                                  * Get the initial setup keep-alive time,
5697                                  * note that this is probably not going to
5698                                  * happen, since rack will be running a rxt timer
5699                                  * if a SYN of some sort is outstanding. It is
5700                                  * actually handled in rack_timeout_rxt().
5701                                  */
5702                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5703                         }
5704                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5705                         if (rack->in_probe_rtt) {
5706                                 /*
5707                                  * We want to instead not wake up a long time from
5708                                  * now but to wake up about the time we would
5709                                  * exit probe-rtt and initiate a keep-alive ack.
5710                                  * This will get us out of probe-rtt and update
5711                                  * our min-rtt.
5712                                  */
5713                                 hpts_timeout = rack_min_probertt_hold;
5714                         }
5715                 }
5716         }
5717         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5718             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5719                 /*
5720                  * RACK, TLP, persists and RXT timers all are restartable
5721                  * based on actions input .. i.e we received a packet (ack
5722                  * or sack) and that changes things (rw, or snd_una etc).
5723                  * Thus we can restart them with a new value. For
5724                  * keep-alive, delayed_ack we keep track of what was left
5725                  * and restart the timer with a smaller value.
5726                  */
5727                 if (left < hpts_timeout)
5728                         hpts_timeout = left;
5729         }
5730         if (hpts_timeout) {
5731                 /*
5732                  * Hack alert for now we can't time-out over 2,147,483
5733                  * seconds (a bit more than 596 hours), which is probably ok
5734                  * :).
5735                  */
5736                 if (hpts_timeout > 0x7ffffffe)
5737                         hpts_timeout = 0x7ffffffe;
5738                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5739         }
5740         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5741         if ((rack->gp_ready == 0) &&
5742             (rack->use_fixed_rate == 0) &&
5743             (hpts_timeout < slot) &&
5744             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5745                 /*
5746                  * We have no good estimate yet for the
5747                  * old clunky burst mitigation or the
5748                  * real pacing. And the tlp or rxt is smaller
5749                  * than the pacing calculation. Lets not
5750                  * pace that long since we know the calculation
5751                  * so far is not accurate.
5752                  */
5753                 slot = hpts_timeout;
5754         }
5755         /**
5756          * Turn off all the flags for queuing by default. The
5757          * flags have important meanings to what happens when
5758          * LRO interacts with the transport. Most likely (by default now)
5759          * mbuf_queueing and ack compression are on. So the transport
5760          * has a couple of flags that control what happens (if those
5761          * are not on then these flags won't have any effect since it
5762          * won't go through the queuing LRO path).
5763          *
5764          * INP_MBUF_QUEUE_READY - This flags says that I am busy
5765          *                        pacing output, so don't disturb. But
5766          *                        it also means LRO can wake me if there
5767          *                        is a SACK arrival.
5768          *
5769          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5770          *                       with the above flag (QUEUE_READY) and
5771          *                       when present it says don't even wake me
5772          *                       if a SACK arrives.
5773          *
5774          * The idea behind these flags is that if we are pacing we
5775          * set the MBUF_QUEUE_READY and only get woken up if
5776          * a SACK arrives (which could change things) or if
5777          * our pacing timer expires. If, however, we have a rack
5778          * timer running, then we don't even want a sack to wake
5779          * us since the rack timer has to expire before we can send.
5780          *
5781          * Other cases should usually have none of the flags set
5782          * so LRO can call into us.
5783          */
5784         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5785         if (slot) {
5786                 rack->r_ctl.rc_last_output_to = us_cts + slot;
5787                 /*
5788                  * A pacing timer (slot) is being set, in
5789                  * such a case we cannot send (we are blocked by
5790                  * the timer). So lets tell LRO that it should not
5791                  * wake us unless there is a SACK. Note this only
5792                  * will be effective if mbuf queueing is on or
5793                  * compressed acks are being processed.
5794                  */
5795                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5796                 /*
5797                  * But wait if we have a Rack timer running
5798                  * even a SACK should not disturb us (with
5799                  * the exception of r_rr_config 3).
5800                  */
5801                 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5802                     (rack->r_rr_config != 3))
5803                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5804                 if (rack->rc_ack_can_sendout_data) {
5805                         /*
5806                          * Ahh but wait, this is that special case
5807                          * where the pacing timer can be disturbed
5808                          * backout the changes (used for non-paced
5809                          * burst limiting).
5810                          */
5811                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5812                 }
5813                 if ((rack->use_rack_rr) &&
5814                     (rack->r_rr_config < 2) &&
5815                     ((hpts_timeout) && (hpts_timeout < slot))) {
5816                         /*
5817                          * Arrange for the hpts to kick back in after the
5818                          * t-o if the t-o does not cause a send.
5819                          */
5820                         (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5821                                                    __LINE__, &diag);
5822                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5823                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5824                 } else {
5825                         (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot),
5826                                                    __LINE__, &diag);
5827                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5828                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5829                 }
5830         } else if (hpts_timeout) {
5831                 /*
5832                  * With respect to inp_flags2 here, lets let any new acks wake
5833                  * us up here. Since we are not pacing (no pacing timer), output
5834                  * can happen so we should let it. If its a Rack timer, then any inbound
5835                  * packet probably won't change the sending (we will be blocked)
5836                  * but it may change the prr stats so letting it in (the set defaults
5837                  * at the start of this block) are good enough.
5838                  */
5839                 (void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5840                                            __LINE__, &diag);
5841                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5842                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5843         } else {
5844                 /* No timer starting */
5845 #ifdef INVARIANTS
5846                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5847                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5848                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5849                 }
5850 #endif
5851         }
5852         rack->rc_tmr_stopped = 0;
5853         if (slot)
5854                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5855 }
5856
5857 /*
5858  * RACK Timer, here we simply do logging and house keeping.
5859  * the normal rack_output() function will call the
5860  * appropriate thing to check if we need to do a RACK retransmit.
5861  * We return 1, saying don't proceed with rack_output only
5862  * when all timers have been stopped (destroyed PCB?).
5863  */
5864 static int
5865 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5866 {
5867         /*
5868          * This timer simply provides an internal trigger to send out data.
5869          * The check_recovery_mode call will see if there are needed
5870          * retransmissions, if so we will enter fast-recovery. The output
5871          * call may or may not do the same thing depending on sysctl
5872          * settings.
5873          */
5874         struct rack_sendmap *rsm;
5875
5876         counter_u64_add(rack_to_tot, 1);
5877         if (rack->r_state && (rack->r_state != tp->t_state))
5878                 rack_set_state(tp, rack);
5879         rack->rc_on_min_to = 0;
5880         rsm = rack_check_recovery_mode(tp, cts);
5881         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5882         if (rsm) {
5883                 rack->r_ctl.rc_resend = rsm;
5884                 rack->r_timer_override = 1;
5885                 if (rack->use_rack_rr) {
5886                         /*
5887                          * Don't accumulate extra pacing delay
5888                          * we are allowing the rack timer to
5889                          * over-ride pacing i.e. rrr takes precedence
5890                          * if the pacing interval is longer than the rrr
5891                          * time (in other words we get the min pacing
5892                          * time versus rrr pacing time).
5893                          */
5894                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5895                 }
5896         }
5897         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5898         if (rsm == NULL) {
5899                 /* restart a timer and return 1 */
5900                 rack_start_hpts_timer(rack, tp, cts,
5901                                       0, 0, 0);
5902                 return (1);
5903         }
5904         return (0);
5905 }
5906
5907 static void
5908 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5909 {
5910         if (rsm->m->m_len > rsm->orig_m_len) {
5911                 /*
5912                  * Mbuf grew, caused by sbcompress, our offset does
5913                  * not change.
5914                  */
5915                 rsm->orig_m_len = rsm->m->m_len;
5916         } else if (rsm->m->m_len < rsm->orig_m_len) {
5917                 /*
5918                  * Mbuf shrank, trimmed off the top by an ack, our
5919                  * offset changes.
5920                  */
5921                 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
5922                 rsm->orig_m_len = rsm->m->m_len;
5923         }
5924 }
5925
5926 static void
5927 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
5928 {
5929         struct mbuf *m;
5930         uint32_t soff;
5931
5932         if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
5933                 /* Fix up the orig_m_len and possibly the mbuf offset */
5934                 rack_adjust_orig_mlen(src_rsm);
5935         }
5936         m = src_rsm->m;
5937         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
5938         while (soff >= m->m_len) {
5939                 /* Move out past this mbuf */
5940                 soff -= m->m_len;
5941                 m = m->m_next;
5942                 KASSERT((m != NULL),
5943                         ("rsm:%p nrsm:%p hit at soff:%u null m",
5944                          src_rsm, rsm, soff));
5945         }
5946         rsm->m = m;
5947         rsm->soff = soff;
5948         rsm->orig_m_len = m->m_len;
5949 }
5950
5951 static __inline void
5952 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5953                struct rack_sendmap *rsm, uint32_t start)
5954 {
5955         int idx;
5956
5957         nrsm->r_start = start;
5958         nrsm->r_end = rsm->r_end;
5959         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
5960         nrsm->r_flags = rsm->r_flags;
5961         nrsm->r_dupack = rsm->r_dupack;
5962         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
5963         nrsm->r_rtr_bytes = 0;
5964         nrsm->r_fas = rsm->r_fas;
5965         rsm->r_end = nrsm->r_start;
5966         nrsm->r_just_ret = rsm->r_just_ret;
5967         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
5968                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
5969         }
5970         /* Now if we have SYN flag we keep it on the left edge */
5971         if (nrsm->r_flags & RACK_HAS_SYN)
5972                 nrsm->r_flags &= ~RACK_HAS_SYN;
5973         /* Now if we have a FIN flag we keep it on the right edge */
5974         if (rsm->r_flags & RACK_HAS_FIN)
5975                 rsm->r_flags &= ~RACK_HAS_FIN;
5976         /* Push bit must go to the right edge as well */
5977         if (rsm->r_flags & RACK_HAD_PUSH)
5978                 rsm->r_flags &= ~RACK_HAD_PUSH;
5979         /* Clone over the state of the hw_tls flag */
5980         nrsm->r_hw_tls = rsm->r_hw_tls;
5981         /*
5982          * Now we need to find nrsm's new location in the mbuf chain
5983          * we basically calculate a new offset, which is soff +
5984          * how much is left in original rsm. Then we walk out the mbuf
5985          * chain to find the righ position, it may be the same mbuf
5986          * or maybe not.
5987          */
5988         KASSERT(((rsm->m != NULL) ||
5989                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
5990                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
5991         if (rsm->m)
5992                 rack_setup_offset_for_rsm(rsm, nrsm);
5993 }
5994
5995 static struct rack_sendmap *
5996 rack_merge_rsm(struct tcp_rack *rack,
5997                struct rack_sendmap *l_rsm,
5998                struct rack_sendmap *r_rsm)
5999 {
6000         /*
6001          * We are merging two ack'd RSM's,
6002          * the l_rsm is on the left (lower seq
6003          * values) and the r_rsm is on the right
6004          * (higher seq value). The simplest way
6005          * to merge these is to move the right
6006          * one into the left. I don't think there
6007          * is any reason we need to try to find
6008          * the oldest (or last oldest retransmitted).
6009          */
6010 #ifdef INVARIANTS
6011         struct rack_sendmap *rm;
6012 #endif
6013         rack_log_map_chg(rack->rc_tp, rack, NULL,
6014                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6015         l_rsm->r_end = r_rsm->r_end;
6016         if (l_rsm->r_dupack < r_rsm->r_dupack)
6017                 l_rsm->r_dupack = r_rsm->r_dupack;
6018         if (r_rsm->r_rtr_bytes)
6019                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6020         if (r_rsm->r_in_tmap) {
6021                 /* This really should not happen */
6022                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6023                 r_rsm->r_in_tmap = 0;
6024         }
6025
6026         /* Now the flags */
6027         if (r_rsm->r_flags & RACK_HAS_FIN)
6028                 l_rsm->r_flags |= RACK_HAS_FIN;
6029         if (r_rsm->r_flags & RACK_TLP)
6030                 l_rsm->r_flags |= RACK_TLP;
6031         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6032                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6033         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6034             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6035                 /*
6036                  * If both are app-limited then let the
6037                  * free lower the count. If right is app
6038                  * limited and left is not, transfer.
6039                  */
6040                 l_rsm->r_flags |= RACK_APP_LIMITED;
6041                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
6042                 if (r_rsm == rack->r_ctl.rc_first_appl)
6043                         rack->r_ctl.rc_first_appl = l_rsm;
6044         }
6045 #ifndef INVARIANTS
6046         (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6047 #else
6048         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6049         if (rm != r_rsm) {
6050                 panic("removing head in rack:%p rsm:%p rm:%p",
6051                       rack, r_rsm, rm);
6052         }
6053 #endif
6054         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6055                 /* Transfer the split limit to the map we free */
6056                 r_rsm->r_limit_type = l_rsm->r_limit_type;
6057                 l_rsm->r_limit_type = 0;
6058         }
6059         rack_free(rack, r_rsm);
6060         return (l_rsm);
6061 }
6062
6063 /*
6064  * TLP Timer, here we simply setup what segment we want to
6065  * have the TLP expire on, the normal rack_output() will then
6066  * send it out.
6067  *
6068  * We return 1, saying don't proceed with rack_output only
6069  * when all timers have been stopped (destroyed PCB?).
6070  */
6071 static int
6072 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6073 {
6074         /*
6075          * Tail Loss Probe.
6076          */
6077         struct rack_sendmap *rsm = NULL;
6078 #ifdef INVARIANTS
6079         struct rack_sendmap *insret;
6080 #endif
6081         struct socket *so = tptosocket(tp);
6082         uint32_t amm;
6083         uint32_t out, avail;
6084         int collapsed_win = 0;
6085
6086         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6087                 /* Its not time yet */
6088                 return (0);
6089         }
6090         if (ctf_progress_timeout_check(tp, true)) {
6091                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6092                 return (-ETIMEDOUT);    /* tcp_drop() */
6093         }
6094         /*
6095          * A TLP timer has expired. We have been idle for 2 rtts. So we now
6096          * need to figure out how to force a full MSS segment out.
6097          */
6098         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6099         rack->r_ctl.retran_during_recovery = 0;
6100         rack->r_ctl.dsack_byte_cnt = 0;
6101         counter_u64_add(rack_tlp_tot, 1);
6102         if (rack->r_state && (rack->r_state != tp->t_state))
6103                 rack_set_state(tp, rack);
6104         avail = sbavail(&so->so_snd);
6105         out = tp->snd_max - tp->snd_una;
6106         if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
6107                 /* special case, we need a retransmission */
6108                 collapsed_win = 1;
6109                 goto need_retran;
6110         }
6111         if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6112                 rack->r_ctl.dsack_persist--;
6113                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6114                         rack->r_ctl.num_dsack = 0;
6115                 }
6116                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6117         }
6118         if ((tp->t_flags & TF_GPUTINPROG) &&
6119             (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6120                 /*
6121                  * If this is the second in a row
6122                  * TLP and we are doing a measurement
6123                  * its time to abandon the measurement.
6124                  * Something is likely broken on
6125                  * the clients network and measuring a
6126                  * broken network does us no good.
6127                  */
6128                 tp->t_flags &= ~TF_GPUTINPROG;
6129                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6130                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6131                                            tp->gput_seq,
6132                                            0, 0, 18, __LINE__, NULL, 0);
6133         }
6134         /*
6135          * Check our send oldest always settings, and if
6136          * there is an oldest to send jump to the need_retran.
6137          */
6138         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6139                 goto need_retran;
6140
6141         if (avail > out) {
6142                 /* New data is available */
6143                 amm = avail - out;
6144                 if (amm > ctf_fixed_maxseg(tp)) {
6145                         amm = ctf_fixed_maxseg(tp);
6146                         if ((amm + out) > tp->snd_wnd) {
6147                                 /* We are rwnd limited */
6148                                 goto need_retran;
6149                         }
6150                 } else if (amm < ctf_fixed_maxseg(tp)) {
6151                         /* not enough to fill a MTU */
6152                         goto need_retran;
6153                 }
6154                 if (IN_FASTRECOVERY(tp->t_flags)) {
6155                         /* Unlikely */
6156                         if (rack->rack_no_prr == 0) {
6157                                 if (out + amm <= tp->snd_wnd) {
6158                                         rack->r_ctl.rc_prr_sndcnt = amm;
6159                                         rack->r_ctl.rc_tlp_new_data = amm;
6160                                         rack_log_to_prr(rack, 4, 0, __LINE__);
6161                                 }
6162                         } else
6163                                 goto need_retran;
6164                 } else {
6165                         /* Set the send-new override */
6166                         if (out + amm <= tp->snd_wnd)
6167                                 rack->r_ctl.rc_tlp_new_data = amm;
6168                         else
6169                                 goto need_retran;
6170                 }
6171                 rack->r_ctl.rc_tlpsend = NULL;
6172                 counter_u64_add(rack_tlp_newdata, 1);
6173                 goto send;
6174         }
6175 need_retran:
6176         /*
6177          * Ok we need to arrange the last un-acked segment to be re-sent, or
6178          * optionally the first un-acked segment.
6179          */
6180         if (collapsed_win == 0) {
6181                 if (rack_always_send_oldest)
6182                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6183                 else {
6184                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6185                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6186                                 rsm = rack_find_high_nonack(rack, rsm);
6187                         }
6188                 }
6189                 if (rsm == NULL) {
6190 #ifdef TCP_BLACKBOX
6191                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6192 #endif
6193                         goto out;
6194                 }
6195         } else {
6196                 /*
6197                  * We must find the last segment
6198                  * that was acceptable by the client.
6199                  */
6200                 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6201                         if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6202                                 /* Found one */
6203                                 break;
6204                         }
6205                 }
6206                 if (rsm == NULL) {
6207                         /* None? if so send the first */
6208                         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6209                         if (rsm == NULL) {
6210 #ifdef TCP_BLACKBOX
6211                                 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6212 #endif
6213                                 goto out;
6214                         }
6215                 }
6216         }
6217         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6218                 /*
6219                  * We need to split this the last segment in two.
6220                  */
6221                 struct rack_sendmap *nrsm;
6222
6223                 nrsm = rack_alloc_full_limit(rack);
6224                 if (nrsm == NULL) {
6225                         /*
6226                          * No memory to split, we will just exit and punt
6227                          * off to the RXT timer.
6228                          */
6229                         goto out;
6230                 }
6231                 rack_clone_rsm(rack, nrsm, rsm,
6232                                (rsm->r_end - ctf_fixed_maxseg(tp)));
6233                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6234 #ifndef INVARIANTS
6235                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6236 #else
6237                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6238                 if (insret != NULL) {
6239                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6240                               nrsm, insret, rack, rsm);
6241                 }
6242 #endif
6243                 if (rsm->r_in_tmap) {
6244                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6245                         nrsm->r_in_tmap = 1;
6246                 }
6247                 rsm = nrsm;
6248         }
6249         rack->r_ctl.rc_tlpsend = rsm;
6250 send:
6251         /* Make sure output path knows we are doing a TLP */
6252         *doing_tlp = 1;
6253         rack->r_timer_override = 1;
6254         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6255         return (0);
6256 out:
6257         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6258         return (0);
6259 }
6260
6261 /*
6262  * Delayed ack Timer, here we simply need to setup the
6263  * ACK_NOW flag and remove the DELACK flag. From there
6264  * the output routine will send the ack out.
6265  *
6266  * We only return 1, saying don't proceed, if all timers
6267  * are stopped (destroyed PCB?).
6268  */
6269 static int
6270 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6271 {
6272
6273         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6274         tp->t_flags &= ~TF_DELACK;
6275         tp->t_flags |= TF_ACKNOW;
6276         KMOD_TCPSTAT_INC(tcps_delack);
6277         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6278         return (0);
6279 }
6280
6281 /*
6282  * Persists timer, here we simply send the
6283  * same thing as a keepalive will.
6284  * the one byte send.
6285  *
6286  * We only return 1, saying don't proceed, if all timers
6287  * are stopped (destroyed PCB?).
6288  */
6289 static int
6290 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6291 {
6292         struct tcptemp *t_template;
6293         int32_t retval = 1;
6294
6295         if (rack->rc_in_persist == 0)
6296                 return (0);
6297         if (ctf_progress_timeout_check(tp, false)) {
6298                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6299                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6300                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6301                 return (-ETIMEDOUT);    /* tcp_drop() */
6302         }
6303         /*
6304          * Persistence timer into zero window. Force a byte to be output, if
6305          * possible.
6306          */
6307         KMOD_TCPSTAT_INC(tcps_persisttimeo);
6308         /*
6309          * Hack: if the peer is dead/unreachable, we do not time out if the
6310          * window is closed.  After a full backoff, drop the connection if
6311          * the idle time (no responses to probes) reaches the maximum
6312          * backoff that we would use if retransmitting.
6313          */
6314         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6315             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6316              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6317                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6318                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6319                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6320                 retval = -ETIMEDOUT;    /* tcp_drop() */
6321                 goto out;
6322         }
6323         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6324             tp->snd_una == tp->snd_max)
6325                 rack_exit_persist(tp, rack, cts);
6326         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6327         /*
6328          * If the user has closed the socket then drop a persisting
6329          * connection after a much reduced timeout.
6330          */
6331         if (tp->t_state > TCPS_CLOSE_WAIT &&
6332             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6333                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6334                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6335                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6336                 retval = -ETIMEDOUT;    /* tcp_drop() */
6337                 goto out;
6338         }
6339         t_template = tcpip_maketemplate(rack->rc_inp);
6340         if (t_template) {
6341                 /* only set it if we were answered */
6342                 if (rack->forced_ack == 0) {
6343                         rack->forced_ack = 1;
6344                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6345                 } else {
6346                         rack->probe_not_answered = 1;
6347                         counter_u64_add(rack_persists_loss, 1);
6348                         rack->r_ctl.persist_lost_ends++;
6349                 }
6350                 counter_u64_add(rack_persists_sends, 1);
6351                 tcp_respond(tp, t_template->tt_ipgen,
6352                             &t_template->tt_t, (struct mbuf *)NULL,
6353                             tp->rcv_nxt, tp->snd_una - 1, 0);
6354                 /* This sends an ack */
6355                 if (tp->t_flags & TF_DELACK)
6356                         tp->t_flags &= ~TF_DELACK;
6357                 free(t_template, M_TEMP);
6358         }
6359         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6360                 tp->t_rxtshift++;
6361 out:
6362         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6363         rack_start_hpts_timer(rack, tp, cts,
6364                               0, 0, 0);
6365         return (retval);
6366 }
6367
6368 /*
6369  * If a keepalive goes off, we had no other timers
6370  * happening. We always return 1 here since this
6371  * routine either drops the connection or sends
6372  * out a segment with respond.
6373  */
6374 static int
6375 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6376 {
6377         struct tcptemp *t_template;
6378         struct inpcb *inp = tptoinpcb(tp);
6379
6380         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6381         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6382         /*
6383          * Keep-alive timer went off; send something or drop connection if
6384          * idle for too long.
6385          */
6386         KMOD_TCPSTAT_INC(tcps_keeptimeo);
6387         if (tp->t_state < TCPS_ESTABLISHED)
6388                 goto dropit;
6389         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6390             tp->t_state <= TCPS_CLOSING) {
6391                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6392                         goto dropit;
6393                 /*
6394                  * Send a packet designed to force a response if the peer is
6395                  * up and reachable: either an ACK if the connection is
6396                  * still alive, or an RST if the peer has closed the
6397                  * connection due to timeout or reboot. Using sequence
6398                  * number tp->snd_una-1 causes the transmitted zero-length
6399                  * segment to lie outside the receive window; by the
6400                  * protocol spec, this requires the correspondent TCP to
6401                  * respond.
6402                  */
6403                 KMOD_TCPSTAT_INC(tcps_keepprobe);
6404                 t_template = tcpip_maketemplate(inp);
6405                 if (t_template) {
6406                         if (rack->forced_ack == 0) {
6407                                 rack->forced_ack = 1;
6408                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6409                         } else {
6410                                 rack->probe_not_answered = 1;
6411                         }
6412                         tcp_respond(tp, t_template->tt_ipgen,
6413                             &t_template->tt_t, (struct mbuf *)NULL,
6414                             tp->rcv_nxt, tp->snd_una - 1, 0);
6415                         free(t_template, M_TEMP);
6416                 }
6417         }
6418         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6419         return (1);
6420 dropit:
6421         KMOD_TCPSTAT_INC(tcps_keepdrops);
6422         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6423         return (-ETIMEDOUT);    /* tcp_drop() */
6424 }
6425
6426 /*
6427  * Retransmit helper function, clear up all the ack
6428  * flags and take care of important book keeping.
6429  */
6430 static void
6431 rack_remxt_tmr(struct tcpcb *tp)
6432 {
6433         /*
6434          * The retransmit timer went off, all sack'd blocks must be
6435          * un-acked.
6436          */
6437         struct rack_sendmap *rsm, *trsm = NULL;
6438         struct tcp_rack *rack;
6439
6440         rack = (struct tcp_rack *)tp->t_fb_ptr;
6441         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6442         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6443         if (rack->r_state && (rack->r_state != tp->t_state))
6444                 rack_set_state(tp, rack);
6445         /*
6446          * Ideally we would like to be able to
6447          * mark SACK-PASS on anything not acked here.
6448          *
6449          * However, if we do that we would burst out
6450          * all that data 1ms apart. This would be unwise,
6451          * so for now we will just let the normal rxt timer
6452          * and tlp timer take care of it.
6453          *
6454          * Also we really need to stick them back in sequence
6455          * order. This way we send in the proper order and any
6456          * sacks that come floating in will "re-ack" the data.
6457          * To do this we zap the tmap with an INIT and then
6458          * walk through and place every rsm in the RB tree
6459          * back in its seq ordered place.
6460          */
6461         TAILQ_INIT(&rack->r_ctl.rc_tmap);
6462         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6463                 rsm->r_dupack = 0;
6464                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6465                 /* We must re-add it back to the tlist */
6466                 if (trsm == NULL) {
6467                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6468                 } else {
6469                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6470                 }
6471                 rsm->r_in_tmap = 1;
6472                 trsm = rsm;
6473                 if (rsm->r_flags & RACK_ACKED)
6474                         rsm->r_flags |= RACK_WAS_ACKED;
6475                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
6476                 rsm->r_flags |= RACK_MUST_RXT;
6477         }
6478         /* Clear the count (we just un-acked them) */
6479         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6480         rack->r_ctl.rc_sacked = 0;
6481         rack->r_ctl.rc_sacklast = NULL;
6482         rack->r_ctl.rc_agg_delayed = 0;
6483         rack->r_early = 0;
6484         rack->r_ctl.rc_agg_early = 0;
6485         rack->r_late = 0;
6486         /* Clear the tlp rtx mark */
6487         rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6488         if (rack->r_ctl.rc_resend != NULL)
6489                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6490         rack->r_ctl.rc_prr_sndcnt = 0;
6491         rack_log_to_prr(rack, 6, 0, __LINE__);
6492         rack->r_timer_override = 1;
6493         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6494 #ifdef NETFLIX_EXP_DETECTION
6495             || (rack->sack_attack_disable != 0)
6496 #endif
6497                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6498                 /*
6499                  * For non-sack customers new data
6500                  * needs to go out as retransmits until
6501                  * we retransmit up to snd_max.
6502                  */
6503                 rack->r_must_retran = 1;
6504                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6505                                                 rack->r_ctl.rc_sacked);
6506         }
6507         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6508 }
6509
6510 static void
6511 rack_convert_rtts(struct tcpcb *tp)
6512 {
6513         if (tp->t_srtt > 1) {
6514                 uint32_t val, frac;
6515
6516                 val = tp->t_srtt >> TCP_RTT_SHIFT;
6517                 frac = tp->t_srtt & 0x1f;
6518                 tp->t_srtt = TICKS_2_USEC(val);
6519                 /*
6520                  * frac is the fractional part of the srtt (if any)
6521                  * but its in ticks and every bit represents
6522                  * 1/32nd of a hz.
6523                  */
6524                 if (frac) {
6525                         if (hz == 1000) {
6526                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6527                         } else {
6528                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6529                         }
6530                         tp->t_srtt += frac;
6531                 }
6532         }
6533         if (tp->t_rttvar) {
6534                 uint32_t val, frac;
6535
6536                 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6537                 frac = tp->t_rttvar & 0x1f;
6538                 tp->t_rttvar = TICKS_2_USEC(val);
6539                 /*
6540                  * frac is the fractional part of the srtt (if any)
6541                  * but its in ticks and every bit represents
6542                  * 1/32nd of a hz.
6543                  */
6544                 if (frac) {
6545                         if (hz == 1000) {
6546                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6547                         } else {
6548                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6549                         }
6550                         tp->t_rttvar += frac;
6551                 }
6552         }
6553         tp->t_rxtcur = RACK_REXMTVAL(tp);
6554         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6555                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6556         }
6557         if (tp->t_rxtcur > rack_rto_max) {
6558                 tp->t_rxtcur = rack_rto_max;
6559         }
6560 }
6561
6562 static void
6563 rack_cc_conn_init(struct tcpcb *tp)
6564 {
6565         struct tcp_rack *rack;
6566         uint32_t srtt;
6567
6568         rack = (struct tcp_rack *)tp->t_fb_ptr;
6569         srtt = tp->t_srtt;
6570         cc_conn_init(tp);
6571         /*
6572          * Now convert to rack's internal format,
6573          * if required.
6574          */
6575         if ((srtt == 0) && (tp->t_srtt != 0))
6576                 rack_convert_rtts(tp);
6577         /*
6578          * We want a chance to stay in slowstart as
6579          * we create a connection. TCP spec says that
6580          * initially ssthresh is infinite. For our
6581          * purposes that is the snd_wnd.
6582          */
6583         if (tp->snd_ssthresh < tp->snd_wnd) {
6584                 tp->snd_ssthresh = tp->snd_wnd;
6585         }
6586         /*
6587          * We also want to assure a IW worth of
6588          * data can get inflight.
6589          */
6590         if (rc_init_window(rack) < tp->snd_cwnd)
6591                 tp->snd_cwnd = rc_init_window(rack);
6592 }
6593
6594 /*
6595  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6596  * we will setup to retransmit the lowest seq number outstanding.
6597  */
6598 static int
6599 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6600 {
6601         struct inpcb *inp = tptoinpcb(tp);
6602         int32_t rexmt;
6603         int32_t retval = 0;
6604         bool isipv6;
6605
6606         if ((tp->t_flags & TF_GPUTINPROG) &&
6607             (tp->t_rxtshift)) {
6608                 /*
6609                  * We have had a second timeout
6610                  * measurements on successive rxt's are not profitable.
6611                  * It is unlikely to be of any use (the network is
6612                  * broken or the client went away).
6613                  */
6614                 tp->t_flags &= ~TF_GPUTINPROG;
6615                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6616                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6617                                            tp->gput_seq,
6618                                            0, 0, 18, __LINE__, NULL, 0);
6619         }
6620         if (ctf_progress_timeout_check(tp, false)) {
6621                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6622                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6623                 return (-ETIMEDOUT);    /* tcp_drop() */
6624         }
6625         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6626         rack->r_ctl.retran_during_recovery = 0;
6627         rack->rc_ack_required = 1;
6628         rack->r_ctl.dsack_byte_cnt = 0;
6629         if (IN_FASTRECOVERY(tp->t_flags))
6630                 tp->t_flags |= TF_WASFRECOVERY;
6631         else
6632                 tp->t_flags &= ~TF_WASFRECOVERY;
6633         if (IN_CONGRECOVERY(tp->t_flags))
6634                 tp->t_flags |= TF_WASCRECOVERY;
6635         else
6636                 tp->t_flags &= ~TF_WASCRECOVERY;
6637         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6638             (tp->snd_una == tp->snd_max)) {
6639                 /* Nothing outstanding .. nothing to do */
6640                 return (0);
6641         }
6642         if (rack->r_ctl.dsack_persist) {
6643                 rack->r_ctl.dsack_persist--;
6644                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6645                         rack->r_ctl.num_dsack = 0;
6646                 }
6647                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6648         }
6649         /*
6650          * Rack can only run one timer  at a time, so we cannot
6651          * run a KEEPINIT (gating SYN sending) and a retransmit
6652          * timer for the SYN. So if we are in a front state and
6653          * have a KEEPINIT timer we need to check the first transmit
6654          * against now to see if we have exceeded the KEEPINIT time
6655          * (if one is set).
6656          */
6657         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6658             (TP_KEEPINIT(tp) != 0)) {
6659                 struct rack_sendmap *rsm;
6660
6661                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6662                 if (rsm) {
6663                         /* Ok we have something outstanding to test keepinit with */
6664                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6665                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6666                                 /* We have exceeded the KEEPINIT time */
6667                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6668                                 goto drop_it;
6669                         }
6670                 }
6671         }
6672         /*
6673          * Retransmission timer went off.  Message has not been acked within
6674          * retransmit interval.  Back off to a longer retransmit interval
6675          * and retransmit one segment.
6676          */
6677         rack_remxt_tmr(tp);
6678         if ((rack->r_ctl.rc_resend == NULL) ||
6679             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6680                 /*
6681                  * If the rwnd collapsed on
6682                  * the one we are retransmitting
6683                  * it does not count against the
6684                  * rxt count.
6685                  */
6686                 tp->t_rxtshift++;
6687         }
6688         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6689                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6690 drop_it:
6691                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6692                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6693                 /* XXXGL: previously t_softerror was casted to uint16_t */
6694                 MPASS(tp->t_softerror >= 0);
6695                 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6696                 goto out;       /* tcp_drop() */
6697         }
6698         if (tp->t_state == TCPS_SYN_SENT) {
6699                 /*
6700                  * If the SYN was retransmitted, indicate CWND to be limited
6701                  * to 1 segment in cc_conn_init().
6702                  */
6703                 tp->snd_cwnd = 1;
6704         } else if (tp->t_rxtshift == 1) {
6705                 /*
6706                  * first retransmit; record ssthresh and cwnd so they can be
6707                  * recovered if this turns out to be a "bad" retransmit. A
6708                  * retransmit is considered "bad" if an ACK for this segment
6709                  * is received within RTT/2 interval; the assumption here is
6710                  * that the ACK was already in flight.  See "On Estimating
6711                  * End-to-End Network Path Properties" by Allman and Paxson
6712                  * for more details.
6713                  */
6714                 tp->snd_cwnd_prev = tp->snd_cwnd;
6715                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6716                 tp->snd_recover_prev = tp->snd_recover;
6717                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6718                 tp->t_flags |= TF_PREVVALID;
6719         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6720                 tp->t_flags &= ~TF_PREVVALID;
6721         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6722         if ((tp->t_state == TCPS_SYN_SENT) ||
6723             (tp->t_state == TCPS_SYN_RECEIVED))
6724                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6725         else
6726                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6727
6728         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6729            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6730         /*
6731          * We enter the path for PLMTUD if connection is established or, if
6732          * connection is FIN_WAIT_1 status, reason for the last is that if
6733          * amount of data we send is very small, we could send it in couple
6734          * of packets and process straight to FIN. In that case we won't
6735          * catch ESTABLISHED state.
6736          */
6737 #ifdef INET6
6738         isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
6739 #else
6740         isipv6 = false;
6741 #endif
6742         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6743             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6744             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6745             ((tp->t_state == TCPS_ESTABLISHED) ||
6746             (tp->t_state == TCPS_FIN_WAIT_1))) {
6747                 /*
6748                  * Idea here is that at each stage of mtu probe (usually,
6749                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6750                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6751                  * should take care of that.
6752                  */
6753                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6754                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6755                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6756                     tp->t_rxtshift % 2 == 0)) {
6757                         /*
6758                          * Enter Path MTU Black-hole Detection mechanism: -
6759                          * Disable Path MTU Discovery (IP "DF" bit). -
6760                          * Reduce MTU to lower value than what we negotiated
6761                          * with peer.
6762                          */
6763                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6764                                 /* Record that we may have found a black hole. */
6765                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6766                                 /* Keep track of previous MSS. */
6767                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6768                         }
6769
6770                         /*
6771                          * Reduce the MSS to blackhole value or to the
6772                          * default in an attempt to retransmit.
6773                          */
6774 #ifdef INET6
6775                         if (isipv6 &&
6776                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6777                                 /* Use the sysctl tuneable blackhole MSS. */
6778                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6779                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6780                         } else if (isipv6) {
6781                                 /* Use the default MSS. */
6782                                 tp->t_maxseg = V_tcp_v6mssdflt;
6783                                 /*
6784                                  * Disable Path MTU Discovery when we switch
6785                                  * to minmss.
6786                                  */
6787                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6788                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6789                         }
6790 #endif
6791 #if defined(INET6) && defined(INET)
6792                         else
6793 #endif
6794 #ifdef INET
6795                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6796                                 /* Use the sysctl tuneable blackhole MSS. */
6797                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6798                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6799                         } else {
6800                                 /* Use the default MSS. */
6801                                 tp->t_maxseg = V_tcp_mssdflt;
6802                                 /*
6803                                  * Disable Path MTU Discovery when we switch
6804                                  * to minmss.
6805                                  */
6806                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6807                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6808                         }
6809 #endif
6810                 } else {
6811                         /*
6812                          * If further retransmissions are still unsuccessful
6813                          * with a lowered MTU, maybe this isn't a blackhole
6814                          * and we restore the previous MSS and blackhole
6815                          * detection flags. The limit '6' is determined by
6816                          * giving each probe stage (1448, 1188, 524) 2
6817                          * chances to recover.
6818                          */
6819                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6820                             (tp->t_rxtshift >= 6)) {
6821                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6822                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6823                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6824                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6825                         }
6826                 }
6827         }
6828         /*
6829          * Disable RFC1323 and SACK if we haven't got any response to
6830          * our third SYN to work-around some broken terminal servers
6831          * (most of which have hopefully been retired) that have bad VJ
6832          * header compression code which trashes TCP segments containing
6833          * unknown-to-them TCP options.
6834          */
6835         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6836             (tp->t_rxtshift == 3))
6837                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6838         /*
6839          * If we backed off this far, our srtt estimate is probably bogus.
6840          * Clobber it so we'll take the next rtt measurement as our srtt;
6841          * move the current srtt into rttvar to keep the current retransmit
6842          * times until then.
6843          */
6844         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6845 #ifdef INET6
6846                 if ((inp->inp_vflag & INP_IPV6) != 0)
6847                         in6_losing(inp);
6848                 else
6849 #endif
6850                         in_losing(inp);
6851                 tp->t_rttvar += tp->t_srtt;
6852                 tp->t_srtt = 0;
6853         }
6854         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6855         tp->snd_recover = tp->snd_max;
6856         tp->t_flags |= TF_ACKNOW;
6857         tp->t_rtttime = 0;
6858         rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
6859 out:
6860         return (retval);
6861 }
6862
6863 static int
6864 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6865 {
6866         int32_t ret = 0;
6867         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6868
6869         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6870             (tp->t_flags & TF_GPUTINPROG)) {
6871                 /*
6872                  * We have a goodput in progress
6873                  * and we have entered a late state.
6874                  * Do we have enough data in the sb
6875                  * to handle the GPUT request?
6876                  */
6877                 uint32_t bytes;
6878
6879                 bytes = tp->gput_ack - tp->gput_seq;
6880                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
6881                         bytes += tp->gput_seq - tp->snd_una;
6882                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
6883                         /*
6884                          * There are not enough bytes in the socket
6885                          * buffer that have been sent to cover this
6886                          * measurement. Cancel it.
6887                          */
6888                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6889                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
6890                                                    tp->gput_seq,
6891                                                    0, 0, 18, __LINE__, NULL, 0);
6892                         tp->t_flags &= ~TF_GPUTINPROG;
6893                 }
6894         }
6895         if (timers == 0) {
6896                 return (0);
6897         }
6898         if (tp->t_state == TCPS_LISTEN) {
6899                 /* no timers on listen sockets */
6900                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6901                         return (0);
6902                 return (1);
6903         }
6904         if ((timers & PACE_TMR_RACK) &&
6905             rack->rc_on_min_to) {
6906                 /*
6907                  * For the rack timer when we
6908                  * are on a min-timeout (which means rrr_conf = 3)
6909                  * we don't want to check the timer. It may
6910                  * be going off for a pace and thats ok we
6911                  * want to send the retransmit (if its ready).
6912                  *
6913                  * If its on a normal rack timer (non-min) then
6914                  * we will check if its expired.
6915                  */
6916                 goto skip_time_check;
6917         }
6918         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6919                 uint32_t left;
6920
6921                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6922                         ret = -1;
6923                         rack_log_to_processing(rack, cts, ret, 0);
6924                         return (0);
6925                 }
6926                 if (hpts_calling == 0) {
6927                         /*
6928                          * A user send or queued mbuf (sack) has called us? We
6929                          * return 0 and let the pacing guards
6930                          * deal with it if they should or
6931                          * should not cause a send.
6932                          */
6933                         ret = -2;
6934                         rack_log_to_processing(rack, cts, ret, 0);
6935                         return (0);
6936                 }
6937                 /*
6938                  * Ok our timer went off early and we are not paced false
6939                  * alarm, go back to sleep.
6940                  */
6941                 ret = -3;
6942                 left = rack->r_ctl.rc_timer_exp - cts;
6943                 tcp_hpts_insert(tptoinpcb(tp), HPTS_MS_TO_SLOTS(left));
6944                 rack_log_to_processing(rack, cts, ret, left);
6945                 return (1);
6946         }
6947 skip_time_check:
6948         rack->rc_tmr_stopped = 0;
6949         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6950         if (timers & PACE_TMR_DELACK) {
6951                 ret = rack_timeout_delack(tp, rack, cts);
6952         } else if (timers & PACE_TMR_RACK) {
6953                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6954                 rack->r_fast_output = 0;
6955                 ret = rack_timeout_rack(tp, rack, cts);
6956         } else if (timers & PACE_TMR_TLP) {
6957                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6958                 ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
6959         } else if (timers & PACE_TMR_RXT) {
6960                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6961                 rack->r_fast_output = 0;
6962                 ret = rack_timeout_rxt(tp, rack, cts);
6963         } else if (timers & PACE_TMR_PERSIT) {
6964                 ret = rack_timeout_persist(tp, rack, cts);
6965         } else if (timers & PACE_TMR_KEEP) {
6966                 ret = rack_timeout_keepalive(tp, rack, cts);
6967         }
6968         rack_log_to_processing(rack, cts, ret, timers);
6969         return (ret);
6970 }
6971
6972 static void
6973 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6974 {
6975         struct timeval tv;
6976         uint32_t us_cts, flags_on_entry;
6977         uint8_t hpts_removed = 0;
6978
6979         flags_on_entry = rack->r_ctl.rc_hpts_flags;
6980         us_cts = tcp_get_usecs(&tv);
6981         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
6982             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
6983              ((tp->snd_max - tp->snd_una) == 0))) {
6984                 tcp_hpts_remove(rack->rc_inp);
6985                 hpts_removed = 1;
6986                 /* If we were not delayed cancel out the flag. */
6987                 if ((tp->snd_max - tp->snd_una) == 0)
6988                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6989                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6990         }
6991         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
6992                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
6993                 if (tcp_in_hpts(rack->rc_inp) &&
6994                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
6995                         /*
6996                          * Canceling timer's when we have no output being
6997                          * paced. We also must remove ourselves from the
6998                          * hpts.
6999                          */
7000                         tcp_hpts_remove(rack->rc_inp);
7001                         hpts_removed = 1;
7002                 }
7003                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7004         }
7005         if (hpts_removed == 0)
7006                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7007 }
7008
7009 static int
7010 rack_stopall(struct tcpcb *tp)
7011 {
7012         struct tcp_rack *rack;
7013         rack = (struct tcp_rack *)tp->t_fb_ptr;
7014         rack->t_timers_stopped = 1;
7015         return (0);
7016 }
7017
7018 static void
7019 rack_stop_all_timers(struct tcpcb *tp)
7020 {
7021         struct tcp_rack *rack;
7022
7023         /*
7024          * Assure no timers are running.
7025          */
7026         if (tcp_timer_active(tp, TT_PERSIST)) {
7027                 /* We enter in persists, set the flag appropriately */
7028                 rack = (struct tcp_rack *)tp->t_fb_ptr;
7029                 rack->rc_in_persist = 1;
7030         }
7031 }
7032
7033 static void
7034 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7035     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7036 {
7037         int32_t idx;
7038
7039         rsm->r_rtr_cnt++;
7040         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7041         rsm->r_dupack = 0;
7042         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7043                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7044                 rsm->r_flags |= RACK_OVERMAX;
7045         }
7046         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7047                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7048                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7049         }
7050         idx = rsm->r_rtr_cnt - 1;
7051         rsm->r_tim_lastsent[idx] = ts;
7052         /*
7053          * Here we don't add in the len of send, since its already
7054          * in snduna <->snd_max.
7055          */
7056         rsm->r_fas = ctf_flight_size(rack->rc_tp,
7057                                      rack->r_ctl.rc_sacked);
7058         if (rsm->r_flags & RACK_ACKED) {
7059                 /* Problably MTU discovery messing with us */
7060                 rsm->r_flags &= ~RACK_ACKED;
7061                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7062         }
7063         if (rsm->r_in_tmap) {
7064                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7065                 rsm->r_in_tmap = 0;
7066         }
7067         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7068         rsm->r_in_tmap = 1;
7069         /* Take off the must retransmit flag, if its on */
7070         if (rsm->r_flags & RACK_MUST_RXT) {
7071                 if (rack->r_must_retran)
7072                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
7073                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
7074                         /*
7075                          * We have retransmitted all we need. Clear
7076                          * any must retransmit flags.
7077                          */
7078                         rack->r_must_retran = 0;
7079                         rack->r_ctl.rc_out_at_rto = 0;
7080                 }
7081                 rsm->r_flags &= ~RACK_MUST_RXT;
7082         }
7083         if (rsm->r_flags & RACK_SACK_PASSED) {
7084                 /* We have retransmitted due to the SACK pass */
7085                 rsm->r_flags &= ~RACK_SACK_PASSED;
7086                 rsm->r_flags |= RACK_WAS_SACKPASS;
7087         }
7088 }
7089
7090 static uint32_t
7091 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7092     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7093 {
7094         /*
7095          * We (re-)transmitted starting at rsm->r_start for some length
7096          * (possibly less than r_end.
7097          */
7098         struct rack_sendmap *nrsm;
7099 #ifdef INVARIANTS
7100         struct rack_sendmap *insret;
7101 #endif
7102         uint32_t c_end;
7103         int32_t len;
7104
7105         len = *lenp;
7106         c_end = rsm->r_start + len;
7107         if (SEQ_GEQ(c_end, rsm->r_end)) {
7108                 /*
7109                  * We retransmitted the whole piece or more than the whole
7110                  * slopping into the next rsm.
7111                  */
7112                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7113                 if (c_end == rsm->r_end) {
7114                         *lenp = 0;
7115                         return (0);
7116                 } else {
7117                         int32_t act_len;
7118
7119                         /* Hangs over the end return whats left */
7120                         act_len = rsm->r_end - rsm->r_start;
7121                         *lenp = (len - act_len);
7122                         return (rsm->r_end);
7123                 }
7124                 /* We don't get out of this block. */
7125         }
7126         /*
7127          * Here we retransmitted less than the whole thing which means we
7128          * have to split this into what was transmitted and what was not.
7129          */
7130         nrsm = rack_alloc_full_limit(rack);
7131         if (nrsm == NULL) {
7132                 /*
7133                  * We can't get memory, so lets not proceed.
7134                  */
7135                 *lenp = 0;
7136                 return (0);
7137         }
7138         /*
7139          * So here we are going to take the original rsm and make it what we
7140          * retransmitted. nrsm will be the tail portion we did not
7141          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7142          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7143          * 1, 6 and the new piece will be 6, 11.
7144          */
7145         rack_clone_rsm(rack, nrsm, rsm, c_end);
7146         nrsm->r_dupack = 0;
7147         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7148 #ifndef INVARIANTS
7149         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7150 #else
7151         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7152         if (insret != NULL) {
7153                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7154                       nrsm, insret, rack, rsm);
7155         }
7156 #endif
7157         if (rsm->r_in_tmap) {
7158                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7159                 nrsm->r_in_tmap = 1;
7160         }
7161         rsm->r_flags &= (~RACK_HAS_FIN);
7162         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7163         /* Log a split of rsm into rsm and nrsm */
7164         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7165         *lenp = 0;
7166         return (0);
7167 }
7168
7169 static void
7170 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7171                 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
7172                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7173 {
7174         struct tcp_rack *rack;
7175         struct rack_sendmap *rsm, *nrsm, fe;
7176 #ifdef INVARIANTS
7177         struct rack_sendmap *insret;
7178 #endif
7179         register uint32_t snd_max, snd_una;
7180
7181         /*
7182          * Add to the RACK log of packets in flight or retransmitted. If
7183          * there is a TS option we will use the TS echoed, if not we will
7184          * grab a TS.
7185          *
7186          * Retransmissions will increment the count and move the ts to its
7187          * proper place. Note that if options do not include TS's then we
7188          * won't be able to effectively use the ACK for an RTT on a retran.
7189          *
7190          * Notes about r_start and r_end. Lets consider a send starting at
7191          * sequence 1 for 10 bytes. In such an example the r_start would be
7192          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7193          * This means that r_end is actually the first sequence for the next
7194          * slot (11).
7195          *
7196          */
7197         /*
7198          * If err is set what do we do XXXrrs? should we not add the thing?
7199          * -- i.e. return if err != 0 or should we pretend we sent it? --
7200          * i.e. proceed with add ** do this for now.
7201          */
7202         INP_WLOCK_ASSERT(tptoinpcb(tp));
7203         if (err)
7204                 /*
7205                  * We don't log errors -- we could but snd_max does not
7206                  * advance in this case either.
7207                  */
7208                 return;
7209
7210         if (th_flags & TH_RST) {
7211                 /*
7212                  * We don't log resets and we return immediately from
7213                  * sending
7214                  */
7215                 return;
7216         }
7217         rack = (struct tcp_rack *)tp->t_fb_ptr;
7218         snd_una = tp->snd_una;
7219         snd_max = tp->snd_max;
7220         if (th_flags & (TH_SYN | TH_FIN)) {
7221                 /*
7222                  * The call to rack_log_output is made before bumping
7223                  * snd_max. This means we can record one extra byte on a SYN
7224                  * or FIN if seq_out is adding more on and a FIN is present
7225                  * (and we are not resending).
7226                  */
7227                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7228                         len++;
7229                 if (th_flags & TH_FIN)
7230                         len++;
7231                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7232                         /*
7233                          * The add/update as not been done for the FIN/SYN
7234                          * yet.
7235                          */
7236                         snd_max = tp->snd_nxt;
7237                 }
7238         }
7239         if (SEQ_LEQ((seq_out + len), snd_una)) {
7240                 /* Are sending an old segment to induce an ack (keep-alive)? */
7241                 return;
7242         }
7243         if (SEQ_LT(seq_out, snd_una)) {
7244                 /* huh? should we panic? */
7245                 uint32_t end;
7246
7247                 end = seq_out + len;
7248                 seq_out = snd_una;
7249                 if (SEQ_GEQ(end, seq_out))
7250                         len = end - seq_out;
7251                 else
7252                         len = 0;
7253         }
7254         if (len == 0) {
7255                 /* We don't log zero window probes */
7256                 return;
7257         }
7258         if (IN_FASTRECOVERY(tp->t_flags)) {
7259                 rack->r_ctl.rc_prr_out += len;
7260         }
7261         /* First question is it a retransmission or new? */
7262         if (seq_out == snd_max) {
7263                 /* Its new */
7264 again:
7265                 rsm = rack_alloc(rack);
7266                 if (rsm == NULL) {
7267                         /*
7268                          * Hmm out of memory and the tcb got destroyed while
7269                          * we tried to wait.
7270                          */
7271                         return;
7272                 }
7273                 if (th_flags & TH_FIN) {
7274                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7275                 } else {
7276                         rsm->r_flags = add_flag;
7277                 }
7278                 if (hw_tls)
7279                         rsm->r_hw_tls = 1;
7280                 rsm->r_tim_lastsent[0] = cts;
7281                 rsm->r_rtr_cnt = 1;
7282                 rsm->r_rtr_bytes = 0;
7283                 if (th_flags & TH_SYN) {
7284                         /* The data space is one beyond snd_una */
7285                         rsm->r_flags |= RACK_HAS_SYN;
7286                 }
7287                 rsm->r_start = seq_out;
7288                 rsm->r_end = rsm->r_start + len;
7289                 rsm->r_dupack = 0;
7290                 /*
7291                  * save off the mbuf location that
7292                  * sndmbuf_noadv returned (which is
7293                  * where we started copying from)..
7294                  */
7295                 rsm->m = s_mb;
7296                 rsm->soff = s_moff;
7297                 /*
7298                  * Here we do add in the len of send, since its not yet
7299                  * reflected in in snduna <->snd_max
7300                  */
7301                 rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7302                                               rack->r_ctl.rc_sacked) +
7303                               (rsm->r_end - rsm->r_start));
7304                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7305                 if (rsm->m) {
7306                         if (rsm->m->m_len <= rsm->soff) {
7307                                 /*
7308                                  * XXXrrs Question, will this happen?
7309                                  *
7310                                  * If sbsndptr is set at the correct place
7311                                  * then s_moff should always be somewhere
7312                                  * within rsm->m. But if the sbsndptr was
7313                                  * off then that won't be true. If it occurs
7314                                  * we need to walkout to the correct location.
7315                                  */
7316                                 struct mbuf *lm;
7317
7318                                 lm = rsm->m;
7319                                 while (lm->m_len <= rsm->soff) {
7320                                         rsm->soff -= lm->m_len;
7321                                         lm = lm->m_next;
7322                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7323                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7324                                 }
7325                                 rsm->m = lm;
7326                         }
7327                         rsm->orig_m_len = rsm->m->m_len;
7328                 } else
7329                         rsm->orig_m_len = 0;
7330                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7331                 /* Log a new rsm */
7332                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7333 #ifndef INVARIANTS
7334                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7335 #else
7336                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7337                 if (insret != NULL) {
7338                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7339                               nrsm, insret, rack, rsm);
7340                 }
7341 #endif
7342                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7343                 rsm->r_in_tmap = 1;
7344                 /*
7345                  * Special case detection, is there just a single
7346                  * packet outstanding when we are not in recovery?
7347                  *
7348                  * If this is true mark it so.
7349                  */
7350                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7351                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7352                         struct rack_sendmap *prsm;
7353
7354                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7355                         if (prsm)
7356                                 prsm->r_one_out_nr = 1;
7357                 }
7358                 return;
7359         }
7360         /*
7361          * If we reach here its a retransmission and we need to find it.
7362          */
7363         memset(&fe, 0, sizeof(fe));
7364 more:
7365         if (hintrsm && (hintrsm->r_start == seq_out)) {
7366                 rsm = hintrsm;
7367                 hintrsm = NULL;
7368         } else {
7369                 /* No hints sorry */
7370                 rsm = NULL;
7371         }
7372         if ((rsm) && (rsm->r_start == seq_out)) {
7373                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7374                 if (len == 0) {
7375                         return;
7376                 } else {
7377                         goto more;
7378                 }
7379         }
7380         /* Ok it was not the last pointer go through it the hard way. */
7381 refind:
7382         fe.r_start = seq_out;
7383         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7384         if (rsm) {
7385                 if (rsm->r_start == seq_out) {
7386                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7387                         if (len == 0) {
7388                                 return;
7389                         } else {
7390                                 goto refind;
7391                         }
7392                 }
7393                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7394                         /* Transmitted within this piece */
7395                         /*
7396                          * Ok we must split off the front and then let the
7397                          * update do the rest
7398                          */
7399                         nrsm = rack_alloc_full_limit(rack);
7400                         if (nrsm == NULL) {
7401                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7402                                 return;
7403                         }
7404                         /*
7405                          * copy rsm to nrsm and then trim the front of rsm
7406                          * to not include this part.
7407                          */
7408                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7409                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7410 #ifndef INVARIANTS
7411                         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7412 #else
7413                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7414                         if (insret != NULL) {
7415                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7416                                       nrsm, insret, rack, rsm);
7417                         }
7418 #endif
7419                         if (rsm->r_in_tmap) {
7420                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7421                                 nrsm->r_in_tmap = 1;
7422                         }
7423                         rsm->r_flags &= (~RACK_HAS_FIN);
7424                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7425                         if (len == 0) {
7426                                 return;
7427                         } else if (len > 0)
7428                                 goto refind;
7429                 }
7430         }
7431         /*
7432          * Hmm not found in map did they retransmit both old and on into the
7433          * new?
7434          */
7435         if (seq_out == tp->snd_max) {
7436                 goto again;
7437         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7438 #ifdef INVARIANTS
7439                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7440                        seq_out, len, tp->snd_una, tp->snd_max);
7441                 printf("Starting Dump of all rack entries\n");
7442                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7443                         printf("rsm:%p start:%u end:%u\n",
7444                                rsm, rsm->r_start, rsm->r_end);
7445                 }
7446                 printf("Dump complete\n");
7447                 panic("seq_out not found rack:%p tp:%p",
7448                       rack, tp);
7449 #endif
7450         } else {
7451 #ifdef INVARIANTS
7452                 /*
7453                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7454                  * flag)
7455                  */
7456                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7457                       seq_out, len, tp->snd_max, tp);
7458 #endif
7459         }
7460 }
7461
7462 /*
7463  * Record one of the RTT updates from an ack into
7464  * our sample structure.
7465  */
7466
7467 static void
7468 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7469                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7470 {
7471         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7472             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7473                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7474         }
7475         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7476             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7477                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7478         }
7479         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7480             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7481                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7482             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7483                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7484         }
7485         if ((confidence == 1) &&
7486             ((rsm == NULL) ||
7487              (rsm->r_just_ret) ||
7488              (rsm->r_one_out_nr &&
7489               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7490                 /*
7491                  * If the rsm had a just return
7492                  * hit it then we can't trust the
7493                  * rtt measurement for buffer deterimination
7494                  * Note that a confidence of 2, indicates
7495                  * SACK'd which overrides the r_just_ret or
7496                  * the r_one_out_nr. If it was a CUM-ACK and
7497                  * we had only two outstanding, but get an
7498                  * ack for only 1. Then that also lowers our
7499                  * confidence.
7500                  */
7501                 confidence = 0;
7502         }
7503         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7504             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7505                 if (rack->r_ctl.rack_rs.confidence == 0) {
7506                         /*
7507                          * We take anything with no current confidence
7508                          * saved.
7509                          */
7510                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7511                         rack->r_ctl.rack_rs.confidence = confidence;
7512                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7513                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7514                         /*
7515                          * Once we have a confident number,
7516                          * we can update it with a smaller
7517                          * value since this confident number
7518                          * may include the DSACK time until
7519                          * the next segment (the second one) arrived.
7520                          */
7521                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7522                         rack->r_ctl.rack_rs.confidence = confidence;
7523                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7524                 }
7525         }
7526         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7527         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7528         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7529         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7530 }
7531
7532 /*
7533  * Collect new round-trip time estimate
7534  * and update averages and current timeout.
7535  */
7536 static void
7537 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7538 {
7539         int32_t delta;
7540         int32_t rtt;
7541
7542         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7543                 /* No valid sample */
7544                 return;
7545         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7546                 /* We are to use the lowest RTT seen in a single ack */
7547                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7548         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7549                 /* We are to use the highest RTT seen in a single ack */
7550                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7551         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7552                 /* We are to use the average RTT seen in a single ack */
7553                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7554                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7555         } else {
7556 #ifdef INVARIANTS
7557                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7558 #endif
7559                 return;
7560         }
7561         if (rtt == 0)
7562                 rtt = 1;
7563         if (rack->rc_gp_rtt_set == 0) {
7564                 /*
7565                  * With no RTT we have to accept
7566                  * even one we are not confident of.
7567                  */
7568                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7569                 rack->rc_gp_rtt_set = 1;
7570         } else if (rack->r_ctl.rack_rs.confidence) {
7571                 /* update the running gp srtt */
7572                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7573                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7574         }
7575         if (rack->r_ctl.rack_rs.confidence) {
7576                 /*
7577                  * record the low and high for highly buffered path computation,
7578                  * we only do this if we are confident (not a retransmission).
7579                  */
7580                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7581                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7582                 }
7583                 if (rack->rc_highly_buffered == 0) {
7584                         /*
7585                          * Currently once we declare a path has
7586                          * highly buffered there is no going
7587                          * back, which may be a problem...
7588                          */
7589                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7590                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7591                                                      rack->r_ctl.rc_highest_us_rtt,
7592                                                      rack->r_ctl.rc_lowest_us_rtt,
7593                                                      RACK_RTTS_SEEHBP);
7594                                 rack->rc_highly_buffered = 1;
7595                         }
7596                 }
7597         }
7598         if ((rack->r_ctl.rack_rs.confidence) ||
7599             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7600                 /*
7601                  * If we are highly confident of it <or> it was
7602                  * never retransmitted we accept it as the last us_rtt.
7603                  */
7604                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7605                 /* The lowest rtt can be set if its was not retransmited */
7606                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7607                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7608                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7609                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7610                 }
7611         }
7612         rack = (struct tcp_rack *)tp->t_fb_ptr;
7613         if (tp->t_srtt != 0) {
7614                 /*
7615                  * We keep a simple srtt in microseconds, like our rtt
7616                  * measurement. We don't need to do any tricks with shifting
7617                  * etc. Instead we just add in 1/8th of the new measurement
7618                  * and subtract out 1/8 of the old srtt. We do the same with
7619                  * the variance after finding the absolute value of the
7620                  * difference between this sample and the current srtt.
7621                  */
7622                 delta = tp->t_srtt - rtt;
7623                 /* Take off 1/8th of the current sRTT */
7624                 tp->t_srtt -= (tp->t_srtt >> 3);
7625                 /* Add in 1/8th of the new RTT just measured */
7626                 tp->t_srtt += (rtt >> 3);
7627                 if (tp->t_srtt <= 0)
7628                         tp->t_srtt = 1;
7629                 /* Now lets make the absolute value of the variance */
7630                 if (delta < 0)
7631                         delta = -delta;
7632                 /* Subtract out 1/8th */
7633                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7634                 /* Add in 1/8th of the new variance we just saw */
7635                 tp->t_rttvar += (delta >> 3);
7636                 if (tp->t_rttvar <= 0)
7637                         tp->t_rttvar = 1;
7638         } else {
7639                 /*
7640                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7641                  * variance to half the rtt (so our first retransmit happens
7642                  * at 3*rtt).
7643                  */
7644                 tp->t_srtt = rtt;
7645                 tp->t_rttvar = rtt >> 1;
7646         }
7647         rack->rc_srtt_measure_made = 1;
7648         KMOD_TCPSTAT_INC(tcps_rttupdated);
7649         if (tp->t_rttupdated < UCHAR_MAX)
7650                 tp->t_rttupdated++;
7651 #ifdef STATS
7652         if (rack_stats_gets_ms_rtt == 0) {
7653                 /* Send in the microsecond rtt used for rxt timeout purposes */
7654                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7655         } else if (rack_stats_gets_ms_rtt == 1) {
7656                 /* Send in the millisecond rtt used for rxt timeout purposes */
7657                 int32_t ms_rtt;
7658
7659                 /* Round up */
7660                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7661                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7662         } else if (rack_stats_gets_ms_rtt == 2) {
7663                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7664                 int32_t ms_rtt;
7665
7666                 /* Round up */
7667                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7668                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7669         }  else {
7670                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7671                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7672         }
7673
7674 #endif
7675         /*
7676          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7677          * way we do the smoothing, srtt and rttvar will each average +1/2
7678          * tick of bias.  When we compute the retransmit timer, we want 1/2
7679          * tick of rounding and 1 extra tick because of +-1/2 tick
7680          * uncertainty in the firing of the timer.  The bias will give us
7681          * exactly the 1.5 tick we need.  But, because the bias is
7682          * statistical, we have to test that we don't drop below the minimum
7683          * feasible timer (which is 2 ticks).
7684          */
7685         tp->t_rxtshift = 0;
7686         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7687                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7688         rack_log_rtt_sample(rack, rtt);
7689         tp->t_softerror = 0;
7690 }
7691
7692
7693 static void
7694 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7695 {
7696         /*
7697          * Apply to filter the inbound us-rtt at us_cts.
7698          */
7699         uint32_t old_rtt;
7700
7701         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7702         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7703                                us_rtt, us_cts);
7704         if (old_rtt > us_rtt) {
7705                 /* We just hit a new lower rtt time */
7706                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7707                                      __LINE__, RACK_RTTS_NEWRTT);
7708                 /*
7709                  * Only count it if its lower than what we saw within our
7710                  * calculated range.
7711                  */
7712                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7713                         if (rack_probertt_lower_within &&
7714                             rack->rc_gp_dyn_mul &&
7715                             (rack->use_fixed_rate == 0) &&
7716                             (rack->rc_always_pace)) {
7717                                 /*
7718                                  * We are seeing a new lower rtt very close
7719                                  * to the time that we would have entered probe-rtt.
7720                                  * This is probably due to the fact that a peer flow
7721                                  * has entered probe-rtt. Lets go in now too.
7722                                  */
7723                                 uint32_t val;
7724
7725                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7726                                 val /= 100;
7727                                 if ((rack->in_probe_rtt == 0)  &&
7728                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7729                                         rack_enter_probertt(rack, us_cts);
7730                                 }
7731                         }
7732                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7733                 }
7734         }
7735 }
7736
7737 static int
7738 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7739     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7740 {
7741         uint32_t us_rtt;
7742         int32_t i, all;
7743         uint32_t t, len_acked;
7744
7745         if ((rsm->r_flags & RACK_ACKED) ||
7746             (rsm->r_flags & RACK_WAS_ACKED))
7747                 /* Already done */
7748                 return (0);
7749         if (rsm->r_no_rtt_allowed) {
7750                 /* Not allowed */
7751                 return (0);
7752         }
7753         if (ack_type == CUM_ACKED) {
7754                 if (SEQ_GT(th_ack, rsm->r_end)) {
7755                         len_acked = rsm->r_end - rsm->r_start;
7756                         all = 1;
7757                 } else {
7758                         len_acked = th_ack - rsm->r_start;
7759                         all = 0;
7760                 }
7761         } else {
7762                 len_acked = rsm->r_end - rsm->r_start;
7763                 all = 0;
7764         }
7765         if (rsm->r_rtr_cnt == 1) {
7766
7767                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7768                 if ((int)t <= 0)
7769                         t = 1;
7770                 if (!tp->t_rttlow || tp->t_rttlow > t)
7771                         tp->t_rttlow = t;
7772                 if (!rack->r_ctl.rc_rack_min_rtt ||
7773                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7774                         rack->r_ctl.rc_rack_min_rtt = t;
7775                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7776                                 rack->r_ctl.rc_rack_min_rtt = 1;
7777                         }
7778                 }
7779                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7780                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7781                 else
7782                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7783                 if (us_rtt == 0)
7784                         us_rtt = 1;
7785                 if (CC_ALGO(tp)->rttsample != NULL) {
7786                         /* Kick the RTT to the CC */
7787                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7788                 }
7789                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7790                 if (ack_type == SACKED) {
7791                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7792                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7793                 } else {
7794                         /*
7795                          * We need to setup what our confidence
7796                          * is in this ack.
7797                          *
7798                          * If the rsm was app limited and it is
7799                          * less than a mss in length (the end
7800                          * of the send) then we have a gap. If we
7801                          * were app limited but say we were sending
7802                          * multiple MSS's then we are more confident
7803                          * int it.
7804                          *
7805                          * When we are not app-limited then we see if
7806                          * the rsm is being included in the current
7807                          * measurement, we tell this by the app_limited_needs_set
7808                          * flag.
7809                          *
7810                          * Note that being cwnd blocked is not applimited
7811                          * as well as the pacing delay between packets which
7812                          * are sending only 1 or 2 MSS's also will show up
7813                          * in the RTT. We probably need to examine this algorithm
7814                          * a bit more and enhance it to account for the delay
7815                          * between rsm's. We could do that by saving off the
7816                          * pacing delay of each rsm (in an rsm) and then
7817                          * factoring that in somehow though for now I am
7818                          * not sure how :)
7819                          */
7820                         int calc_conf = 0;
7821
7822                         if (rsm->r_flags & RACK_APP_LIMITED) {
7823                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7824                                         calc_conf = 0;
7825                                 else
7826                                         calc_conf = 1;
7827                         } else if (rack->app_limited_needs_set == 0) {
7828                                 calc_conf = 1;
7829                         } else {
7830                                 calc_conf = 0;
7831                         }
7832                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7833                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7834                                             calc_conf, rsm, rsm->r_rtr_cnt);
7835                 }
7836                 if ((rsm->r_flags & RACK_TLP) &&
7837                     (!IN_FASTRECOVERY(tp->t_flags))) {
7838                         /* Segment was a TLP and our retrans matched */
7839                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7840                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
7841                         }
7842                 }
7843                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7844                         /* New more recent rack_tmit_time */
7845                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7846                         rack->rc_rack_rtt = t;
7847                 }
7848                 return (1);
7849         }
7850         /*
7851          * We clear the soft/rxtshift since we got an ack.
7852          * There is no assurance we will call the commit() function
7853          * so we need to clear these to avoid incorrect handling.
7854          */
7855         tp->t_rxtshift = 0;
7856         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7857                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7858         tp->t_softerror = 0;
7859         if (to && (to->to_flags & TOF_TS) &&
7860             (ack_type == CUM_ACKED) &&
7861             (to->to_tsecr) &&
7862             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7863                 /*
7864                  * Now which timestamp does it match? In this block the ACK
7865                  * must be coming from a previous transmission.
7866                  */
7867                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7868                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7869                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7870                                 if ((int)t <= 0)
7871                                         t = 1;
7872                                 if (CC_ALGO(tp)->rttsample != NULL) {
7873                                         /*
7874                                          * Kick the RTT to the CC, here
7875                                          * we lie a bit in that we know the
7876                                          * retransmission is correct even though
7877                                          * we retransmitted. This is because
7878                                          * we match the timestamps.
7879                                          */
7880                                         if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
7881                                                 us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
7882                                         else
7883                                                 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
7884                                         CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7885                                 }
7886                                 if ((i + 1) < rsm->r_rtr_cnt) {
7887                                         /*
7888                                          * The peer ack'd from our previous
7889                                          * transmission. We have a spurious
7890                                          * retransmission and thus we dont
7891                                          * want to update our rack_rtt.
7892                                          *
7893                                          * Hmm should there be a CC revert here?
7894                                          *
7895                                          */
7896                                         return (0);
7897                                 }
7898                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7899                                         tp->t_rttlow = t;
7900                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7901                                         rack->r_ctl.rc_rack_min_rtt = t;
7902                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7903                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7904                                         }
7905                                 }
7906                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7907                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7908                                         /* New more recent rack_tmit_time */
7909                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7910                                         rack->rc_rack_rtt = t;
7911                                 }
7912                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7913                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7914                                                     rsm->r_rtr_cnt);
7915                                 return (1);
7916                         }
7917                 }
7918                 goto ts_not_found;
7919         } else {
7920                 /*
7921                  * Ok its a SACK block that we retransmitted. or a windows
7922                  * machine without timestamps. We can tell nothing from the
7923                  * time-stamp since its not there or the time the peer last
7924                  * recieved a segment that moved forward its cum-ack point.
7925                  */
7926 ts_not_found:
7927                 i = rsm->r_rtr_cnt - 1;
7928                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7929                 if ((int)t <= 0)
7930                         t = 1;
7931                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7932                         /*
7933                          * We retransmitted and the ack came back in less
7934                          * than the smallest rtt we have observed. We most
7935                          * likely did an improper retransmit as outlined in
7936                          * 6.2 Step 2 point 2 in the rack-draft so we
7937                          * don't want to update our rack_rtt. We in
7938                          * theory (in future) might want to think about reverting our
7939                          * cwnd state but we won't for now.
7940                          */
7941                         return (0);
7942                 } else if (rack->r_ctl.rc_rack_min_rtt) {
7943                         /*
7944                          * We retransmitted it and the retransmit did the
7945                          * job.
7946                          */
7947                         if (!rack->r_ctl.rc_rack_min_rtt ||
7948                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7949                                 rack->r_ctl.rc_rack_min_rtt = t;
7950                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7951                                         rack->r_ctl.rc_rack_min_rtt = 1;
7952                                 }
7953                         }
7954                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7955                                 /* New more recent rack_tmit_time */
7956                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7957                                 rack->rc_rack_rtt = t;
7958                         }
7959                         return (1);
7960                 }
7961         }
7962         return (0);
7963 }
7964
7965 /*
7966  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7967  */
7968 static void
7969 rack_log_sack_passed(struct tcpcb *tp,
7970     struct tcp_rack *rack, struct rack_sendmap *rsm)
7971 {
7972         struct rack_sendmap *nrsm;
7973
7974         nrsm = rsm;
7975         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7976             rack_head, r_tnext) {
7977                 if (nrsm == rsm) {
7978                         /* Skip original segment he is acked */
7979                         continue;
7980                 }
7981                 if (nrsm->r_flags & RACK_ACKED) {
7982                         /*
7983                          * Skip ack'd segments, though we
7984                          * should not see these, since tmap
7985                          * should not have ack'd segments.
7986                          */
7987                         continue;
7988                 }
7989                 if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
7990                         /*
7991                          * If the peer dropped the rwnd on
7992                          * these then we don't worry about them.
7993                          */
7994                         continue;
7995                 }
7996                 if (nrsm->r_flags & RACK_SACK_PASSED) {
7997                         /*
7998                          * We found one that is already marked
7999                          * passed, we have been here before and
8000                          * so all others below this are marked.
8001                          */
8002                         break;
8003                 }
8004                 nrsm->r_flags |= RACK_SACK_PASSED;
8005                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8006         }
8007 }
8008
8009 static void
8010 rack_need_set_test(struct tcpcb *tp,
8011                    struct tcp_rack *rack,
8012                    struct rack_sendmap *rsm,
8013                    tcp_seq th_ack,
8014                    int line,
8015                    int use_which)
8016 {
8017
8018         if ((tp->t_flags & TF_GPUTINPROG) &&
8019             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8020                 /*
8021                  * We were app limited, and this ack
8022                  * butts up or goes beyond the point where we want
8023                  * to start our next measurement. We need
8024                  * to record the new gput_ts as here and
8025                  * possibly update the start sequence.
8026                  */
8027                 uint32_t seq, ts;
8028
8029                 if (rsm->r_rtr_cnt > 1) {
8030                         /*
8031                          * This is a retransmit, can we
8032                          * really make any assessment at this
8033                          * point?  We are not really sure of
8034                          * the timestamp, is it this or the
8035                          * previous transmission?
8036                          *
8037                          * Lets wait for something better that
8038                          * is not retransmitted.
8039                          */
8040                         return;
8041                 }
8042                 seq = tp->gput_seq;
8043                 ts = tp->gput_ts;
8044                 rack->app_limited_needs_set = 0;
8045                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8046                 /* Do we start at a new end? */
8047                 if ((use_which == RACK_USE_BEG) &&
8048                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8049                         /*
8050                          * When we get an ACK that just eats
8051                          * up some of the rsm, we set RACK_USE_BEG
8052                          * since whats at r_start (i.e. th_ack)
8053                          * is left unacked and thats where the
8054                          * measurement not starts.
8055                          */
8056                         tp->gput_seq = rsm->r_start;
8057                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8058                 }
8059                 if ((use_which == RACK_USE_END) &&
8060                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8061                             /*
8062                              * We use the end when the cumack
8063                              * is moving forward and completely
8064                              * deleting the rsm passed so basically
8065                              * r_end holds th_ack.
8066                              *
8067                              * For SACK's we also want to use the end
8068                              * since this piece just got sacked and
8069                              * we want to target anything after that
8070                              * in our measurement.
8071                              */
8072                             tp->gput_seq = rsm->r_end;
8073                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8074                 }
8075                 if (use_which == RACK_USE_END_OR_THACK) {
8076                         /*
8077                          * special case for ack moving forward,
8078                          * not a sack, we need to move all the
8079                          * way up to where this ack cum-ack moves
8080                          * to.
8081                          */
8082                         if (SEQ_GT(th_ack, rsm->r_end))
8083                                 tp->gput_seq = th_ack;
8084                         else
8085                                 tp->gput_seq = rsm->r_end;
8086                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8087                 }
8088                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8089                         /*
8090                          * We moved beyond this guy's range, re-calculate
8091                          * the new end point.
8092                          */
8093                         if (rack->rc_gp_filled == 0) {
8094                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8095                         } else {
8096                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8097                         }
8098                 }
8099                 /*
8100                  * We are moving the goal post, we may be able to clear the
8101                  * measure_saw_probe_rtt flag.
8102                  */
8103                 if ((rack->in_probe_rtt == 0) &&
8104                     (rack->measure_saw_probe_rtt) &&
8105                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8106                         rack->measure_saw_probe_rtt = 0;
8107                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8108                                            seq, tp->gput_seq, 0, 5, line, NULL, 0);
8109                 if (rack->rc_gp_filled &&
8110                     ((tp->gput_ack - tp->gput_seq) <
8111                      max(rc_init_window(rack), (MIN_GP_WIN *
8112                                                 ctf_fixed_maxseg(tp))))) {
8113                         uint32_t ideal_amount;
8114
8115                         ideal_amount = rack_get_measure_window(tp, rack);
8116                         if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
8117                                 /*
8118                                  * There is no sense of continuing this measurement
8119                                  * because its too small to gain us anything we
8120                                  * trust. Skip it and that way we can start a new
8121                                  * measurement quicker.
8122                                  */
8123                                 tp->t_flags &= ~TF_GPUTINPROG;
8124                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8125                                                            0, 0, 0, 6, __LINE__, NULL, 0);
8126                         } else {
8127                                 /*
8128                                  * Reset the window further out.
8129                                  */
8130                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8131                         }
8132                 }
8133         }
8134 }
8135
8136 static inline int
8137 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8138 {
8139         if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8140                 /* Behind our TLP definition or right at */
8141                 return (0);
8142         }
8143         if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8144                 /* The start is beyond or right at our end of TLP definition */
8145                 return (0);
8146         }
8147         /* It has to be a sub-part of the original TLP recorded */
8148         return (1);
8149 }
8150
8151
8152 static uint32_t
8153 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8154                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8155 {
8156         uint32_t start, end, changed = 0;
8157         struct rack_sendmap stack_map;
8158         struct rack_sendmap *rsm, *nrsm, fe, *prev, *next;
8159 #ifdef INVARIANTS
8160         struct rack_sendmap *insret;
8161 #endif
8162         int32_t used_ref = 1;
8163         int moved = 0;
8164
8165         start = sack->start;
8166         end = sack->end;
8167         rsm = *prsm;
8168         memset(&fe, 0, sizeof(fe));
8169 do_rest_ofb:
8170         if ((rsm == NULL) ||
8171             (SEQ_LT(end, rsm->r_start)) ||
8172             (SEQ_GEQ(start, rsm->r_end)) ||
8173             (SEQ_LT(start, rsm->r_start))) {
8174                 /*
8175                  * We are not in the right spot,
8176                  * find the correct spot in the tree.
8177                  */
8178                 used_ref = 0;
8179                 fe.r_start = start;
8180                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8181                 moved++;
8182         }
8183         if (rsm == NULL) {
8184                 /* TSNH */
8185                 goto out;
8186         }
8187         /* Ok we have an ACK for some piece of this rsm */
8188         if (rsm->r_start != start) {
8189                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8190                         /*
8191                          * Before any splitting or hookery is
8192                          * done is it a TLP of interest i.e. rxt?
8193                          */
8194                         if ((rsm->r_flags & RACK_TLP) &&
8195                             (rsm->r_rtr_cnt > 1)) {
8196                                 /*
8197                                  * We are splitting a rxt TLP, check
8198                                  * if we need to save off the start/end
8199                                  */
8200                                 if (rack->rc_last_tlp_acked_set &&
8201                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8202                                         /*
8203                                          * We already turned this on since we are inside
8204                                          * the previous one was a partially sack now we
8205                                          * are getting another one (maybe all of it).
8206                                          *
8207                                          */
8208                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8209                                         /*
8210                                          * Lets make sure we have all of it though.
8211                                          */
8212                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8213                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8214                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8215                                                                      rack->r_ctl.last_tlp_acked_end);
8216                                         }
8217                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8218                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8219                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8220                                                                      rack->r_ctl.last_tlp_acked_end);
8221                                         }
8222                                 } else {
8223                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8224                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8225                                         rack->rc_last_tlp_past_cumack = 0;
8226                                         rack->rc_last_tlp_acked_set = 1;
8227                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8228                                 }
8229                         }
8230                         /**
8231                          * Need to split this in two pieces the before and after,
8232                          * the before remains in the map, the after must be
8233                          * added. In other words we have:
8234                          * rsm        |--------------|
8235                          * sackblk        |------->
8236                          * rsm will become
8237                          *     rsm    |---|
8238                          * and nrsm will be  the sacked piece
8239                          *     nrsm       |----------|
8240                          *
8241                          * But before we start down that path lets
8242                          * see if the sack spans over on top of
8243                          * the next guy and it is already sacked.
8244                          *
8245                          */
8246                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8247                         if (next && (next->r_flags & RACK_ACKED) &&
8248                             SEQ_GEQ(end, next->r_start)) {
8249                                 /**
8250                                  * So the next one is already acked, and
8251                                  * we can thus by hookery use our stack_map
8252                                  * to reflect the piece being sacked and
8253                                  * then adjust the two tree entries moving
8254                                  * the start and ends around. So we start like:
8255                                  *  rsm     |------------|             (not-acked)
8256                                  *  next                 |-----------| (acked)
8257                                  *  sackblk        |-------->
8258                                  *  We want to end like so:
8259                                  *  rsm     |------|                   (not-acked)
8260                                  *  next           |-----------------| (acked)
8261                                  *  nrsm           |-----|
8262                                  * Where nrsm is a temporary stack piece we
8263                                  * use to update all the gizmos.
8264                                  */
8265                                 /* Copy up our fudge block */
8266                                 nrsm = &stack_map;
8267                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8268                                 /* Now adjust our tree blocks */
8269                                 rsm->r_end = start;
8270                                 next->r_start = start;
8271                                 /* Now we must adjust back where next->m is */
8272                                 rack_setup_offset_for_rsm(rsm, next);
8273
8274                                 /* We don't need to adjust rsm, it did not change */
8275                                 /* Clear out the dup ack count of the remainder */
8276                                 rsm->r_dupack = 0;
8277                                 rsm->r_just_ret = 0;
8278                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8279                                 /* Now lets make sure our fudge block is right */
8280                                 nrsm->r_start = start;
8281                                 /* Now lets update all the stats and such */
8282                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8283                                 if (rack->app_limited_needs_set)
8284                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8285                                 changed += (nrsm->r_end - nrsm->r_start);
8286                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8287                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8288                                         rack->r_ctl.rc_reorder_ts = cts;
8289                                 }
8290                                 /*
8291                                  * Now we want to go up from rsm (the
8292                                  * one left un-acked) to the next one
8293                                  * in the tmap. We do this so when
8294                                  * we walk backwards we include marking
8295                                  * sack-passed on rsm (The one passed in
8296                                  * is skipped since it is generally called
8297                                  * on something sacked before removing it
8298                                  * from the tmap).
8299                                  */
8300                                 if (rsm->r_in_tmap) {
8301                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8302                                         /*
8303                                          * Now that we have the next
8304                                          * one walk backwards from there.
8305                                          */
8306                                         if (nrsm && nrsm->r_in_tmap)
8307                                                 rack_log_sack_passed(tp, rack, nrsm);
8308                                 }
8309                                 /* Now are we done? */
8310                                 if (SEQ_LT(end, next->r_end) ||
8311                                     (end == next->r_end)) {
8312                                         /* Done with block */
8313                                         goto out;
8314                                 }
8315                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8316                                 counter_u64_add(rack_sack_used_next_merge, 1);
8317                                 /* Postion for the next block */
8318                                 start = next->r_end;
8319                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8320                                 if (rsm == NULL)
8321                                         goto out;
8322                         } else {
8323                                 /**
8324                                  * We can't use any hookery here, so we
8325                                  * need to split the map. We enter like
8326                                  * so:
8327                                  *  rsm      |--------|
8328                                  *  sackblk       |----->
8329                                  * We will add the new block nrsm and
8330                                  * that will be the new portion, and then
8331                                  * fall through after reseting rsm. So we
8332                                  * split and look like this:
8333                                  *  rsm      |----|
8334                                  *  sackblk       |----->
8335                                  *  nrsm          |---|
8336                                  * We then fall through reseting
8337                                  * rsm to nrsm, so the next block
8338                                  * picks it up.
8339                                  */
8340                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8341                                 if (nrsm == NULL) {
8342                                         /*
8343                                          * failed XXXrrs what can we do but loose the sack
8344                                          * info?
8345                                          */
8346                                         goto out;
8347                                 }
8348                                 counter_u64_add(rack_sack_splits, 1);
8349                                 rack_clone_rsm(rack, nrsm, rsm, start);
8350                                 rsm->r_just_ret = 0;
8351 #ifndef INVARIANTS
8352                                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8353 #else
8354                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8355                                 if (insret != NULL) {
8356                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8357                                               nrsm, insret, rack, rsm);
8358                                 }
8359 #endif
8360                                 if (rsm->r_in_tmap) {
8361                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8362                                         nrsm->r_in_tmap = 1;
8363                                 }
8364                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8365                                 rsm->r_flags &= (~RACK_HAS_FIN);
8366                                 /* Position us to point to the new nrsm that starts the sack blk */
8367                                 rsm = nrsm;
8368                         }
8369                 } else {
8370                         /* Already sacked this piece */
8371                         counter_u64_add(rack_sack_skipped_acked, 1);
8372                         moved++;
8373                         if (end == rsm->r_end) {
8374                                 /* Done with block */
8375                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8376                                 goto out;
8377                         } else if (SEQ_LT(end, rsm->r_end)) {
8378                                 /* A partial sack to a already sacked block */
8379                                 moved++;
8380                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8381                                 goto out;
8382                         } else {
8383                                 /*
8384                                  * The end goes beyond this guy
8385                                  * reposition the start to the
8386                                  * next block.
8387                                  */
8388                                 start = rsm->r_end;
8389                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8390                                 if (rsm == NULL)
8391                                         goto out;
8392                         }
8393                 }
8394         }
8395         if (SEQ_GEQ(end, rsm->r_end)) {
8396                 /**
8397                  * The end of this block is either beyond this guy or right
8398                  * at this guy. I.e.:
8399                  *  rsm ---                 |-----|
8400                  *  end                     |-----|
8401                  *  <or>
8402                  *  end                     |---------|
8403                  */
8404                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8405                         /*
8406                          * Is it a TLP of interest?
8407                          */
8408                         if ((rsm->r_flags & RACK_TLP) &&
8409                             (rsm->r_rtr_cnt > 1)) {
8410                                 /*
8411                                  * We are splitting a rxt TLP, check
8412                                  * if we need to save off the start/end
8413                                  */
8414                                 if (rack->rc_last_tlp_acked_set &&
8415                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8416                                         /*
8417                                          * We already turned this on since we are inside
8418                                          * the previous one was a partially sack now we
8419                                          * are getting another one (maybe all of it).
8420                                          */
8421                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8422                                         /*
8423                                          * Lets make sure we have all of it though.
8424                                          */
8425                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8426                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8427                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8428                                                                      rack->r_ctl.last_tlp_acked_end);
8429                                         }
8430                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8431                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8432                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8433                                                                      rack->r_ctl.last_tlp_acked_end);
8434                                         }
8435                                 } else {
8436                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8437                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8438                                         rack->rc_last_tlp_past_cumack = 0;
8439                                         rack->rc_last_tlp_acked_set = 1;
8440                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8441                                 }
8442                         }
8443                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8444                         changed += (rsm->r_end - rsm->r_start);
8445                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8446                         if (rsm->r_in_tmap) /* should be true */
8447                                 rack_log_sack_passed(tp, rack, rsm);
8448                         /* Is Reordering occuring? */
8449                         if (rsm->r_flags & RACK_SACK_PASSED) {
8450                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8451                                 rack->r_ctl.rc_reorder_ts = cts;
8452                         }
8453                         if (rack->app_limited_needs_set)
8454                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8455                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8456                         rsm->r_flags |= RACK_ACKED;
8457                         if (rsm->r_in_tmap) {
8458                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8459                                 rsm->r_in_tmap = 0;
8460                         }
8461                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8462                 } else {
8463                         counter_u64_add(rack_sack_skipped_acked, 1);
8464                         moved++;
8465                 }
8466                 if (end == rsm->r_end) {
8467                         /* This block only - done, setup for next */
8468                         goto out;
8469                 }
8470                 /*
8471                  * There is more not coverend by this rsm move on
8472                  * to the next block in the RB tree.
8473                  */
8474                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8475                 start = rsm->r_end;
8476                 rsm = nrsm;
8477                 if (rsm == NULL)
8478                         goto out;
8479                 goto do_rest_ofb;
8480         }
8481         /**
8482          * The end of this sack block is smaller than
8483          * our rsm i.e.:
8484          *  rsm ---                 |-----|
8485          *  end                     |--|
8486          */
8487         if ((rsm->r_flags & RACK_ACKED) == 0) {
8488                 /*
8489                  * Is it a TLP of interest?
8490                  */
8491                 if ((rsm->r_flags & RACK_TLP) &&
8492                     (rsm->r_rtr_cnt > 1)) {
8493                         /*
8494                          * We are splitting a rxt TLP, check
8495                          * if we need to save off the start/end
8496                          */
8497                         if (rack->rc_last_tlp_acked_set &&
8498                             (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8499                                 /*
8500                                  * We already turned this on since we are inside
8501                                  * the previous one was a partially sack now we
8502                                  * are getting another one (maybe all of it).
8503                                  */
8504                                 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8505                                 /*
8506                                  * Lets make sure we have all of it though.
8507                                  */
8508                                 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8509                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8510                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8511                                                              rack->r_ctl.last_tlp_acked_end);
8512                                 }
8513                                 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8514                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8515                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8516                                                              rack->r_ctl.last_tlp_acked_end);
8517                                 }
8518                         } else {
8519                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8520                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8521                                 rack->rc_last_tlp_past_cumack = 0;
8522                                 rack->rc_last_tlp_acked_set = 1;
8523                                 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8524                         }
8525                 }
8526                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8527                 if (prev &&
8528                     (prev->r_flags & RACK_ACKED)) {
8529                         /**
8530                          * Goal, we want the right remainder of rsm to shrink
8531                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8532                          * We want to expand prev to go all the way
8533                          * to prev->r_end <- end.
8534                          * so in the tree we have before:
8535                          *   prev     |--------|         (acked)
8536                          *   rsm               |-------| (non-acked)
8537                          *   sackblk           |-|
8538                          * We churn it so we end up with
8539                          *   prev     |----------|       (acked)
8540                          *   rsm                 |-----| (non-acked)
8541                          *   nrsm              |-| (temporary)
8542                          *
8543                          * Note if either prev/rsm is a TLP we don't
8544                          * do this.
8545                          */
8546                         nrsm = &stack_map;
8547                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8548                         prev->r_end = end;
8549                         rsm->r_start = end;
8550                         /* Now adjust nrsm (stack copy) to be
8551                          * the one that is the small
8552                          * piece that was "sacked".
8553                          */
8554                         nrsm->r_end = end;
8555                         rsm->r_dupack = 0;
8556                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8557                         /*
8558                          * Now that the rsm has had its start moved forward
8559                          * lets go ahead and get its new place in the world.
8560                          */
8561                         rack_setup_offset_for_rsm(prev, rsm);
8562                         /*
8563                          * Now nrsm is our new little piece
8564                          * that is acked (which was merged
8565                          * to prev). Update the rtt and changed
8566                          * based on that. Also check for reordering.
8567                          */
8568                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8569                         if (rack->app_limited_needs_set)
8570                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8571                         changed += (nrsm->r_end - nrsm->r_start);
8572                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8573                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8574                                 rack->r_ctl.rc_reorder_ts = cts;
8575                         }
8576                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8577                         rsm = prev;
8578                         counter_u64_add(rack_sack_used_prev_merge, 1);
8579                 } else {
8580                         /**
8581                          * This is the case where our previous
8582                          * block is not acked either, so we must
8583                          * split the block in two.
8584                          */
8585                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8586                         if (nrsm == NULL) {
8587                                 /* failed rrs what can we do but loose the sack info? */
8588                                 goto out;
8589                         }
8590                         if ((rsm->r_flags & RACK_TLP) &&
8591                             (rsm->r_rtr_cnt > 1)) {
8592                                 /*
8593                                  * We are splitting a rxt TLP, check
8594                                  * if we need to save off the start/end
8595                                  */
8596                                 if (rack->rc_last_tlp_acked_set &&
8597                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8598                                             /*
8599                                              * We already turned this on since this block is inside
8600                                              * the previous one was a partially sack now we
8601                                              * are getting another one (maybe all of it).
8602                                              */
8603                                             rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8604                                             /*
8605                                              * Lets make sure we have all of it though.
8606                                              */
8607                                             if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8608                                                     rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8609                                                     rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8610                                                                          rack->r_ctl.last_tlp_acked_end);
8611                                             }
8612                                             if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8613                                                     rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8614                                                     rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8615                                                                          rack->r_ctl.last_tlp_acked_end);
8616                                             }
8617                                     } else {
8618                                             rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8619                                             rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8620                                             rack->rc_last_tlp_acked_set = 1;
8621                                             rack->rc_last_tlp_past_cumack = 0;
8622                                             rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8623                                     }
8624                         }
8625                         /**
8626                          * In this case nrsm becomes
8627                          * nrsm->r_start = end;
8628                          * nrsm->r_end = rsm->r_end;
8629                          * which is un-acked.
8630                          * <and>
8631                          * rsm->r_end = nrsm->r_start;
8632                          * i.e. the remaining un-acked
8633                          * piece is left on the left
8634                          * hand side.
8635                          *
8636                          * So we start like this
8637                          * rsm      |----------| (not acked)
8638                          * sackblk  |---|
8639                          * build it so we have
8640                          * rsm      |---|         (acked)
8641                          * nrsm         |------|  (not acked)
8642                          */
8643                         counter_u64_add(rack_sack_splits, 1);
8644                         rack_clone_rsm(rack, nrsm, rsm, end);
8645                         rsm->r_flags &= (~RACK_HAS_FIN);
8646                         rsm->r_just_ret = 0;
8647 #ifndef INVARIANTS
8648                         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8649 #else
8650                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8651                         if (insret != NULL) {
8652                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8653                                       nrsm, insret, rack, rsm);
8654                         }
8655 #endif
8656                         if (rsm->r_in_tmap) {
8657                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8658                                 nrsm->r_in_tmap = 1;
8659                         }
8660                         nrsm->r_dupack = 0;
8661                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8662                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8663                         changed += (rsm->r_end - rsm->r_start);
8664                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8665                         if (rsm->r_in_tmap) /* should be true */
8666                                 rack_log_sack_passed(tp, rack, rsm);
8667                         /* Is Reordering occuring? */
8668                         if (rsm->r_flags & RACK_SACK_PASSED) {
8669                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8670                                 rack->r_ctl.rc_reorder_ts = cts;
8671                         }
8672                         if (rack->app_limited_needs_set)
8673                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8674                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8675                         rsm->r_flags |= RACK_ACKED;
8676                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8677                         if (rsm->r_in_tmap) {
8678                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8679                                 rsm->r_in_tmap = 0;
8680                         }
8681                 }
8682         } else if (start != end){
8683                 /*
8684                  * The block was already acked.
8685                  */
8686                 counter_u64_add(rack_sack_skipped_acked, 1);
8687                 moved++;
8688         }
8689 out:
8690         if (rsm &&
8691             ((rsm->r_flags & RACK_TLP) == 0) &&
8692             (rsm->r_flags & RACK_ACKED)) {
8693                 /*
8694                  * Now can we merge where we worked
8695                  * with either the previous or
8696                  * next block?
8697                  */
8698                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8699                 while (next) {
8700                         if (next->r_flags & RACK_TLP)
8701                                 break;
8702                         if (next->r_flags & RACK_ACKED) {
8703                         /* yep this and next can be merged */
8704                                 rsm = rack_merge_rsm(rack, rsm, next);
8705                                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8706                         } else
8707                                 break;
8708                 }
8709                 /* Now what about the previous? */
8710                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8711                 while (prev) {
8712                         if (prev->r_flags & RACK_TLP)
8713                                 break;
8714                         if (prev->r_flags & RACK_ACKED) {
8715                                 /* yep the previous and this can be merged */
8716                                 rsm = rack_merge_rsm(rack, prev, rsm);
8717                                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8718                         } else
8719                                 break;
8720                 }
8721         }
8722         if (used_ref == 0) {
8723                 counter_u64_add(rack_sack_proc_all, 1);
8724         } else {
8725                 counter_u64_add(rack_sack_proc_short, 1);
8726         }
8727         /* Save off the next one for quick reference. */
8728         if (rsm)
8729                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8730         else
8731                 nrsm = NULL;
8732         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8733         /* Pass back the moved. */
8734         *moved_two = moved;
8735         return (changed);
8736 }
8737
8738 static void inline
8739 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8740 {
8741         struct rack_sendmap *tmap;
8742
8743         tmap = NULL;
8744         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8745                 /* Its no longer sacked, mark it so */
8746                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8747 #ifdef INVARIANTS
8748                 if (rsm->r_in_tmap) {
8749                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8750                               rack, rsm, rsm->r_flags);
8751                 }
8752 #endif
8753                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8754                 /* Rebuild it into our tmap */
8755                 if (tmap == NULL) {
8756                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8757                         tmap = rsm;
8758                 } else {
8759                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8760                         tmap = rsm;
8761                 }
8762                 tmap->r_in_tmap = 1;
8763                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8764         }
8765         /*
8766          * Now lets possibly clear the sack filter so we start
8767          * recognizing sacks that cover this area.
8768          */
8769         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8770
8771 }
8772
8773 static void
8774 rack_do_decay(struct tcp_rack *rack)
8775 {
8776         struct timeval res;
8777
8778 #define timersub(tvp, uvp, vvp)                                         \
8779         do {                                                            \
8780                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8781                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8782                 if ((vvp)->tv_usec < 0) {                               \
8783                         (vvp)->tv_sec--;                                \
8784                         (vvp)->tv_usec += 1000000;                      \
8785                 }                                                       \
8786         } while (0)
8787
8788         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8789 #undef timersub
8790
8791         rack->r_ctl.input_pkt++;
8792         if ((rack->rc_in_persist) ||
8793             (res.tv_sec >= 1) ||
8794             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8795                 /*
8796                  * Check for decay of non-SAD,
8797                  * we want all SAD detection metrics to
8798                  * decay 1/4 per second (or more) passed.
8799                  */
8800 #ifdef NETFLIX_EXP_DETECTION
8801                 uint32_t pkt_delta;
8802
8803                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8804 #endif
8805                 /* Update our saved tracking values */
8806                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8807                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8808                 /* Now do we escape without decay? */
8809 #ifdef NETFLIX_EXP_DETECTION
8810                 if (rack->rc_in_persist ||
8811                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8812                     (pkt_delta < tcp_sad_low_pps)){
8813                         /*
8814                          * We don't decay idle connections
8815                          * or ones that have a low input pps.
8816                          */
8817                         return;
8818                 }
8819                 /* Decay the counters */
8820                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8821                                                         tcp_sad_decay_val);
8822                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8823                                                          tcp_sad_decay_val);
8824                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8825                                                                tcp_sad_decay_val);
8826                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8827                                                                 tcp_sad_decay_val);
8828 #endif
8829         }
8830 }
8831
8832 static void
8833 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8834 {
8835         struct rack_sendmap *rsm;
8836 #ifdef INVARIANTS
8837         struct rack_sendmap *rm;
8838 #endif
8839
8840         /*
8841          * The ACK point is advancing to th_ack, we must drop off
8842          * the packets in the rack log and calculate any eligble
8843          * RTT's.
8844          */
8845         rack->r_wanted_output = 1;
8846
8847         /* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
8848         if ((rack->rc_last_tlp_acked_set == 1)&&
8849             (rack->rc_last_tlp_past_cumack == 1) &&
8850             (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
8851                 /*
8852                  * We have reached the point where our last rack
8853                  * tlp retransmit sequence is ahead of the cum-ack.
8854                  * This can only happen when the cum-ack moves all
8855                  * the way around (its been a full 2^^31+1 bytes
8856                  * or more since we sent a retransmitted TLP). Lets
8857                  * turn off the valid flag since its not really valid.
8858                  *
8859                  * Note since sack's also turn on this event we have
8860                  * a complication, we have to wait to age it out until
8861                  * the cum-ack is by the TLP before checking which is
8862                  * what the next else clause does.
8863                  */
8864                 rack_log_dsack_event(rack, 9, __LINE__,
8865                                      rack->r_ctl.last_tlp_acked_start,
8866                                      rack->r_ctl.last_tlp_acked_end);
8867                 rack->rc_last_tlp_acked_set = 0;
8868                 rack->rc_last_tlp_past_cumack = 0;
8869         } else if ((rack->rc_last_tlp_acked_set == 1) &&
8870                    (rack->rc_last_tlp_past_cumack == 0) &&
8871                    (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
8872                 /*
8873                  * It is safe to start aging TLP's out.
8874                  */
8875                 rack->rc_last_tlp_past_cumack = 1;
8876         }
8877         /* We do the same for the tlp send seq as well */
8878         if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8879             (rack->rc_last_sent_tlp_past_cumack == 1) &&
8880             (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
8881                 rack_log_dsack_event(rack, 9, __LINE__,
8882                                      rack->r_ctl.last_sent_tlp_seq,
8883                                      (rack->r_ctl.last_sent_tlp_seq +
8884                                       rack->r_ctl.last_sent_tlp_len));
8885                 rack->rc_last_sent_tlp_seq_valid = 0;
8886                 rack->rc_last_sent_tlp_past_cumack = 0;
8887         } else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8888                    (rack->rc_last_sent_tlp_past_cumack == 0) &&
8889                    (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
8890                 /*
8891                  * It is safe to start aging TLP's send.
8892                  */
8893                 rack->rc_last_sent_tlp_past_cumack = 1;
8894         }
8895 more:
8896         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8897         if (rsm == NULL) {
8898                 if ((th_ack - 1) == tp->iss) {
8899                         /*
8900                          * For the SYN incoming case we will not
8901                          * have called tcp_output for the sending of
8902                          * the SYN, so there will be no map. All
8903                          * other cases should probably be a panic.
8904                          */
8905                         return;
8906                 }
8907                 if (tp->t_flags & TF_SENTFIN) {
8908                         /* if we sent a FIN we often will not have map */
8909                         return;
8910                 }
8911 #ifdef INVARIANTS
8912                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8913                       tp,
8914                       tp->t_state, th_ack, rack,
8915                       tp->snd_una, tp->snd_max, tp->snd_nxt);
8916 #endif
8917                 return;
8918         }
8919         if (SEQ_LT(th_ack, rsm->r_start)) {
8920                 /* Huh map is missing this */
8921 #ifdef INVARIANTS
8922                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8923                        rsm->r_start,
8924                        th_ack, tp->t_state, rack->r_state);
8925 #endif
8926                 return;
8927         }
8928         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8929
8930         /* Now was it a retransmitted TLP? */
8931         if ((rsm->r_flags & RACK_TLP) &&
8932             (rsm->r_rtr_cnt > 1)) {
8933                 /*
8934                  * Yes, this rsm was a TLP and retransmitted, remember that
8935                  * since if a DSACK comes back on this we don't want
8936                  * to think of it as a reordered segment. This may
8937                  * get updated again with possibly even other TLPs
8938                  * in flight, but thats ok. Only when we don't send
8939                  * a retransmitted TLP for 1/2 the sequences space
8940                  * will it get turned off (above).
8941                  */
8942                 if (rack->rc_last_tlp_acked_set &&
8943                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8944                         /*
8945                          * We already turned this on since the end matches,
8946                          * the previous one was a partially ack now we
8947                          * are getting another one (maybe all of it).
8948                          */
8949                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8950                         /*
8951                          * Lets make sure we have all of it though.
8952                          */
8953                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8954                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8955                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8956                                                      rack->r_ctl.last_tlp_acked_end);
8957                         }
8958                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8959                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8960                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8961                                                      rack->r_ctl.last_tlp_acked_end);
8962                         }
8963                 } else {
8964                         rack->rc_last_tlp_past_cumack = 1;
8965                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8966                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8967                         rack->rc_last_tlp_acked_set = 1;
8968                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8969                 }
8970         }
8971         /* Now do we consume the whole thing? */
8972         if (SEQ_GEQ(th_ack, rsm->r_end)) {
8973                 /* Its all consumed. */
8974                 uint32_t left;
8975                 uint8_t newly_acked;
8976
8977                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8978                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8979                 rsm->r_rtr_bytes = 0;
8980                 /* Record the time of highest cumack sent */
8981                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8982 #ifndef INVARIANTS
8983                 (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8984 #else
8985                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8986                 if (rm != rsm) {
8987                         panic("removing head in rack:%p rsm:%p rm:%p",
8988                               rack, rsm, rm);
8989                 }
8990 #endif
8991                 if (rsm->r_in_tmap) {
8992                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8993                         rsm->r_in_tmap = 0;
8994                 }
8995                 newly_acked = 1;
8996                 if (rsm->r_flags & RACK_ACKED) {
8997                         /*
8998                          * It was acked on the scoreboard -- remove
8999                          * it from total
9000                          */
9001                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
9002                         newly_acked = 0;
9003                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
9004                         /*
9005                          * There are segments ACKED on the
9006                          * scoreboard further up. We are seeing
9007                          * reordering.
9008                          */
9009                         rsm->r_flags &= ~RACK_SACK_PASSED;
9010                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9011                         rsm->r_flags |= RACK_ACKED;
9012                         rack->r_ctl.rc_reorder_ts = cts;
9013                         if (rack->r_ent_rec_ns) {
9014                                 /*
9015                                  * We have sent no more, and we saw an sack
9016                                  * then ack arrive.
9017                                  */
9018                                 rack->r_might_revert = 1;
9019                         }
9020                 }
9021                 if ((rsm->r_flags & RACK_TO_REXT) &&
9022                     (tp->t_flags & TF_RCVD_TSTMP) &&
9023                     (to->to_flags & TOF_TS) &&
9024                     (to->to_tsecr != 0) &&
9025                     (tp->t_flags & TF_PREVVALID)) {
9026                         /*
9027                          * We can use the timestamp to see
9028                          * if this retransmission was from the
9029                          * first transmit. If so we made a mistake.
9030                          */
9031                         tp->t_flags &= ~TF_PREVVALID;
9032                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9033                                 /* The first transmit is what this ack is for */
9034                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
9035                         }
9036                 }
9037                 left = th_ack - rsm->r_end;
9038                 if (rack->app_limited_needs_set && newly_acked)
9039                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9040                 /* Free back to zone */
9041                 rack_free(rack, rsm);
9042                 if (left) {
9043                         goto more;
9044                 }
9045                 /* Check for reneging */
9046                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9047                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9048                         /*
9049                          * The peer has moved snd_una up to
9050                          * the edge of this send, i.e. one
9051                          * that it had previously acked. The only
9052                          * way that can be true if the peer threw
9053                          * away data (space issues) that it had
9054                          * previously sacked (else it would have
9055                          * given us snd_una up to (rsm->r_end).
9056                          * We need to undo the acked markings here.
9057                          *
9058                          * Note we have to look to make sure th_ack is
9059                          * our rsm->r_start in case we get an old ack
9060                          * where th_ack is behind snd_una.
9061                          */
9062                         rack_peer_reneges(rack, rsm, th_ack);
9063                 }
9064                 return;
9065         }
9066         if (rsm->r_flags & RACK_ACKED) {
9067                 /*
9068                  * It was acked on the scoreboard -- remove it from
9069                  * total for the part being cum-acked.
9070                  */
9071                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9072         }
9073         /*
9074          * Clear the dup ack count for
9075          * the piece that remains.
9076          */
9077         rsm->r_dupack = 0;
9078         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9079         if (rsm->r_rtr_bytes) {
9080                 /*
9081                  * It was retransmitted adjust the
9082                  * sack holes for what was acked.
9083                  */
9084                 int ack_am;
9085
9086                 ack_am = (th_ack - rsm->r_start);
9087                 if (ack_am >= rsm->r_rtr_bytes) {
9088                         rack->r_ctl.rc_holes_rxt -= ack_am;
9089                         rsm->r_rtr_bytes -= ack_am;
9090                 }
9091         }
9092         /*
9093          * Update where the piece starts and record
9094          * the time of send of highest cumack sent.
9095          */
9096         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9097         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9098         /* Now we need to move our offset forward too */
9099         if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9100                 /* Fix up the orig_m_len and possibly the mbuf offset */
9101                 rack_adjust_orig_mlen(rsm);
9102         }
9103         rsm->soff += (th_ack - rsm->r_start);
9104         rsm->r_start = th_ack;
9105         /* Now do we need to move the mbuf fwd too? */
9106         if (rsm->m) {
9107                 while (rsm->soff >= rsm->m->m_len) {
9108                         rsm->soff -= rsm->m->m_len;
9109                         rsm->m = rsm->m->m_next;
9110                         KASSERT((rsm->m != NULL),
9111                                 (" nrsm:%p hit at soff:%u null m",
9112                                  rsm, rsm->soff));
9113                 }
9114                 rsm->orig_m_len = rsm->m->m_len;
9115         }
9116         if (rack->app_limited_needs_set)
9117                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9118 }
9119
9120 static void
9121 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9122 {
9123         struct rack_sendmap *rsm;
9124         int sack_pass_fnd = 0;
9125
9126         if (rack->r_might_revert) {
9127                 /*
9128                  * Ok we have reordering, have not sent anything, we
9129                  * might want to revert the congestion state if nothing
9130                  * further has SACK_PASSED on it. Lets check.
9131                  *
9132                  * We also get here when we have DSACKs come in for
9133                  * all the data that we FR'd. Note that a rxt or tlp
9134                  * timer clears this from happening.
9135                  */
9136
9137                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9138                         if (rsm->r_flags & RACK_SACK_PASSED) {
9139                                 sack_pass_fnd = 1;
9140                                 break;
9141                         }
9142                 }
9143                 if (sack_pass_fnd == 0) {
9144                         /*
9145                          * We went into recovery
9146                          * incorrectly due to reordering!
9147                          */
9148                         int orig_cwnd;
9149
9150                         rack->r_ent_rec_ns = 0;
9151                         orig_cwnd = tp->snd_cwnd;
9152                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9153                         tp->snd_recover = tp->snd_una;
9154                         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
9155                         EXIT_RECOVERY(tp->t_flags);
9156                 }
9157                 rack->r_might_revert = 0;
9158         }
9159 }
9160
9161 #ifdef NETFLIX_EXP_DETECTION
9162 static void
9163 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9164 {
9165         if ((rack->do_detection || tcp_force_detection) &&
9166             tcp_sack_to_ack_thresh &&
9167             tcp_sack_to_move_thresh &&
9168             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9169                 /*
9170                  * We have thresholds set to find
9171                  * possible attackers and disable sack.
9172                  * Check them.
9173                  */
9174                 uint64_t ackratio, moveratio, movetotal;
9175
9176                 /* Log detecting */
9177                 rack_log_sad(rack, 1);
9178                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
9179                 ackratio *= (uint64_t)(1000);
9180                 if (rack->r_ctl.ack_count)
9181                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9182                 else {
9183                         /* We really should not hit here */
9184                         ackratio = 1000;
9185                 }
9186                 if ((rack->sack_attack_disable == 0) &&
9187                     (ackratio > rack_highest_sack_thresh_seen))
9188                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9189                 movetotal = rack->r_ctl.sack_moved_extra;
9190                 movetotal += rack->r_ctl.sack_noextra_move;
9191                 moveratio = rack->r_ctl.sack_moved_extra;
9192                 moveratio *= (uint64_t)1000;
9193                 if (movetotal)
9194                         moveratio /= movetotal;
9195                 else {
9196                         /* No moves, thats pretty good */
9197                         moveratio = 0;
9198                 }
9199                 if ((rack->sack_attack_disable == 0) &&
9200                     (moveratio > rack_highest_move_thresh_seen))
9201                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
9202                 if (rack->sack_attack_disable == 0) {
9203                         if ((ackratio > tcp_sack_to_ack_thresh) &&
9204                             (moveratio > tcp_sack_to_move_thresh)) {
9205                                 /* Disable sack processing */
9206                                 rack->sack_attack_disable = 1;
9207                                 if (rack->r_rep_attack == 0) {
9208                                         rack->r_rep_attack = 1;
9209                                         counter_u64_add(rack_sack_attacks_detected, 1);
9210                                 }
9211                                 if (tcp_attack_on_turns_on_logging) {
9212                                         /*
9213                                          * Turn on logging, used for debugging
9214                                          * false positives.
9215                                          */
9216                                         rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9217                                 }
9218                                 /* Clamp the cwnd at flight size */
9219                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9220                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9221                                 rack_log_sad(rack, 2);
9222                         }
9223                 } else {
9224                         /* We are sack-disabled check for false positives */
9225                         if ((ackratio <= tcp_restoral_thresh) ||
9226                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9227                                 rack->sack_attack_disable = 0;
9228                                 rack_log_sad(rack, 3);
9229                                 /* Restart counting */
9230                                 rack->r_ctl.sack_count = 0;
9231                                 rack->r_ctl.sack_moved_extra = 0;
9232                                 rack->r_ctl.sack_noextra_move = 1;
9233                                 rack->r_ctl.ack_count = max(1,
9234                                       (bytes_this_ack / segsiz));
9235
9236                                 if (rack->r_rep_reverse == 0) {
9237                                         rack->r_rep_reverse = 1;
9238                                         counter_u64_add(rack_sack_attacks_reversed, 1);
9239                                 }
9240                                 /* Restore the cwnd */
9241                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9242                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9243                         }
9244                 }
9245         }
9246 }
9247 #endif
9248
9249 static int
9250 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9251 {
9252
9253         uint32_t am, l_end;
9254         int was_tlp = 0;
9255
9256         if (SEQ_GT(end, start))
9257                 am = end - start;
9258         else
9259                 am = 0;
9260         if ((rack->rc_last_tlp_acked_set ) &&
9261             (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9262             (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9263                 /*
9264                  * The DSACK is because of a TLP which we don't
9265                  * do anything with the reordering window over since
9266                  * it was not reordering that caused the DSACK but
9267                  * our previous retransmit TLP.
9268                  */
9269                 rack_log_dsack_event(rack, 7, __LINE__, start, end);
9270                 was_tlp = 1;
9271                 goto skip_dsack_round;
9272         }
9273         if (rack->rc_last_sent_tlp_seq_valid) {
9274                 l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9275                 if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9276                     (SEQ_LEQ(end, l_end))) {
9277                         /*
9278                          * This dsack is from the last sent TLP, ignore it
9279                          * for reordering purposes.
9280                          */
9281                         rack_log_dsack_event(rack, 7, __LINE__, start, end);
9282                         was_tlp = 1;
9283                         goto skip_dsack_round;
9284                 }
9285         }
9286         if (rack->rc_dsack_round_seen == 0) {
9287                 rack->rc_dsack_round_seen = 1;
9288                 rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9289                 rack->r_ctl.num_dsack++;
9290                 rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */
9291                 rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9292         }
9293 skip_dsack_round:
9294         /*
9295          * We keep track of how many DSACK blocks we get
9296          * after a recovery incident.
9297          */
9298         rack->r_ctl.dsack_byte_cnt += am;
9299         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9300             rack->r_ctl.retran_during_recovery &&
9301             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9302                 /*
9303                  * False recovery most likely culprit is reordering. If
9304                  * nothing else is missing we need to revert.
9305                  */
9306                 rack->r_might_revert = 1;
9307                 rack_handle_might_revert(rack->rc_tp, rack);
9308                 rack->r_might_revert = 0;
9309                 rack->r_ctl.retran_during_recovery = 0;
9310                 rack->r_ctl.dsack_byte_cnt = 0;
9311         }
9312         return (was_tlp);
9313 }
9314
9315 static uint32_t
9316 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
9317 {
9318         return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
9319 }
9320
9321 static int32_t
9322 rack_compute_pipe(struct tcpcb *tp)
9323 {
9324         return ((int32_t)do_rack_compute_pipe(tp,
9325                                               (struct tcp_rack *)tp->t_fb_ptr,
9326                                               tp->snd_una));
9327 }
9328
9329 static void
9330 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9331 {
9332         /* Deal with changed and PRR here (in recovery only) */
9333         uint32_t pipe, snd_una;
9334
9335         rack->r_ctl.rc_prr_delivered += changed;
9336
9337         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9338                 /*
9339                  * It is all outstanding, we are application limited
9340                  * and thus we don't need more room to send anything.
9341                  * Note we use tp->snd_una here and not th_ack because
9342                  * the data as yet not been cut from the sb.
9343                  */
9344                 rack->r_ctl.rc_prr_sndcnt = 0;
9345                 return;
9346         }
9347         /* Compute prr_sndcnt */
9348         if (SEQ_GT(tp->snd_una, th_ack)) {
9349                 snd_una = tp->snd_una;
9350         } else {
9351                 snd_una = th_ack;
9352         }
9353         pipe = do_rack_compute_pipe(tp, rack, snd_una);
9354         if (pipe > tp->snd_ssthresh) {
9355                 long sndcnt;
9356
9357                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9358                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9359                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9360                 else {
9361                         rack->r_ctl.rc_prr_sndcnt = 0;
9362                         rack_log_to_prr(rack, 9, 0, __LINE__);
9363                         sndcnt = 0;
9364                 }
9365                 sndcnt++;
9366                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9367                         sndcnt -= rack->r_ctl.rc_prr_out;
9368                 else
9369                         sndcnt = 0;
9370                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9371                 rack_log_to_prr(rack, 10, 0, __LINE__);
9372         } else {
9373                 uint32_t limit;
9374
9375                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9376                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9377                 else
9378                         limit = 0;
9379                 if (changed > limit)
9380                         limit = changed;
9381                 limit += ctf_fixed_maxseg(tp);
9382                 if (tp->snd_ssthresh > pipe) {
9383                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9384                         rack_log_to_prr(rack, 11, 0, __LINE__);
9385                 } else {
9386                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9387                         rack_log_to_prr(rack, 12, 0, __LINE__);
9388                 }
9389         }
9390 }
9391
9392 static void
9393 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9394 {
9395         uint32_t changed;
9396         struct tcp_rack *rack;
9397         struct rack_sendmap *rsm;
9398         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9399         register uint32_t th_ack;
9400         int32_t i, j, k, num_sack_blks = 0;
9401         uint32_t cts, acked, ack_point;
9402         int loop_start = 0, moved_two = 0;
9403         uint32_t tsused;
9404
9405
9406         INP_WLOCK_ASSERT(tptoinpcb(tp));
9407         if (tcp_get_flags(th) & TH_RST) {
9408                 /* We don't log resets */
9409                 return;
9410         }
9411         rack = (struct tcp_rack *)tp->t_fb_ptr;
9412         cts = tcp_get_usecs(NULL);
9413         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9414         changed = 0;
9415         th_ack = th->th_ack;
9416         if (rack->sack_attack_disable == 0)
9417                 rack_do_decay(rack);
9418         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9419                 /*
9420                  * You only get credit for
9421                  * MSS and greater (and you get extra
9422                  * credit for larger cum-ack moves).
9423                  */
9424                 int ac;
9425
9426                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9427                 rack->r_ctl.ack_count += ac;
9428                 counter_u64_add(rack_ack_total, ac);
9429         }
9430         if (rack->r_ctl.ack_count > 0xfff00000) {
9431                 /*
9432                  * reduce the number to keep us under
9433                  * a uint32_t.
9434                  */
9435                 rack->r_ctl.ack_count /= 2;
9436                 rack->r_ctl.sack_count /= 2;
9437         }
9438         if (SEQ_GT(th_ack, tp->snd_una)) {
9439                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9440                 tp->t_acktime = ticks;
9441         }
9442         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9443                 changed = th_ack - rsm->r_start;
9444         if (changed) {
9445                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9446         }
9447         if ((to->to_flags & TOF_SACK) == 0) {
9448                 /* We are done nothing left and no sack. */
9449                 rack_handle_might_revert(tp, rack);
9450                 /*
9451                  * For cases where we struck a dup-ack
9452                  * with no SACK, add to the changes so
9453                  * PRR will work right.
9454                  */
9455                 if (dup_ack_struck && (changed == 0)) {
9456                         changed += ctf_fixed_maxseg(rack->rc_tp);
9457                 }
9458                 goto out;
9459         }
9460         /* Sack block processing */
9461         if (SEQ_GT(th_ack, tp->snd_una))
9462                 ack_point = th_ack;
9463         else
9464                 ack_point = tp->snd_una;
9465         for (i = 0; i < to->to_nsacks; i++) {
9466                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9467                       &sack, sizeof(sack));
9468                 sack.start = ntohl(sack.start);
9469                 sack.end = ntohl(sack.end);
9470                 if (SEQ_GT(sack.end, sack.start) &&
9471                     SEQ_GT(sack.start, ack_point) &&
9472                     SEQ_LT(sack.start, tp->snd_max) &&
9473                     SEQ_GT(sack.end, ack_point) &&
9474                     SEQ_LEQ(sack.end, tp->snd_max)) {
9475                         sack_blocks[num_sack_blks] = sack;
9476                         num_sack_blks++;
9477                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9478                            SEQ_LEQ(sack.end, th_ack)) {
9479                         int was_tlp;
9480
9481                         was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9482                         /*
9483                          * Its a D-SACK block.
9484                          */
9485                         tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9486                 }
9487         }
9488         if (rack->rc_dsack_round_seen) {
9489                 /* Is the dsack roound over? */
9490                 if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9491                         /* Yes it is */
9492                         rack->rc_dsack_round_seen = 0;
9493                         rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9494                 }
9495         }
9496         /*
9497          * Sort the SACK blocks so we can update the rack scoreboard with
9498          * just one pass.
9499          */
9500         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9501                                          num_sack_blks, th->th_ack);
9502         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9503         if (num_sack_blks == 0) {
9504                 /* Nothing to sack (DSACKs?) */
9505                 goto out_with_totals;
9506         }
9507         if (num_sack_blks < 2) {
9508                 /* Only one, we don't need to sort */
9509                 goto do_sack_work;
9510         }
9511         /* Sort the sacks */
9512         for (i = 0; i < num_sack_blks; i++) {
9513                 for (j = i + 1; j < num_sack_blks; j++) {
9514                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9515                                 sack = sack_blocks[i];
9516                                 sack_blocks[i] = sack_blocks[j];
9517                                 sack_blocks[j] = sack;
9518                         }
9519                 }
9520         }
9521         /*
9522          * Now are any of the sack block ends the same (yes some
9523          * implementations send these)?
9524          */
9525 again:
9526         if (num_sack_blks == 0)
9527                 goto out_with_totals;
9528         if (num_sack_blks > 1) {
9529                 for (i = 0; i < num_sack_blks; i++) {
9530                         for (j = i + 1; j < num_sack_blks; j++) {
9531                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9532                                         /*
9533                                          * Ok these two have the same end we
9534                                          * want the smallest end and then
9535                                          * throw away the larger and start
9536                                          * again.
9537                                          */
9538                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9539                                                 /*
9540                                                  * The second block covers
9541                                                  * more area use that
9542                                                  */
9543                                                 sack_blocks[i].start = sack_blocks[j].start;
9544                                         }
9545                                         /*
9546                                          * Now collapse out the dup-sack and
9547                                          * lower the count
9548                                          */
9549                                         for (k = (j + 1); k < num_sack_blks; k++) {
9550                                                 sack_blocks[j].start = sack_blocks[k].start;
9551                                                 sack_blocks[j].end = sack_blocks[k].end;
9552                                                 j++;
9553                                         }
9554                                         num_sack_blks--;
9555                                         goto again;
9556                                 }
9557                         }
9558                 }
9559         }
9560 do_sack_work:
9561         /*
9562          * First lets look to see if
9563          * we have retransmitted and
9564          * can use the transmit next?
9565          */
9566         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9567         if (rsm &&
9568             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9569             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9570                 /*
9571                  * We probably did the FR and the next
9572                  * SACK in continues as we would expect.
9573                  */
9574                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9575                 if (acked) {
9576                         rack->r_wanted_output = 1;
9577                         changed += acked;
9578                 }
9579                 if (num_sack_blks == 1) {
9580                         /*
9581                          * This is what we would expect from
9582                          * a normal implementation to happen
9583                          * after we have retransmitted the FR,
9584                          * i.e the sack-filter pushes down
9585                          * to 1 block and the next to be retransmitted
9586                          * is the sequence in the sack block (has more
9587                          * are acked). Count this as ACK'd data to boost
9588                          * up the chances of recovering any false positives.
9589                          */
9590                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9591                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9592                         counter_u64_add(rack_express_sack, 1);
9593                         if (rack->r_ctl.ack_count > 0xfff00000) {
9594                                 /*
9595                                  * reduce the number to keep us under
9596                                  * a uint32_t.
9597                                  */
9598                                 rack->r_ctl.ack_count /= 2;
9599                                 rack->r_ctl.sack_count /= 2;
9600                         }
9601                         goto out_with_totals;
9602                 } else {
9603                         /*
9604                          * Start the loop through the
9605                          * rest of blocks, past the first block.
9606                          */
9607                         moved_two = 0;
9608                         loop_start = 1;
9609                 }
9610         }
9611         /* Its a sack of some sort */
9612         rack->r_ctl.sack_count++;
9613         if (rack->r_ctl.sack_count > 0xfff00000) {
9614                 /*
9615                  * reduce the number to keep us under
9616                  * a uint32_t.
9617                  */
9618                 rack->r_ctl.ack_count /= 2;
9619                 rack->r_ctl.sack_count /= 2;
9620         }
9621         counter_u64_add(rack_sack_total, 1);
9622         if (rack->sack_attack_disable) {
9623                 /* An attacker disablement is in place */
9624                 if (num_sack_blks > 1) {
9625                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9626                         rack->r_ctl.sack_moved_extra++;
9627                         counter_u64_add(rack_move_some, 1);
9628                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9629                                 rack->r_ctl.sack_moved_extra /= 2;
9630                                 rack->r_ctl.sack_noextra_move /= 2;
9631                         }
9632                 }
9633                 goto out;
9634         }
9635         rsm = rack->r_ctl.rc_sacklast;
9636         for (i = loop_start; i < num_sack_blks; i++) {
9637                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9638                 if (acked) {
9639                         rack->r_wanted_output = 1;
9640                         changed += acked;
9641                 }
9642                 if (moved_two) {
9643                         /*
9644                          * If we did not get a SACK for at least a MSS and
9645                          * had to move at all, or if we moved more than our
9646                          * threshold, it counts against the "extra" move.
9647                          */
9648                         rack->r_ctl.sack_moved_extra += moved_two;
9649                         counter_u64_add(rack_move_some, 1);
9650                 } else {
9651                         /*
9652                          * else we did not have to move
9653                          * any more than we would expect.
9654                          */
9655                         rack->r_ctl.sack_noextra_move++;
9656                         counter_u64_add(rack_move_none, 1);
9657                 }
9658                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9659                         /*
9660                          * If the SACK was not a full MSS then
9661                          * we add to sack_count the number of
9662                          * MSS's (or possibly more than
9663                          * a MSS if its a TSO send) we had to skip by.
9664                          */
9665                         rack->r_ctl.sack_count += moved_two;
9666                         counter_u64_add(rack_sack_total, moved_two);
9667                 }
9668                 /*
9669                  * Now we need to setup for the next
9670                  * round. First we make sure we won't
9671                  * exceed the size of our uint32_t on
9672                  * the various counts, and then clear out
9673                  * moved_two.
9674                  */
9675                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9676                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9677                         rack->r_ctl.sack_moved_extra /= 2;
9678                         rack->r_ctl.sack_noextra_move /= 2;
9679                 }
9680                 if (rack->r_ctl.sack_count > 0xfff00000) {
9681                         rack->r_ctl.ack_count /= 2;
9682                         rack->r_ctl.sack_count /= 2;
9683                 }
9684                 moved_two = 0;
9685         }
9686 out_with_totals:
9687         if (num_sack_blks > 1) {
9688                 /*
9689                  * You get an extra stroke if
9690                  * you have more than one sack-blk, this
9691                  * could be where we are skipping forward
9692                  * and the sack-filter is still working, or
9693                  * it could be an attacker constantly
9694                  * moving us.
9695                  */
9696                 rack->r_ctl.sack_moved_extra++;
9697                 counter_u64_add(rack_move_some, 1);
9698         }
9699 out:
9700 #ifdef NETFLIX_EXP_DETECTION
9701         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9702 #endif
9703         if (changed) {
9704                 /* Something changed cancel the rack timer */
9705                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9706         }
9707         tsused = tcp_get_usecs(NULL);
9708         rsm = tcp_rack_output(tp, rack, tsused);
9709         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9710             rsm &&
9711             ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
9712                 /* Enter recovery */
9713                 entered_recovery = 1;
9714                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9715                 /*
9716                  * When we enter recovery we need to assure we send
9717                  * one packet.
9718                  */
9719                 if (rack->rack_no_prr == 0) {
9720                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9721                         rack_log_to_prr(rack, 8, 0, __LINE__);
9722                 }
9723                 rack->r_timer_override = 1;
9724                 rack->r_early = 0;
9725                 rack->r_ctl.rc_agg_early = 0;
9726         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9727                    rsm &&
9728                    (rack->r_rr_config == 3)) {
9729                 /*
9730                  * Assure we can output and we get no
9731                  * remembered pace time except the retransmit.
9732                  */
9733                 rack->r_timer_override = 1;
9734                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9735                 rack->r_ctl.rc_resend = rsm;
9736         }
9737         if (IN_FASTRECOVERY(tp->t_flags) &&
9738             (rack->rack_no_prr == 0) &&
9739             (entered_recovery == 0)) {
9740                 rack_update_prr(tp, rack, changed, th_ack);
9741                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9742                      ((tcp_in_hpts(rack->rc_inp) == 0) &&
9743                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9744                         /*
9745                          * If you are pacing output you don't want
9746                          * to override.
9747                          */
9748                         rack->r_early = 0;
9749                         rack->r_ctl.rc_agg_early = 0;
9750                         rack->r_timer_override = 1;
9751                 }
9752         }
9753 }
9754
9755 static void
9756 rack_strike_dupack(struct tcp_rack *rack)
9757 {
9758         struct rack_sendmap *rsm;
9759
9760         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9761         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9762                 rsm = TAILQ_NEXT(rsm, r_tnext);
9763                 if (rsm->r_flags & RACK_MUST_RXT) {
9764                         /* Sendmap entries that are marked to
9765                          * be retransmitted do not need dupack's
9766                          * struck. We get these marks for a number
9767                          * of reasons (rxt timeout with no sack,
9768                          * mtu change, or rwnd collapses). When
9769                          * these events occur, we know we must retransmit
9770                          * them and mark the sendmap entries. Dupack counting
9771                          * is not needed since we are already set to retransmit
9772                          * it as soon as we can.
9773                          */
9774                         continue;
9775                 }
9776         }
9777         if (rsm && (rsm->r_dupack < 0xff)) {
9778                 rsm->r_dupack++;
9779                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9780                         struct timeval tv;
9781                         uint32_t cts;
9782                         /*
9783                          * Here we see if we need to retransmit. For
9784                          * a SACK type connection if enough time has passed
9785                          * we will get a return of the rsm. For a non-sack
9786                          * connection we will get the rsm returned if the
9787                          * dupack value is 3 or more.
9788                          */
9789                         cts = tcp_get_usecs(&tv);
9790                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9791                         if (rack->r_ctl.rc_resend != NULL) {
9792                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9793                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9794                                                          rack->rc_tp->snd_una, __LINE__);
9795                                 }
9796                                 rack->r_wanted_output = 1;
9797                                 rack->r_timer_override = 1;
9798                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9799                         }
9800                 } else {
9801                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9802                 }
9803         }
9804 }
9805
9806 static void
9807 rack_check_bottom_drag(struct tcpcb *tp,
9808                        struct tcp_rack *rack,
9809                        struct socket *so, int32_t acked)
9810 {
9811         uint32_t segsiz, minseg;
9812
9813         segsiz = ctf_fixed_maxseg(tp);
9814         minseg = segsiz;
9815
9816         if (tp->snd_max == tp->snd_una) {
9817                 /*
9818                  * We are doing dynamic pacing and we are way
9819                  * under. Basically everything got acked while
9820                  * we were still waiting on the pacer to expire.
9821                  *
9822                  * This means we need to boost the b/w in
9823                  * addition to any earlier boosting of
9824                  * the multiplier.
9825                  */
9826                 rack->rc_dragged_bottom = 1;
9827                 rack_validate_multipliers_at_or_above100(rack);
9828                 /*
9829                  * Lets use the segment bytes acked plus
9830                  * the lowest RTT seen as the basis to
9831                  * form a b/w estimate. This will be off
9832                  * due to the fact that the true estimate
9833                  * should be around 1/2 the time of the RTT
9834                  * but we can settle for that.
9835                  */
9836                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9837                     acked) {
9838                         uint64_t bw, calc_bw, rtt;
9839
9840                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9841                         if (rtt == 0) {
9842                                 /* no us sample is there a ms one? */
9843                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9844                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9845                                 } else {
9846                                         goto no_measurement;
9847                                 }
9848                         }
9849                         bw = acked;
9850                         calc_bw = bw * 1000000;
9851                         calc_bw /= rtt;
9852                         if (rack->r_ctl.last_max_bw &&
9853                             (rack->r_ctl.last_max_bw < calc_bw)) {
9854                                 /*
9855                                  * If we have a last calculated max bw
9856                                  * enforce it.
9857                                  */
9858                                 calc_bw = rack->r_ctl.last_max_bw;
9859                         }
9860                         /* now plop it in */
9861                         if (rack->rc_gp_filled == 0) {
9862                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9863                                         /*
9864                                          * If we have no measurement
9865                                          * don't let us set in more than
9866                                          * 1.2Mbps. If we are still too
9867                                          * low after pacing with this we
9868                                          * will hopefully have a max b/w
9869                                          * available to sanity check things.
9870                                          */
9871                                         calc_bw = ONE_POINT_TWO_MEG;
9872                                 }
9873                                 rack->r_ctl.rc_rtt_diff = 0;
9874                                 rack->r_ctl.gp_bw = calc_bw;
9875                                 rack->rc_gp_filled = 1;
9876                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9877                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9878                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9879                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9880                                 rack->r_ctl.rc_rtt_diff = 0;
9881                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9882                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9883                                 rack->r_ctl.gp_bw = calc_bw;
9884                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9885                         } else
9886                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9887                         if ((rack->gp_ready == 0) &&
9888                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9889                                 /* We have enough measurements now */
9890                                 rack->gp_ready = 1;
9891                                 rack_set_cc_pacing(rack);
9892                                 if (rack->defer_options)
9893                                         rack_apply_deferred_options(rack);
9894                         }
9895                         /*
9896                          * For acks over 1mss we do a extra boost to simulate
9897                          * where we would get 2 acks (we want 110 for the mul).
9898                          */
9899                         if (acked > segsiz)
9900                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9901                 } else {
9902                         /*
9903                          * zero rtt possibly?, settle for just an old increase.
9904                          */
9905 no_measurement:
9906                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9907                 }
9908         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9909                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9910                                                minseg)) &&
9911                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9912                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9913                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9914                     (segsiz * rack_req_segs))) {
9915                 /*
9916                  * We are doing dynamic GP pacing and
9917                  * we have everything except 1MSS or less
9918                  * bytes left out. We are still pacing away.
9919                  * And there is data that could be sent, This
9920                  * means we are inserting delayed ack time in
9921                  * our measurements because we are pacing too slow.
9922                  */
9923                 rack_validate_multipliers_at_or_above100(rack);
9924                 rack->rc_dragged_bottom = 1;
9925                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9926         }
9927 }
9928
9929
9930
9931 static void
9932 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9933 {
9934         /*
9935          * The fast output path is enabled and we
9936          * have moved the cumack forward. Lets see if
9937          * we can expand forward the fast path length by
9938          * that amount. What we would ideally like to
9939          * do is increase the number of bytes in the
9940          * fast path block (left_to_send) by the
9941          * acked amount. However we have to gate that
9942          * by two factors:
9943          * 1) The amount outstanding and the rwnd of the peer
9944          *    (i.e. we don't want to exceed the rwnd of the peer).
9945          *    <and>
9946          * 2) The amount of data left in the socket buffer (i.e.
9947          *    we can't send beyond what is in the buffer).
9948          *
9949          * Note that this does not take into account any increase
9950          * in the cwnd. We will only extend the fast path by
9951          * what was acked.
9952          */
9953         uint32_t new_total, gating_val;
9954
9955         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9956         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9957                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9958         if (new_total <= gating_val) {
9959                 /* We can increase left_to_send by the acked amount */
9960                 counter_u64_add(rack_extended_rfo, 1);
9961                 rack->r_ctl.fsb.left_to_send = new_total;
9962                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9963                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
9964                          rack, rack->r_ctl.fsb.left_to_send,
9965                          sbavail(&rack->rc_inp->inp_socket->so_snd),
9966                          (tp->snd_max - tp->snd_una)));
9967
9968         }
9969 }
9970
9971 static void
9972 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9973 {
9974         /*
9975          * Here any sendmap entry that points to the
9976          * beginning mbuf must be adjusted to the correct
9977          * offset. This must be called with:
9978          * 1) The socket buffer locked
9979          * 2) snd_una adjusted to its new position.
9980          *
9981          * Note that (2) implies rack_ack_received has also
9982          * been called.
9983          *
9984          * We grab the first mbuf in the socket buffer and
9985          * then go through the front of the sendmap, recalculating
9986          * the stored offset for any sendmap entry that has
9987          * that mbuf. We must use the sb functions to do this
9988          * since its possible an add was done has well as
9989          * the subtraction we may have just completed. This should
9990          * not be a penalty though, since we just referenced the sb
9991          * to go in and trim off the mbufs that we freed (of course
9992          * there will be a penalty for the sendmap references though).
9993          */
9994         struct mbuf *m;
9995         struct rack_sendmap *rsm;
9996
9997         SOCKBUF_LOCK_ASSERT(sb);
9998         m = sb->sb_mb;
9999         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
10000         if ((rsm == NULL) || (m == NULL)) {
10001                 /* Nothing outstanding */
10002                 return;
10003         }
10004         while (rsm->m && (rsm->m == m)) {
10005                 /* one to adjust */
10006 #ifdef INVARIANTS
10007                 struct mbuf *tm;
10008                 uint32_t soff;
10009
10010                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
10011                 if (rsm->orig_m_len != m->m_len) {
10012                         rack_adjust_orig_mlen(rsm);
10013                 }
10014                 if (rsm->soff != soff) {
10015                         /*
10016                          * This is not a fatal error, we anticipate it
10017                          * might happen (the else code), so we count it here
10018                          * so that under invariant we can see that it really
10019                          * does happen.
10020                          */
10021                         counter_u64_add(rack_adjust_map_bw, 1);
10022                 }
10023                 rsm->m = tm;
10024                 rsm->soff = soff;
10025                 if (tm)
10026                         rsm->orig_m_len = rsm->m->m_len;
10027                 else
10028                         rsm->orig_m_len = 0;
10029 #else
10030                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
10031                 if (rsm->m)
10032                         rsm->orig_m_len = rsm->m->m_len;
10033                 else
10034                         rsm->orig_m_len = 0;
10035 #endif
10036                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
10037                               rsm);
10038                 if (rsm == NULL)
10039                         break;
10040         }
10041 }
10042
10043 /*
10044  * Return value of 1, we do not need to call rack_process_data().
10045  * return value of 0, rack_process_data can be called.
10046  * For ret_val if its 0 the TCP is locked, if its non-zero
10047  * its unlocked and probably unsafe to touch the TCB.
10048  */
10049 static int
10050 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10051     struct tcpcb *tp, struct tcpopt *to,
10052     uint32_t tiwin, int32_t tlen,
10053     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10054 {
10055         int32_t ourfinisacked = 0;
10056         int32_t nsegs, acked_amount;
10057         int32_t acked;
10058         struct mbuf *mfree;
10059         struct tcp_rack *rack;
10060         int32_t under_pacing = 0;
10061         int32_t recovery = 0;
10062
10063         INP_WLOCK_ASSERT(tptoinpcb(tp));
10064
10065         rack = (struct tcp_rack *)tp->t_fb_ptr;
10066         if (SEQ_GT(th->th_ack, tp->snd_max)) {
10067                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10068                                       &rack->r_ctl.challenge_ack_ts,
10069                                       &rack->r_ctl.challenge_ack_cnt);
10070                 rack->r_wanted_output = 1;
10071                 return (1);
10072         }
10073         if (rack->gp_ready &&
10074             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10075                 under_pacing = 1;
10076         }
10077         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10078                 int in_rec, dup_ack_struck = 0;
10079
10080                 in_rec = IN_FASTRECOVERY(tp->t_flags);
10081                 if (rack->rc_in_persist) {
10082                         tp->t_rxtshift = 0;
10083                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10084                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10085                 }
10086                 if ((th->th_ack == tp->snd_una) &&
10087                     (tiwin == tp->snd_wnd) &&
10088                     ((to->to_flags & TOF_SACK) == 0)) {
10089                         rack_strike_dupack(rack);
10090                         dup_ack_struck = 1;
10091                 }
10092                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10093         }
10094         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10095                 /*
10096                  * Old ack, behind (or duplicate to) the last one rcv'd
10097                  * Note: We mark reordering is occuring if its
10098                  * less than and we have not closed our window.
10099                  */
10100                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10101                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10102                 }
10103                 return (0);
10104         }
10105         /*
10106          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10107          * something we sent.
10108          */
10109         if (tp->t_flags & TF_NEEDSYN) {
10110                 /*
10111                  * T/TCP: Connection was half-synchronized, and our SYN has
10112                  * been ACK'd (so connection is now fully synchronized).  Go
10113                  * to non-starred state, increment snd_una for ACK of SYN,
10114                  * and check if we can do window scaling.
10115                  */
10116                 tp->t_flags &= ~TF_NEEDSYN;
10117                 tp->snd_una++;
10118                 /* Do window scaling? */
10119                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10120                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10121                         tp->rcv_scale = tp->request_r_scale;
10122                         /* Send window already scaled. */
10123                 }
10124         }
10125         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10126
10127         acked = BYTES_THIS_ACK(tp, th);
10128         if (acked) {
10129                 /*
10130                  * Any time we move the cum-ack forward clear
10131                  * keep-alive tied probe-not-answered. The
10132                  * persists clears its own on entry.
10133                  */
10134                 rack->probe_not_answered = 0;
10135         }
10136         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10137         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10138         /*
10139          * If we just performed our first retransmit, and the ACK arrives
10140          * within our recovery window, then it was a mistake to do the
10141          * retransmit in the first place.  Recover our original cwnd and
10142          * ssthresh, and proceed to transmit where we left off.
10143          */
10144         if ((tp->t_flags & TF_PREVVALID) &&
10145             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10146                 tp->t_flags &= ~TF_PREVVALID;
10147                 if (tp->t_rxtshift == 1 &&
10148                     (int)(ticks - tp->t_badrxtwin) < 0)
10149                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10150         }
10151         if (acked) {
10152                 /* assure we are not backed off */
10153                 tp->t_rxtshift = 0;
10154                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10155                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10156                 rack->rc_tlp_in_progress = 0;
10157                 rack->r_ctl.rc_tlp_cnt_out = 0;
10158                 /*
10159                  * If it is the RXT timer we want to
10160                  * stop it, so we can restart a TLP.
10161                  */
10162                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10163                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10164 #ifdef NETFLIX_HTTP_LOGGING
10165                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10166 #endif
10167         }
10168         /*
10169          * If we have a timestamp reply, update smoothed round trip time. If
10170          * no timestamp is present but transmit timer is running and timed
10171          * sequence number was acked, update smoothed round trip time. Since
10172          * we now have an rtt measurement, cancel the timer backoff (cf.,
10173          * Phil Karn's retransmit alg.). Recompute the initial retransmit
10174          * timer.
10175          *
10176          * Some boxes send broken timestamp replies during the SYN+ACK
10177          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10178          * and blow up the retransmit timer.
10179          */
10180         /*
10181          * If all outstanding data is acked, stop retransmit timer and
10182          * remember to restart (more output or persist). If there is more
10183          * data to be acked, restart retransmit timer, using current
10184          * (possibly backed-off) value.
10185          */
10186         if (acked == 0) {
10187                 if (ofia)
10188                         *ofia = ourfinisacked;
10189                 return (0);
10190         }
10191         if (IN_RECOVERY(tp->t_flags)) {
10192                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10193                     (SEQ_LT(th->th_ack, tp->snd_max))) {
10194                         tcp_rack_partialack(tp);
10195                 } else {
10196                         rack_post_recovery(tp, th->th_ack);
10197                         recovery = 1;
10198                 }
10199         }
10200         /*
10201          * Let the congestion control algorithm update congestion control
10202          * related information. This typically means increasing the
10203          * congestion window.
10204          */
10205         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10206         SOCKBUF_LOCK(&so->so_snd);
10207         acked_amount = min(acked, (int)sbavail(&so->so_snd));
10208         tp->snd_wnd -= acked_amount;
10209         mfree = sbcut_locked(&so->so_snd, acked_amount);
10210         if ((sbused(&so->so_snd) == 0) &&
10211             (acked > acked_amount) &&
10212             (tp->t_state >= TCPS_FIN_WAIT_1) &&
10213             (tp->t_flags & TF_SENTFIN)) {
10214                 /*
10215                  * We must be sure our fin
10216                  * was sent and acked (we can be
10217                  * in FIN_WAIT_1 without having
10218                  * sent the fin).
10219                  */
10220                 ourfinisacked = 1;
10221         }
10222         tp->snd_una = th->th_ack;
10223         if (acked_amount && sbavail(&so->so_snd))
10224                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10225         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10226         /* NB: sowwakeup_locked() does an implicit unlock. */
10227         sowwakeup_locked(so);
10228         m_freem(mfree);
10229         if (SEQ_GT(tp->snd_una, tp->snd_recover))
10230                 tp->snd_recover = tp->snd_una;
10231
10232         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10233                 tp->snd_nxt = tp->snd_una;
10234         }
10235         if (under_pacing &&
10236             (rack->use_fixed_rate == 0) &&
10237             (rack->in_probe_rtt == 0) &&
10238             rack->rc_gp_dyn_mul &&
10239             rack->rc_always_pace) {
10240                 /* Check if we are dragging bottom */
10241                 rack_check_bottom_drag(tp, rack, so, acked);
10242         }
10243         if (tp->snd_una == tp->snd_max) {
10244                 /* Nothing left outstanding */
10245                 tp->t_flags &= ~TF_PREVVALID;
10246                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10247                 rack->r_ctl.retran_during_recovery = 0;
10248                 rack->r_ctl.dsack_byte_cnt = 0;
10249                 if (rack->r_ctl.rc_went_idle_time == 0)
10250                         rack->r_ctl.rc_went_idle_time = 1;
10251                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10252                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
10253                         tp->t_acktime = 0;
10254                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10255                 /* Set need output so persist might get set */
10256                 rack->r_wanted_output = 1;
10257                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10258                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10259                     (sbavail(&so->so_snd) == 0) &&
10260                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10261                         /*
10262                          * The socket was gone and the
10263                          * peer sent data (now or in the past), time to
10264                          * reset him.
10265                          */
10266                         *ret_val = 1;
10267                         /* tcp_close will kill the inp pre-log the Reset */
10268                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10269                         tp = tcp_close(tp);
10270                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10271                         return (1);
10272                 }
10273         }
10274         if (ofia)
10275                 *ofia = ourfinisacked;
10276         return (0);
10277 }
10278
10279
10280 static void
10281 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
10282                   int dir, uint32_t flags, struct rack_sendmap *rsm)
10283 {
10284         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
10285                 union tcp_log_stackspecific log;
10286                 struct timeval tv;
10287
10288                 memset(&log, 0, sizeof(log));
10289                 log.u_bbr.flex1 = cnt;
10290                 log.u_bbr.flex2 = split;
10291                 log.u_bbr.flex3 = out;
10292                 log.u_bbr.flex4 = line;
10293                 log.u_bbr.flex5 = rack->r_must_retran;
10294                 log.u_bbr.flex6 = flags;
10295                 log.u_bbr.flex7 = rack->rc_has_collapsed;
10296                 log.u_bbr.flex8 = dir;  /*
10297                                          * 1 is collapsed, 0 is uncollapsed,
10298                                          * 2 is log of a rsm being marked, 3 is a split.
10299                                          */
10300                 if (rsm == NULL)
10301                         log.u_bbr.rttProp = 0;
10302                 else
10303                         log.u_bbr.rttProp = (uint64_t)rsm;
10304                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
10305                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10306                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
10307                     &rack->rc_inp->inp_socket->so_rcv,
10308                     &rack->rc_inp->inp_socket->so_snd,
10309                     TCP_RACK_LOG_COLLAPSE, 0,
10310                     0, &log, false, &tv);
10311         }
10312 }
10313
10314 static void
10315 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, int line)
10316 {
10317         /*
10318          * Here all we do is mark the collapsed point and set the flag.
10319          * This may happen again and again, but there is no
10320          * sense splitting our map until we know where the
10321          * peer finally lands in the collapse.
10322          */
10323         rack_trace_point(rack, RACK_TP_COLLAPSED_WND);
10324         if ((rack->rc_has_collapsed == 0) ||
10325             (rack->r_ctl.last_collapse_point != (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)))
10326                 counter_u64_add(rack_collapsed_win_seen, 1);
10327         rack->r_ctl.last_collapse_point = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10328         rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
10329         rack->rc_has_collapsed = 1;
10330         rack->r_collapse_point_valid = 1;
10331         rack_log_collapse(rack, 0, 0, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
10332 }
10333
10334 static void
10335 rack_un_collapse_window(struct tcp_rack *rack, int line)
10336 {
10337         struct rack_sendmap *nrsm, *rsm, fe;
10338         int cnt = 0, split = 0;
10339 #ifdef INVARIANTS
10340         struct rack_sendmap *insret;
10341 #endif
10342
10343         memset(&fe, 0, sizeof(fe));
10344         rack->rc_has_collapsed = 0;
10345         fe.r_start = rack->r_ctl.last_collapse_point;
10346         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10347         if (rsm == NULL) {
10348                 /* Nothing to do maybe the peer ack'ed it all */
10349                 rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10350                 return;
10351         }
10352         /* Now do we need to split this one? */
10353         if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
10354                 rack_log_collapse(rack, rsm->r_start, rsm->r_end,
10355                                   rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
10356                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10357                 if (nrsm == NULL) {
10358                         /* We can't get a rsm, mark all? */
10359                         nrsm = rsm;
10360                         goto no_split;
10361                 }
10362                 /* Clone it */
10363                 split = 1;
10364                 rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
10365 #ifndef INVARIANTS
10366                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10367 #else
10368                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10369                 if (insret != NULL) {
10370                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10371                               nrsm, insret, rack, rsm);
10372                 }
10373 #endif
10374                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
10375                                  rack->r_ctl.last_collapse_point, __LINE__);
10376                 if (rsm->r_in_tmap) {
10377                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10378                         nrsm->r_in_tmap = 1;
10379                 }
10380                 /*
10381                  * Set in the new RSM as the
10382                  * collapsed starting point
10383                  */
10384                 rsm = nrsm;
10385         }
10386 no_split:
10387         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10388                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
10389                 rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
10390                 cnt++;
10391         }
10392         if (cnt) {
10393                 counter_u64_add(rack_collapsed_win, 1);
10394         }
10395         rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10396 }
10397
10398 static void
10399 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10400                         int32_t tlen, int32_t tfo_syn)
10401 {
10402         if (DELAY_ACK(tp, tlen) || tfo_syn) {
10403                 if (rack->rc_dack_mode &&
10404                     (tlen > 500) &&
10405                     (rack->rc_dack_toggle == 1)) {
10406                         goto no_delayed_ack;
10407                 }
10408                 rack_timer_cancel(tp, rack,
10409                                   rack->r_ctl.rc_rcvtime, __LINE__);
10410                 tp->t_flags |= TF_DELACK;
10411         } else {
10412 no_delayed_ack:
10413                 rack->r_wanted_output = 1;
10414                 tp->t_flags |= TF_ACKNOW;
10415                 if (rack->rc_dack_mode) {
10416                         if (tp->t_flags & TF_DELACK)
10417                                 rack->rc_dack_toggle = 1;
10418                         else
10419                                 rack->rc_dack_toggle = 0;
10420                 }
10421         }
10422 }
10423
10424 static void
10425 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10426 {
10427         /*
10428          * If fast output is in progress, lets validate that
10429          * the new window did not shrink on us and make it
10430          * so fast output should end.
10431          */
10432         if (rack->r_fast_output) {
10433                 uint32_t out;
10434
10435                 /*
10436                  * Calculate what we will send if left as is
10437                  * and compare that to our send window.
10438                  */
10439                 out = ctf_outstanding(tp);
10440                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10441                         /* ok we have an issue */
10442                         if (out >= tp->snd_wnd) {
10443                                 /* Turn off fast output the window is met or collapsed */
10444                                 rack->r_fast_output = 0;
10445                         } else {
10446                                 /* we have some room left */
10447                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10448                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10449                                         /* If not at least 1 full segment never mind */
10450                                         rack->r_fast_output = 0;
10451                                 }
10452                         }
10453                 }
10454         }
10455 }
10456
10457
10458 /*
10459  * Return value of 1, the TCB is unlocked and most
10460  * likely gone, return value of 0, the TCP is still
10461  * locked.
10462  */
10463 static int
10464 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10465     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10466     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10467 {
10468         /*
10469          * Update window information. Don't look at window if no ACK: TAC's
10470          * send garbage on first SYN.
10471          */
10472         int32_t nsegs;
10473         int32_t tfo_syn;
10474         struct tcp_rack *rack;
10475
10476         INP_WLOCK_ASSERT(tptoinpcb(tp));
10477
10478         rack = (struct tcp_rack *)tp->t_fb_ptr;
10479         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10480         if ((thflags & TH_ACK) &&
10481             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10482             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10483             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10484                 /* keep track of pure window updates */
10485                 if (tlen == 0 &&
10486                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10487                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10488                 tp->snd_wnd = tiwin;
10489                 rack_validate_fo_sendwin_up(tp, rack);
10490                 tp->snd_wl1 = th->th_seq;
10491                 tp->snd_wl2 = th->th_ack;
10492                 if (tp->snd_wnd > tp->max_sndwnd)
10493                         tp->max_sndwnd = tp->snd_wnd;
10494                 rack->r_wanted_output = 1;
10495         } else if (thflags & TH_ACK) {
10496                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10497                         tp->snd_wnd = tiwin;
10498                         rack_validate_fo_sendwin_up(tp, rack);
10499                         tp->snd_wl1 = th->th_seq;
10500                         tp->snd_wl2 = th->th_ack;
10501                 }
10502         }
10503         if (tp->snd_wnd < ctf_outstanding(tp))
10504                 /* The peer collapsed the window */
10505                 rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
10506         else if (rack->rc_has_collapsed)
10507                 rack_un_collapse_window(rack, __LINE__);
10508         if ((rack->r_collapse_point_valid) &&
10509             (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
10510                 rack->r_collapse_point_valid = 0;
10511         /* Was persist timer active and now we have window space? */
10512         if ((rack->rc_in_persist != 0) &&
10513             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10514                                 rack->r_ctl.rc_pace_min_segs))) {
10515                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10516                 tp->snd_nxt = tp->snd_max;
10517                 /* Make sure we output to start the timer */
10518                 rack->r_wanted_output = 1;
10519         }
10520         /* Do we enter persists? */
10521         if ((rack->rc_in_persist == 0) &&
10522             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10523             TCPS_HAVEESTABLISHED(tp->t_state) &&
10524             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10525             sbavail(&tptosocket(tp)->so_snd) &&
10526             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10527                 /*
10528                  * Here the rwnd is less than
10529                  * the pacing size, we are established,
10530                  * nothing is outstanding, and there is
10531                  * data to send. Enter persists.
10532                  */
10533                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10534         }
10535         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10536                 m_freem(m);
10537                 return (0);
10538         }
10539         /*
10540          * don't process the URG bit, ignore them drag
10541          * along the up.
10542          */
10543         tp->rcv_up = tp->rcv_nxt;
10544
10545         /*
10546          * Process the segment text, merging it into the TCP sequencing
10547          * queue, and arranging for acknowledgment of receipt if necessary.
10548          * This process logically involves adjusting tp->rcv_wnd as data is
10549          * presented to the user (this happens in tcp_usrreq.c, case
10550          * PRU_RCVD).  If a FIN has already been received on this connection
10551          * then we just ignore the text.
10552          */
10553         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10554                    IS_FASTOPEN(tp->t_flags));
10555         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10556             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10557                 tcp_seq save_start = th->th_seq;
10558                 tcp_seq save_rnxt  = tp->rcv_nxt;
10559                 int     save_tlen  = tlen;
10560
10561                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10562                 /*
10563                  * Insert segment which includes th into TCP reassembly
10564                  * queue with control block tp.  Set thflags to whether
10565                  * reassembly now includes a segment with FIN.  This handles
10566                  * the common case inline (segment is the next to be
10567                  * received on an established connection, and the queue is
10568                  * empty), avoiding linkage into and removal from the queue
10569                  * and repetition of various conversions. Set DELACK for
10570                  * segments received in order, but ack immediately when
10571                  * segments are out of order (so fast retransmit can work).
10572                  */
10573                 if (th->th_seq == tp->rcv_nxt &&
10574                     SEGQ_EMPTY(tp) &&
10575                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10576                     tfo_syn)) {
10577 #ifdef NETFLIX_SB_LIMITS
10578                         u_int mcnt, appended;
10579
10580                         if (so->so_rcv.sb_shlim) {
10581                                 mcnt = m_memcnt(m);
10582                                 appended = 0;
10583                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10584                                     CFO_NOSLEEP, NULL) == false) {
10585                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10586                                         m_freem(m);
10587                                         return (0);
10588                                 }
10589                         }
10590 #endif
10591                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10592                         tp->rcv_nxt += tlen;
10593                         if (tlen &&
10594                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10595                             (tp->t_fbyte_in == 0)) {
10596                                 tp->t_fbyte_in = ticks;
10597                                 if (tp->t_fbyte_in == 0)
10598                                         tp->t_fbyte_in = 1;
10599                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10600                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10601                         }
10602                         thflags = tcp_get_flags(th) & TH_FIN;
10603                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10604                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10605                         SOCKBUF_LOCK(&so->so_rcv);
10606                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10607                                 m_freem(m);
10608                         } else
10609 #ifdef NETFLIX_SB_LIMITS
10610                                 appended =
10611 #endif
10612                                         sbappendstream_locked(&so->so_rcv, m, 0);
10613
10614                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10615                         /* NB: sorwakeup_locked() does an implicit unlock. */
10616                         sorwakeup_locked(so);
10617 #ifdef NETFLIX_SB_LIMITS
10618                         if (so->so_rcv.sb_shlim && appended != mcnt)
10619                                 counter_fo_release(so->so_rcv.sb_shlim,
10620                                     mcnt - appended);
10621 #endif
10622                 } else {
10623                         /*
10624                          * XXX: Due to the header drop above "th" is
10625                          * theoretically invalid by now.  Fortunately
10626                          * m_adj() doesn't actually frees any mbufs when
10627                          * trimming from the head.
10628                          */
10629                         tcp_seq temp = save_start;
10630
10631                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10632                         tp->t_flags |= TF_ACKNOW;
10633                         if (tp->t_flags & TF_WAKESOR) {
10634                                 tp->t_flags &= ~TF_WAKESOR;
10635                                 /* NB: sorwakeup_locked() does an implicit unlock. */
10636                                 sorwakeup_locked(so);
10637                         }
10638                 }
10639                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10640                     (save_tlen > 0) &&
10641                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10642                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10643                                 /*
10644                                  * DSACK actually handled in the fastpath
10645                                  * above.
10646                                  */
10647                                 RACK_OPTS_INC(tcp_sack_path_1);
10648                                 tcp_update_sack_list(tp, save_start,
10649                                     save_start + save_tlen);
10650                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10651                                 if ((tp->rcv_numsacks >= 1) &&
10652                                     (tp->sackblks[0].end == save_start)) {
10653                                         /*
10654                                          * Partial overlap, recorded at todrop
10655                                          * above.
10656                                          */
10657                                         RACK_OPTS_INC(tcp_sack_path_2a);
10658                                         tcp_update_sack_list(tp,
10659                                             tp->sackblks[0].start,
10660                                             tp->sackblks[0].end);
10661                                 } else {
10662                                         RACK_OPTS_INC(tcp_sack_path_2b);
10663                                         tcp_update_dsack_list(tp, save_start,
10664                                             save_start + save_tlen);
10665                                 }
10666                         } else if (tlen >= save_tlen) {
10667                                 /* Update of sackblks. */
10668                                 RACK_OPTS_INC(tcp_sack_path_3);
10669                                 tcp_update_dsack_list(tp, save_start,
10670                                     save_start + save_tlen);
10671                         } else if (tlen > 0) {
10672                                 RACK_OPTS_INC(tcp_sack_path_4);
10673                                 tcp_update_dsack_list(tp, save_start,
10674                                     save_start + tlen);
10675                         }
10676                 }
10677         } else {
10678                 m_freem(m);
10679                 thflags &= ~TH_FIN;
10680         }
10681
10682         /*
10683          * If FIN is received ACK the FIN and let the user know that the
10684          * connection is closing.
10685          */
10686         if (thflags & TH_FIN) {
10687                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10688                         /* The socket upcall is handled by socantrcvmore. */
10689                         socantrcvmore(so);
10690                         /*
10691                          * If connection is half-synchronized (ie NEEDSYN
10692                          * flag on) then delay ACK, so it may be piggybacked
10693                          * when SYN is sent. Otherwise, since we received a
10694                          * FIN then no more input can be expected, send ACK
10695                          * now.
10696                          */
10697                         if (tp->t_flags & TF_NEEDSYN) {
10698                                 rack_timer_cancel(tp, rack,
10699                                     rack->r_ctl.rc_rcvtime, __LINE__);
10700                                 tp->t_flags |= TF_DELACK;
10701                         } else {
10702                                 tp->t_flags |= TF_ACKNOW;
10703                         }
10704                         tp->rcv_nxt++;
10705                 }
10706                 switch (tp->t_state) {
10707                         /*
10708                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10709                          * CLOSE_WAIT state.
10710                          */
10711                 case TCPS_SYN_RECEIVED:
10712                         tp->t_starttime = ticks;
10713                         /* FALLTHROUGH */
10714                 case TCPS_ESTABLISHED:
10715                         rack_timer_cancel(tp, rack,
10716                             rack->r_ctl.rc_rcvtime, __LINE__);
10717                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10718                         break;
10719
10720                         /*
10721                          * If still in FIN_WAIT_1 STATE FIN has not been
10722                          * acked so enter the CLOSING state.
10723                          */
10724                 case TCPS_FIN_WAIT_1:
10725                         rack_timer_cancel(tp, rack,
10726                             rack->r_ctl.rc_rcvtime, __LINE__);
10727                         tcp_state_change(tp, TCPS_CLOSING);
10728                         break;
10729
10730                         /*
10731                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10732                          * starting the time-wait timer, turning off the
10733                          * other standard timers.
10734                          */
10735                 case TCPS_FIN_WAIT_2:
10736                         rack_timer_cancel(tp, rack,
10737                             rack->r_ctl.rc_rcvtime, __LINE__);
10738                         tcp_twstart(tp);
10739                         return (1);
10740                 }
10741         }
10742         /*
10743          * Return any desired output.
10744          */
10745         if ((tp->t_flags & TF_ACKNOW) ||
10746             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10747                 rack->r_wanted_output = 1;
10748         }
10749         return (0);
10750 }
10751
10752 /*
10753  * Here nothing is really faster, its just that we
10754  * have broken out the fast-data path also just like
10755  * the fast-ack.
10756  */
10757 static int
10758 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10759     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10760     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10761 {
10762         int32_t nsegs;
10763         int32_t newsize = 0;    /* automatic sockbuf scaling */
10764         struct tcp_rack *rack;
10765 #ifdef NETFLIX_SB_LIMITS
10766         u_int mcnt, appended;
10767 #endif
10768
10769         /*
10770          * If last ACK falls within this segment's sequence numbers, record
10771          * the timestamp. NOTE that the test is modified according to the
10772          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10773          */
10774         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10775                 return (0);
10776         }
10777         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10778                 return (0);
10779         }
10780         if (tiwin && tiwin != tp->snd_wnd) {
10781                 return (0);
10782         }
10783         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10784                 return (0);
10785         }
10786         if (__predict_false((to->to_flags & TOF_TS) &&
10787             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10788                 return (0);
10789         }
10790         if (__predict_false((th->th_ack != tp->snd_una))) {
10791                 return (0);
10792         }
10793         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10794                 return (0);
10795         }
10796         if ((to->to_flags & TOF_TS) != 0 &&
10797             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10798                 tp->ts_recent_age = tcp_ts_getticks();
10799                 tp->ts_recent = to->to_tsval;
10800         }
10801         rack = (struct tcp_rack *)tp->t_fb_ptr;
10802         /*
10803          * This is a pure, in-sequence data packet with nothing on the
10804          * reassembly queue and we have enough buffer space to take it.
10805          */
10806         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10807
10808 #ifdef NETFLIX_SB_LIMITS
10809         if (so->so_rcv.sb_shlim) {
10810                 mcnt = m_memcnt(m);
10811                 appended = 0;
10812                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10813                     CFO_NOSLEEP, NULL) == false) {
10814                         counter_u64_add(tcp_sb_shlim_fails, 1);
10815                         m_freem(m);
10816                         return (1);
10817                 }
10818         }
10819 #endif
10820         /* Clean receiver SACK report if present */
10821         if (tp->rcv_numsacks)
10822                 tcp_clean_sackreport(tp);
10823         KMOD_TCPSTAT_INC(tcps_preddat);
10824         tp->rcv_nxt += tlen;
10825         if (tlen &&
10826             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10827             (tp->t_fbyte_in == 0)) {
10828                 tp->t_fbyte_in = ticks;
10829                 if (tp->t_fbyte_in == 0)
10830                         tp->t_fbyte_in = 1;
10831                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10832                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10833         }
10834         /*
10835          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10836          */
10837         tp->snd_wl1 = th->th_seq;
10838         /*
10839          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10840          */
10841         tp->rcv_up = tp->rcv_nxt;
10842         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10843         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10844         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10845
10846         /* Add data to socket buffer. */
10847         SOCKBUF_LOCK(&so->so_rcv);
10848         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10849                 m_freem(m);
10850         } else {
10851                 /*
10852                  * Set new socket buffer size. Give up when limit is
10853                  * reached.
10854                  */
10855                 if (newsize)
10856                         if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
10857                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10858                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10859 #ifdef NETFLIX_SB_LIMITS
10860                 appended =
10861 #endif
10862                         sbappendstream_locked(&so->so_rcv, m, 0);
10863                 ctf_calc_rwin(so, tp);
10864         }
10865         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10866         /* NB: sorwakeup_locked() does an implicit unlock. */
10867         sorwakeup_locked(so);
10868 #ifdef NETFLIX_SB_LIMITS
10869         if (so->so_rcv.sb_shlim && mcnt != appended)
10870                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10871 #endif
10872         rack_handle_delayed_ack(tp, rack, tlen, 0);
10873         if (tp->snd_una == tp->snd_max)
10874                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10875         return (1);
10876 }
10877
10878 /*
10879  * This subfunction is used to try to highly optimize the
10880  * fast path. We again allow window updates that are
10881  * in sequence to remain in the fast-path. We also add
10882  * in the __predict's to attempt to help the compiler.
10883  * Note that if we return a 0, then we can *not* process
10884  * it and the caller should push the packet into the
10885  * slow-path.
10886  */
10887 static int
10888 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10889     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10890     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10891 {
10892         int32_t acked;
10893         int32_t nsegs;
10894         int32_t under_pacing = 0;
10895         struct tcp_rack *rack;
10896
10897         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10898                 /* Old ack, behind (or duplicate to) the last one rcv'd */
10899                 return (0);
10900         }
10901         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10902                 /* Above what we have sent? */
10903                 return (0);
10904         }
10905         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10906                 /* We are retransmitting */
10907                 return (0);
10908         }
10909         if (__predict_false(tiwin == 0)) {
10910                 /* zero window */
10911                 return (0);
10912         }
10913         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10914                 /* We need a SYN or a FIN, unlikely.. */
10915                 return (0);
10916         }
10917         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10918                 /* Timestamp is behind .. old ack with seq wrap? */
10919                 return (0);
10920         }
10921         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10922                 /* Still recovering */
10923                 return (0);
10924         }
10925         rack = (struct tcp_rack *)tp->t_fb_ptr;
10926         if (rack->r_ctl.rc_sacked) {
10927                 /* We have sack holes on our scoreboard */
10928                 return (0);
10929         }
10930         /* Ok if we reach here, we can process a fast-ack */
10931         if (rack->gp_ready &&
10932             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10933                 under_pacing = 1;
10934         }
10935         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10936         rack_log_ack(tp, to, th, 0, 0);
10937         /* Did the window get updated? */
10938         if (tiwin != tp->snd_wnd) {
10939                 tp->snd_wnd = tiwin;
10940                 rack_validate_fo_sendwin_up(tp, rack);
10941                 tp->snd_wl1 = th->th_seq;
10942                 if (tp->snd_wnd > tp->max_sndwnd)
10943                         tp->max_sndwnd = tp->snd_wnd;
10944         }
10945         /* Do we exit persists? */
10946         if ((rack->rc_in_persist != 0) &&
10947             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10948                                rack->r_ctl.rc_pace_min_segs))) {
10949                 rack_exit_persist(tp, rack, cts);
10950         }
10951         /* Do we enter persists? */
10952         if ((rack->rc_in_persist == 0) &&
10953             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10954             TCPS_HAVEESTABLISHED(tp->t_state) &&
10955             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10956             sbavail(&tptosocket(tp)->so_snd) &&
10957             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10958                 /*
10959                  * Here the rwnd is less than
10960                  * the pacing size, we are established,
10961                  * nothing is outstanding, and there is
10962                  * data to send. Enter persists.
10963                  */
10964                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10965         }
10966         /*
10967          * If last ACK falls within this segment's sequence numbers, record
10968          * the timestamp. NOTE that the test is modified according to the
10969          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10970          */
10971         if ((to->to_flags & TOF_TS) != 0 &&
10972             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10973                 tp->ts_recent_age = tcp_ts_getticks();
10974                 tp->ts_recent = to->to_tsval;
10975         }
10976         /*
10977          * This is a pure ack for outstanding data.
10978          */
10979         KMOD_TCPSTAT_INC(tcps_predack);
10980
10981         /*
10982          * "bad retransmit" recovery.
10983          */
10984         if ((tp->t_flags & TF_PREVVALID) &&
10985             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10986                 tp->t_flags &= ~TF_PREVVALID;
10987                 if (tp->t_rxtshift == 1 &&
10988                     (int)(ticks - tp->t_badrxtwin) < 0)
10989                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10990         }
10991         /*
10992          * Recalculate the transmit timer / rtt.
10993          *
10994          * Some boxes send broken timestamp replies during the SYN+ACK
10995          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10996          * and blow up the retransmit timer.
10997          */
10998         acked = BYTES_THIS_ACK(tp, th);
10999
11000 #ifdef TCP_HHOOK
11001         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
11002         hhook_run_tcp_est_in(tp, th, to);
11003 #endif
11004         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11005         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11006         if (acked) {
11007                 struct mbuf *mfree;
11008
11009                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
11010                 SOCKBUF_LOCK(&so->so_snd);
11011                 mfree = sbcut_locked(&so->so_snd, acked);
11012                 tp->snd_una = th->th_ack;
11013                 /* Note we want to hold the sb lock through the sendmap adjust */
11014                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
11015                 /* Wake up the socket if we have room to write more */
11016                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11017                 sowwakeup_locked(so);
11018                 m_freem(mfree);
11019                 tp->t_rxtshift = 0;
11020                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11021                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11022                 rack->rc_tlp_in_progress = 0;
11023                 rack->r_ctl.rc_tlp_cnt_out = 0;
11024                 /*
11025                  * If it is the RXT timer we want to
11026                  * stop it, so we can restart a TLP.
11027                  */
11028                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11029                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11030 #ifdef NETFLIX_HTTP_LOGGING
11031                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
11032 #endif
11033         }
11034         /*
11035          * Let the congestion control algorithm update congestion control
11036          * related information. This typically means increasing the
11037          * congestion window.
11038          */
11039         if (tp->snd_wnd < ctf_outstanding(tp)) {
11040                 /* The peer collapsed the window */
11041                 rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
11042         } else if (rack->rc_has_collapsed)
11043                 rack_un_collapse_window(rack, __LINE__);
11044         if ((rack->r_collapse_point_valid) &&
11045             (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
11046                 rack->r_collapse_point_valid = 0;
11047         /*
11048          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
11049          */
11050         tp->snd_wl2 = th->th_ack;
11051         tp->t_dupacks = 0;
11052         m_freem(m);
11053         /* ND6_HINT(tp);         *//* Some progress has been made. */
11054
11055         /*
11056          * If all outstanding data are acked, stop retransmit timer,
11057          * otherwise restart timer using current (possibly backed-off)
11058          * value. If process is waiting for space, wakeup/selwakeup/signal.
11059          * If data are ready to send, let tcp_output decide between more
11060          * output or persist.
11061          */
11062         if (under_pacing &&
11063             (rack->use_fixed_rate == 0) &&
11064             (rack->in_probe_rtt == 0) &&
11065             rack->rc_gp_dyn_mul &&
11066             rack->rc_always_pace) {
11067                 /* Check if we are dragging bottom */
11068                 rack_check_bottom_drag(tp, rack, so, acked);
11069         }
11070         if (tp->snd_una == tp->snd_max) {
11071                 tp->t_flags &= ~TF_PREVVALID;
11072                 rack->r_ctl.retran_during_recovery = 0;
11073                 rack->r_ctl.dsack_byte_cnt = 0;
11074                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11075                 if (rack->r_ctl.rc_went_idle_time == 0)
11076                         rack->r_ctl.rc_went_idle_time = 1;
11077                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11078                 if (sbavail(&tptosocket(tp)->so_snd) == 0)
11079                         tp->t_acktime = 0;
11080                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11081         }
11082         if (acked && rack->r_fast_output)
11083                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11084         if (sbavail(&so->so_snd)) {
11085                 rack->r_wanted_output = 1;
11086         }
11087         return (1);
11088 }
11089
11090 /*
11091  * Return value of 1, the TCB is unlocked and most
11092  * likely gone, return value of 0, the TCP is still
11093  * locked.
11094  */
11095 static int
11096 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11097     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11098     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11099 {
11100         int32_t ret_val = 0;
11101         int32_t todrop;
11102         int32_t ourfinisacked = 0;
11103         struct tcp_rack *rack;
11104
11105         INP_WLOCK_ASSERT(tptoinpcb(tp));
11106
11107         ctf_calc_rwin(so, tp);
11108         /*
11109          * If the state is SYN_SENT: if seg contains an ACK, but not for our
11110          * SYN, drop the input. if seg contains a RST, then drop the
11111          * connection. if seg does not contain SYN, then drop it. Otherwise
11112          * this is an acceptable SYN segment initialize tp->rcv_nxt and
11113          * tp->irs if seg contains ack then advance tp->snd_una if seg
11114          * contains an ECE and ECN support is enabled, the stream is ECN
11115          * capable. if SYN has been acked change to ESTABLISHED else
11116          * SYN_RCVD state arrange for segment to be acked (eventually)
11117          * continue processing rest of data/controls.
11118          */
11119         if ((thflags & TH_ACK) &&
11120             (SEQ_LEQ(th->th_ack, tp->iss) ||
11121             SEQ_GT(th->th_ack, tp->snd_max))) {
11122                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11123                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11124                 return (1);
11125         }
11126         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11127                 TCP_PROBE5(connect__refused, NULL, tp,
11128                     mtod(m, const char *), tp, th);
11129                 tp = tcp_drop(tp, ECONNREFUSED);
11130                 ctf_do_drop(m, tp);
11131                 return (1);
11132         }
11133         if (thflags & TH_RST) {
11134                 ctf_do_drop(m, tp);
11135                 return (1);
11136         }
11137         if (!(thflags & TH_SYN)) {
11138                 ctf_do_drop(m, tp);
11139                 return (1);
11140         }
11141         tp->irs = th->th_seq;
11142         tcp_rcvseqinit(tp);
11143         rack = (struct tcp_rack *)tp->t_fb_ptr;
11144         if (thflags & TH_ACK) {
11145                 int tfo_partial = 0;
11146
11147                 KMOD_TCPSTAT_INC(tcps_connects);
11148                 soisconnected(so);
11149 #ifdef MAC
11150                 mac_socketpeer_set_from_mbuf(m, so);
11151 #endif
11152                 /* Do window scaling on this connection? */
11153                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11154                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11155                         tp->rcv_scale = tp->request_r_scale;
11156                 }
11157                 tp->rcv_adv += min(tp->rcv_wnd,
11158                     TCP_MAXWIN << tp->rcv_scale);
11159                 /*
11160                  * If not all the data that was sent in the TFO SYN
11161                  * has been acked, resend the remainder right away.
11162                  */
11163                 if (IS_FASTOPEN(tp->t_flags) &&
11164                     (tp->snd_una != tp->snd_max)) {
11165                         tp->snd_nxt = th->th_ack;
11166                         tfo_partial = 1;
11167                 }
11168                 /*
11169                  * If there's data, delay ACK; if there's also a FIN ACKNOW
11170                  * will be turned on later.
11171                  */
11172                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11173                         rack_timer_cancel(tp, rack,
11174                                           rack->r_ctl.rc_rcvtime, __LINE__);
11175                         tp->t_flags |= TF_DELACK;
11176                 } else {
11177                         rack->r_wanted_output = 1;
11178                         tp->t_flags |= TF_ACKNOW;
11179                         rack->rc_dack_toggle = 0;
11180                 }
11181
11182                 tcp_ecn_input_syn_sent(tp, thflags, iptos);
11183
11184                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
11185                         /*
11186                          * We advance snd_una for the
11187                          * fast open case. If th_ack is
11188                          * acknowledging data beyond
11189                          * snd_una we can't just call
11190                          * ack-processing since the
11191                          * data stream in our send-map
11192                          * will start at snd_una + 1 (one
11193                          * beyond the SYN). If its just
11194                          * equal we don't need to do that
11195                          * and there is no send_map.
11196                          */
11197                         tp->snd_una++;
11198                 }
11199                 /*
11200                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11201                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11202                  */
11203                 tp->t_starttime = ticks;
11204                 if (tp->t_flags & TF_NEEDFIN) {
11205                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
11206                         tp->t_flags &= ~TF_NEEDFIN;
11207                         thflags &= ~TH_SYN;
11208                 } else {
11209                         tcp_state_change(tp, TCPS_ESTABLISHED);
11210                         TCP_PROBE5(connect__established, NULL, tp,
11211                             mtod(m, const char *), tp, th);
11212                         rack_cc_conn_init(tp);
11213                 }
11214         } else {
11215                 /*
11216                  * Received initial SYN in SYN-SENT[*] state => simultaneous
11217                  * open.  If segment contains CC option and there is a
11218                  * cached CC, apply TAO test. If it succeeds, connection is *
11219                  * half-synchronized. Otherwise, do 3-way handshake:
11220                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11221                  * there was no CC option, clear cached CC value.
11222                  */
11223                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
11224                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
11225         }
11226         /*
11227          * Advance th->th_seq to correspond to first data byte. If data,
11228          * trim to stay within window, dropping FIN if necessary.
11229          */
11230         th->th_seq++;
11231         if (tlen > tp->rcv_wnd) {
11232                 todrop = tlen - tp->rcv_wnd;
11233                 m_adj(m, -todrop);
11234                 tlen = tp->rcv_wnd;
11235                 thflags &= ~TH_FIN;
11236                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11237                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11238         }
11239         tp->snd_wl1 = th->th_seq - 1;
11240         tp->rcv_up = th->th_seq;
11241         /*
11242          * Client side of transaction: already sent SYN and data. If the
11243          * remote host used T/TCP to validate the SYN, our data will be
11244          * ACK'd; if so, enter normal data segment processing in the middle
11245          * of step 5, ack processing. Otherwise, goto step 6.
11246          */
11247         if (thflags & TH_ACK) {
11248                 /* For syn-sent we need to possibly update the rtt */
11249                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11250                         uint32_t t, mcts;
11251
11252                         mcts = tcp_ts_getticks();
11253                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11254                         if (!tp->t_rttlow || tp->t_rttlow > t)
11255                                 tp->t_rttlow = t;
11256                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11257                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11258                         tcp_rack_xmit_timer_commit(rack, tp);
11259                 }
11260                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11261                         return (ret_val);
11262                 /* We may have changed to FIN_WAIT_1 above */
11263                 if (tp->t_state == TCPS_FIN_WAIT_1) {
11264                         /*
11265                          * In FIN_WAIT_1 STATE in addition to the processing
11266                          * for the ESTABLISHED state if our FIN is now
11267                          * acknowledged then enter FIN_WAIT_2.
11268                          */
11269                         if (ourfinisacked) {
11270                                 /*
11271                                  * If we can't receive any more data, then
11272                                  * closing user can proceed. Starting the
11273                                  * timer is contrary to the specification,
11274                                  * but if we don't get a FIN we'll hang
11275                                  * forever.
11276                                  *
11277                                  * XXXjl: we should release the tp also, and
11278                                  * use a compressed state.
11279                                  */
11280                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11281                                         soisdisconnected(so);
11282                                         tcp_timer_activate(tp, TT_2MSL,
11283                                             (tcp_fast_finwait2_recycle ?
11284                                             tcp_finwait2_timeout :
11285                                             TP_MAXIDLE(tp)));
11286                                 }
11287                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11288                         }
11289                 }
11290         }
11291         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11292            tiwin, thflags, nxt_pkt));
11293 }
11294
11295 /*
11296  * Return value of 1, the TCB is unlocked and most
11297  * likely gone, return value of 0, the TCP is still
11298  * locked.
11299  */
11300 static int
11301 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11302     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11303     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11304 {
11305         struct tcp_rack *rack;
11306         int32_t ret_val = 0;
11307         int32_t ourfinisacked = 0;
11308
11309         ctf_calc_rwin(so, tp);
11310         if ((thflags & TH_ACK) &&
11311             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11312             SEQ_GT(th->th_ack, tp->snd_max))) {
11313                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11314                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11315                 return (1);
11316         }
11317         rack = (struct tcp_rack *)tp->t_fb_ptr;
11318         if (IS_FASTOPEN(tp->t_flags)) {
11319                 /*
11320                  * When a TFO connection is in SYN_RECEIVED, the
11321                  * only valid packets are the initial SYN, a
11322                  * retransmit/copy of the initial SYN (possibly with
11323                  * a subset of the original data), a valid ACK, a
11324                  * FIN, or a RST.
11325                  */
11326                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11327                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11328                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11329                         return (1);
11330                 } else if (thflags & TH_SYN) {
11331                         /* non-initial SYN is ignored */
11332                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11333                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11334                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11335                                 ctf_do_drop(m, NULL);
11336                                 return (0);
11337                         }
11338                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11339                         ctf_do_drop(m, NULL);
11340                         return (0);
11341                 }
11342         }
11343
11344         if ((thflags & TH_RST) ||
11345             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11346                 return (__ctf_process_rst(m, th, so, tp,
11347                                           &rack->r_ctl.challenge_ack_ts,
11348                                           &rack->r_ctl.challenge_ack_cnt));
11349         /*
11350          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11351          * it's less than ts_recent, drop it.
11352          */
11353         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11354             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11355                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11356                         return (ret_val);
11357         }
11358         /*
11359          * In the SYN-RECEIVED state, validate that the packet belongs to
11360          * this connection before trimming the data to fit the receive
11361          * window.  Check the sequence number versus IRS since we know the
11362          * sequence numbers haven't wrapped.  This is a partial fix for the
11363          * "LAND" DoS attack.
11364          */
11365         if (SEQ_LT(th->th_seq, tp->irs)) {
11366                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11367                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11368                 return (1);
11369         }
11370         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11371                               &rack->r_ctl.challenge_ack_ts,
11372                               &rack->r_ctl.challenge_ack_cnt)) {
11373                 return (ret_val);
11374         }
11375         /*
11376          * If last ACK falls within this segment's sequence numbers, record
11377          * its timestamp. NOTE: 1) That the test incorporates suggestions
11378          * from the latest proposal of the tcplw@cray.com list (Braden
11379          * 1993/04/26). 2) That updating only on newer timestamps interferes
11380          * with our earlier PAWS tests, so this check should be solely
11381          * predicated on the sequence space of this segment. 3) That we
11382          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11383          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11384          * SEG.Len, This modified check allows us to overcome RFC1323's
11385          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11386          * p.869. In such cases, we can still calculate the RTT correctly
11387          * when RCV.NXT == Last.ACK.Sent.
11388          */
11389         if ((to->to_flags & TOF_TS) != 0 &&
11390             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11391             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11392             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11393                 tp->ts_recent_age = tcp_ts_getticks();
11394                 tp->ts_recent = to->to_tsval;
11395         }
11396         tp->snd_wnd = tiwin;
11397         rack_validate_fo_sendwin_up(tp, rack);
11398         /*
11399          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11400          * is on (half-synchronized state), then queue data for later
11401          * processing; else drop segment and return.
11402          */
11403         if ((thflags & TH_ACK) == 0) {
11404                 if (IS_FASTOPEN(tp->t_flags)) {
11405                         rack_cc_conn_init(tp);
11406                 }
11407                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11408                     tiwin, thflags, nxt_pkt));
11409         }
11410         KMOD_TCPSTAT_INC(tcps_connects);
11411         if (tp->t_flags & TF_SONOTCONN) {
11412                 tp->t_flags &= ~TF_SONOTCONN;
11413                 soisconnected(so);
11414         }
11415         /* Do window scaling? */
11416         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11417             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11418                 tp->rcv_scale = tp->request_r_scale;
11419         }
11420         /*
11421          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11422          * FIN-WAIT-1
11423          */
11424         tp->t_starttime = ticks;
11425         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11426                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11427                 tp->t_tfo_pending = NULL;
11428         }
11429         if (tp->t_flags & TF_NEEDFIN) {
11430                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11431                 tp->t_flags &= ~TF_NEEDFIN;
11432         } else {
11433                 tcp_state_change(tp, TCPS_ESTABLISHED);
11434                 TCP_PROBE5(accept__established, NULL, tp,
11435                     mtod(m, const char *), tp, th);
11436                 /*
11437                  * TFO connections call cc_conn_init() during SYN
11438                  * processing.  Calling it again here for such connections
11439                  * is not harmless as it would undo the snd_cwnd reduction
11440                  * that occurs when a TFO SYN|ACK is retransmitted.
11441                  */
11442                 if (!IS_FASTOPEN(tp->t_flags))
11443                         rack_cc_conn_init(tp);
11444         }
11445         /*
11446          * Account for the ACK of our SYN prior to
11447          * regular ACK processing below, except for
11448          * simultaneous SYN, which is handled later.
11449          */
11450         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11451                 tp->snd_una++;
11452         /*
11453          * If segment contains data or ACK, will call tcp_reass() later; if
11454          * not, do so now to pass queued data to user.
11455          */
11456         if (tlen == 0 && (thflags & TH_FIN) == 0) {
11457                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11458                     (struct mbuf *)0);
11459                 if (tp->t_flags & TF_WAKESOR) {
11460                         tp->t_flags &= ~TF_WAKESOR;
11461                         /* NB: sorwakeup_locked() does an implicit unlock. */
11462                         sorwakeup_locked(so);
11463                 }
11464         }
11465         tp->snd_wl1 = th->th_seq - 1;
11466         /* For syn-recv we need to possibly update the rtt */
11467         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11468                 uint32_t t, mcts;
11469
11470                 mcts = tcp_ts_getticks();
11471                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11472                 if (!tp->t_rttlow || tp->t_rttlow > t)
11473                         tp->t_rttlow = t;
11474                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11475                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11476                 tcp_rack_xmit_timer_commit(rack, tp);
11477         }
11478         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11479                 return (ret_val);
11480         }
11481         if (tp->t_state == TCPS_FIN_WAIT_1) {
11482                 /* We could have went to FIN_WAIT_1 (or EST) above */
11483                 /*
11484                  * In FIN_WAIT_1 STATE in addition to the processing for the
11485                  * ESTABLISHED state if our FIN is now acknowledged then
11486                  * enter FIN_WAIT_2.
11487                  */
11488                 if (ourfinisacked) {
11489                         /*
11490                          * If we can't receive any more data, then closing
11491                          * user can proceed. Starting the timer is contrary
11492                          * to the specification, but if we don't get a FIN
11493                          * we'll hang forever.
11494                          *
11495                          * XXXjl: we should release the tp also, and use a
11496                          * compressed state.
11497                          */
11498                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11499                                 soisdisconnected(so);
11500                                 tcp_timer_activate(tp, TT_2MSL,
11501                                     (tcp_fast_finwait2_recycle ?
11502                                     tcp_finwait2_timeout :
11503                                     TP_MAXIDLE(tp)));
11504                         }
11505                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11506                 }
11507         }
11508         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11509             tiwin, thflags, nxt_pkt));
11510 }
11511
11512 /*
11513  * Return value of 1, the TCB is unlocked and most
11514  * likely gone, return value of 0, the TCP is still
11515  * locked.
11516  */
11517 static int
11518 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11519     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11520     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11521 {
11522         int32_t ret_val = 0;
11523         struct tcp_rack *rack;
11524
11525         /*
11526          * Header prediction: check for the two common cases of a
11527          * uni-directional data xfer.  If the packet has no control flags,
11528          * is in-sequence, the window didn't change and we're not
11529          * retransmitting, it's a candidate.  If the length is zero and the
11530          * ack moved forward, we're the sender side of the xfer.  Just free
11531          * the data acked & wake any higher level process that was blocked
11532          * waiting for space.  If the length is non-zero and the ack didn't
11533          * move, we're the receiver side.  If we're getting packets in-order
11534          * (the reassembly queue is empty), add the data toc The socket
11535          * buffer and note that we need a delayed ack. Make sure that the
11536          * hidden state-flags are also off. Since we check for
11537          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11538          */
11539         rack = (struct tcp_rack *)tp->t_fb_ptr;
11540         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11541             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11542             __predict_true(SEGQ_EMPTY(tp)) &&
11543             __predict_true(th->th_seq == tp->rcv_nxt)) {
11544                 if (tlen == 0) {
11545                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11546                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11547                                 return (0);
11548                         }
11549                 } else {
11550                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11551                             tiwin, nxt_pkt, iptos)) {
11552                                 return (0);
11553                         }
11554                 }
11555         }
11556         ctf_calc_rwin(so, tp);
11557
11558         if ((thflags & TH_RST) ||
11559             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11560                 return (__ctf_process_rst(m, th, so, tp,
11561                                           &rack->r_ctl.challenge_ack_ts,
11562                                           &rack->r_ctl.challenge_ack_cnt));
11563
11564         /*
11565          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11566          * synchronized state.
11567          */
11568         if (thflags & TH_SYN) {
11569                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11570                 return (ret_val);
11571         }
11572         /*
11573          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11574          * it's less than ts_recent, drop it.
11575          */
11576         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11577             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11578                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11579                         return (ret_val);
11580         }
11581         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11582                               &rack->r_ctl.challenge_ack_ts,
11583                               &rack->r_ctl.challenge_ack_cnt)) {
11584                 return (ret_val);
11585         }
11586         /*
11587          * If last ACK falls within this segment's sequence numbers, record
11588          * its timestamp. NOTE: 1) That the test incorporates suggestions
11589          * from the latest proposal of the tcplw@cray.com list (Braden
11590          * 1993/04/26). 2) That updating only on newer timestamps interferes
11591          * with our earlier PAWS tests, so this check should be solely
11592          * predicated on the sequence space of this segment. 3) That we
11593          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11594          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11595          * SEG.Len, This modified check allows us to overcome RFC1323's
11596          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11597          * p.869. In such cases, we can still calculate the RTT correctly
11598          * when RCV.NXT == Last.ACK.Sent.
11599          */
11600         if ((to->to_flags & TOF_TS) != 0 &&
11601             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11602             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11603             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11604                 tp->ts_recent_age = tcp_ts_getticks();
11605                 tp->ts_recent = to->to_tsval;
11606         }
11607         /*
11608          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11609          * is on (half-synchronized state), then queue data for later
11610          * processing; else drop segment and return.
11611          */
11612         if ((thflags & TH_ACK) == 0) {
11613                 if (tp->t_flags & TF_NEEDSYN) {
11614                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11615                             tiwin, thflags, nxt_pkt));
11616
11617                 } else if (tp->t_flags & TF_ACKNOW) {
11618                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11619                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11620                         return (ret_val);
11621                 } else {
11622                         ctf_do_drop(m, NULL);
11623                         return (0);
11624                 }
11625         }
11626         /*
11627          * Ack processing.
11628          */
11629         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11630                 return (ret_val);
11631         }
11632         if (sbavail(&so->so_snd)) {
11633                 if (ctf_progress_timeout_check(tp, true)) {
11634                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11635                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11636                         return (1);
11637                 }
11638         }
11639         /* State changes only happen in rack_process_data() */
11640         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11641             tiwin, thflags, nxt_pkt));
11642 }
11643
11644 /*
11645  * Return value of 1, the TCB is unlocked and most
11646  * likely gone, return value of 0, the TCP is still
11647  * locked.
11648  */
11649 static int
11650 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11651     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11652     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11653 {
11654         int32_t ret_val = 0;
11655         struct tcp_rack *rack;
11656
11657         rack = (struct tcp_rack *)tp->t_fb_ptr;
11658         ctf_calc_rwin(so, tp);
11659         if ((thflags & TH_RST) ||
11660             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11661                 return (__ctf_process_rst(m, th, so, tp,
11662                                           &rack->r_ctl.challenge_ack_ts,
11663                                           &rack->r_ctl.challenge_ack_cnt));
11664         /*
11665          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11666          * synchronized state.
11667          */
11668         if (thflags & TH_SYN) {
11669                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11670                 return (ret_val);
11671         }
11672         /*
11673          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11674          * it's less than ts_recent, drop it.
11675          */
11676         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11677             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11678                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11679                         return (ret_val);
11680         }
11681         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11682                               &rack->r_ctl.challenge_ack_ts,
11683                               &rack->r_ctl.challenge_ack_cnt)) {
11684                 return (ret_val);
11685         }
11686         /*
11687          * If last ACK falls within this segment's sequence numbers, record
11688          * its timestamp. NOTE: 1) That the test incorporates suggestions
11689          * from the latest proposal of the tcplw@cray.com list (Braden
11690          * 1993/04/26). 2) That updating only on newer timestamps interferes
11691          * with our earlier PAWS tests, so this check should be solely
11692          * predicated on the sequence space of this segment. 3) That we
11693          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11694          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11695          * SEG.Len, This modified check allows us to overcome RFC1323's
11696          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11697          * p.869. In such cases, we can still calculate the RTT correctly
11698          * when RCV.NXT == Last.ACK.Sent.
11699          */
11700         if ((to->to_flags & TOF_TS) != 0 &&
11701             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11702             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11703             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11704                 tp->ts_recent_age = tcp_ts_getticks();
11705                 tp->ts_recent = to->to_tsval;
11706         }
11707         /*
11708          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11709          * is on (half-synchronized state), then queue data for later
11710          * processing; else drop segment and return.
11711          */
11712         if ((thflags & TH_ACK) == 0) {
11713                 if (tp->t_flags & TF_NEEDSYN) {
11714                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11715                             tiwin, thflags, nxt_pkt));
11716
11717                 } else if (tp->t_flags & TF_ACKNOW) {
11718                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11719                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11720                         return (ret_val);
11721                 } else {
11722                         ctf_do_drop(m, NULL);
11723                         return (0);
11724                 }
11725         }
11726         /*
11727          * Ack processing.
11728          */
11729         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11730                 return (ret_val);
11731         }
11732         if (sbavail(&so->so_snd)) {
11733                 if (ctf_progress_timeout_check(tp, true)) {
11734                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11735                                                 tp, tick, PROGRESS_DROP, __LINE__);
11736                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11737                         return (1);
11738                 }
11739         }
11740         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11741             tiwin, thflags, nxt_pkt));
11742 }
11743
11744 static int
11745 rack_check_data_after_close(struct mbuf *m,
11746     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11747 {
11748         struct tcp_rack *rack;
11749
11750         rack = (struct tcp_rack *)tp->t_fb_ptr;
11751         if (rack->rc_allow_data_af_clo == 0) {
11752         close_now:
11753                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11754                 /* tcp_close will kill the inp pre-log the Reset */
11755                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11756                 tp = tcp_close(tp);
11757                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11758                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11759                 return (1);
11760         }
11761         if (sbavail(&so->so_snd) == 0)
11762                 goto close_now;
11763         /* Ok we allow data that is ignored and a followup reset */
11764         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11765         tp->rcv_nxt = th->th_seq + *tlen;
11766         tp->t_flags2 |= TF2_DROP_AF_DATA;
11767         rack->r_wanted_output = 1;
11768         *tlen = 0;
11769         return (0);
11770 }
11771
11772 /*
11773  * Return value of 1, the TCB is unlocked and most
11774  * likely gone, return value of 0, the TCP is still
11775  * locked.
11776  */
11777 static int
11778 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11779     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11780     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11781 {
11782         int32_t ret_val = 0;
11783         int32_t ourfinisacked = 0;
11784         struct tcp_rack *rack;
11785
11786         rack = (struct tcp_rack *)tp->t_fb_ptr;
11787         ctf_calc_rwin(so, tp);
11788
11789         if ((thflags & TH_RST) ||
11790             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11791                 return (__ctf_process_rst(m, th, so, tp,
11792                                           &rack->r_ctl.challenge_ack_ts,
11793                                           &rack->r_ctl.challenge_ack_cnt));
11794         /*
11795          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11796          * synchronized state.
11797          */
11798         if (thflags & TH_SYN) {
11799                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11800                 return (ret_val);
11801         }
11802         /*
11803          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11804          * it's less than ts_recent, drop it.
11805          */
11806         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11807             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11808                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11809                         return (ret_val);
11810         }
11811         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11812                               &rack->r_ctl.challenge_ack_ts,
11813                               &rack->r_ctl.challenge_ack_cnt)) {
11814                 return (ret_val);
11815         }
11816         /*
11817          * If new data are received on a connection after the user processes
11818          * are gone, then RST the other end.
11819          */
11820         if ((tp->t_flags & TF_CLOSED) && tlen &&
11821             rack_check_data_after_close(m, tp, &tlen, th, so))
11822                 return (1);
11823         /*
11824          * If last ACK falls within this segment's sequence numbers, record
11825          * its timestamp. NOTE: 1) That the test incorporates suggestions
11826          * from the latest proposal of the tcplw@cray.com list (Braden
11827          * 1993/04/26). 2) That updating only on newer timestamps interferes
11828          * with our earlier PAWS tests, so this check should be solely
11829          * predicated on the sequence space of this segment. 3) That we
11830          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11831          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11832          * SEG.Len, This modified check allows us to overcome RFC1323's
11833          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11834          * p.869. In such cases, we can still calculate the RTT correctly
11835          * when RCV.NXT == Last.ACK.Sent.
11836          */
11837         if ((to->to_flags & TOF_TS) != 0 &&
11838             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11839             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11840             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11841                 tp->ts_recent_age = tcp_ts_getticks();
11842                 tp->ts_recent = to->to_tsval;
11843         }
11844         /*
11845          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11846          * is on (half-synchronized state), then queue data for later
11847          * processing; else drop segment and return.
11848          */
11849         if ((thflags & TH_ACK) == 0) {
11850                 if (tp->t_flags & TF_NEEDSYN) {
11851                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11852                             tiwin, thflags, nxt_pkt));
11853                 } else if (tp->t_flags & TF_ACKNOW) {
11854                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11855                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11856                         return (ret_val);
11857                 } else {
11858                         ctf_do_drop(m, NULL);
11859                         return (0);
11860                 }
11861         }
11862         /*
11863          * Ack processing.
11864          */
11865         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11866                 return (ret_val);
11867         }
11868         if (ourfinisacked) {
11869                 /*
11870                  * If we can't receive any more data, then closing user can
11871                  * proceed. Starting the timer is contrary to the
11872                  * specification, but if we don't get a FIN we'll hang
11873                  * forever.
11874                  *
11875                  * XXXjl: we should release the tp also, and use a
11876                  * compressed state.
11877                  */
11878                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11879                         soisdisconnected(so);
11880                         tcp_timer_activate(tp, TT_2MSL,
11881                             (tcp_fast_finwait2_recycle ?
11882                             tcp_finwait2_timeout :
11883                             TP_MAXIDLE(tp)));
11884                 }
11885                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11886         }
11887         if (sbavail(&so->so_snd)) {
11888                 if (ctf_progress_timeout_check(tp, true)) {
11889                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11890                                                 tp, tick, PROGRESS_DROP, __LINE__);
11891                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11892                         return (1);
11893                 }
11894         }
11895         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11896             tiwin, thflags, nxt_pkt));
11897 }
11898
11899 /*
11900  * Return value of 1, the TCB is unlocked and most
11901  * likely gone, return value of 0, the TCP is still
11902  * locked.
11903  */
11904 static int
11905 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11906     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11907     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11908 {
11909         int32_t ret_val = 0;
11910         int32_t ourfinisacked = 0;
11911         struct tcp_rack *rack;
11912
11913         rack = (struct tcp_rack *)tp->t_fb_ptr;
11914         ctf_calc_rwin(so, tp);
11915
11916         if ((thflags & TH_RST) ||
11917             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11918                 return (__ctf_process_rst(m, th, so, tp,
11919                                           &rack->r_ctl.challenge_ack_ts,
11920                                           &rack->r_ctl.challenge_ack_cnt));
11921         /*
11922          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11923          * synchronized state.
11924          */
11925         if (thflags & TH_SYN) {
11926                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11927                 return (ret_val);
11928         }
11929         /*
11930          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11931          * it's less than ts_recent, drop it.
11932          */
11933         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11934             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11935                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11936                         return (ret_val);
11937         }
11938         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11939                               &rack->r_ctl.challenge_ack_ts,
11940                               &rack->r_ctl.challenge_ack_cnt)) {
11941                 return (ret_val);
11942         }
11943         /*
11944          * If new data are received on a connection after the user processes
11945          * are gone, then RST the other end.
11946          */
11947         if ((tp->t_flags & TF_CLOSED) && tlen &&
11948             rack_check_data_after_close(m, tp, &tlen, th, so))
11949                 return (1);
11950         /*
11951          * If last ACK falls within this segment's sequence numbers, record
11952          * its timestamp. NOTE: 1) That the test incorporates suggestions
11953          * from the latest proposal of the tcplw@cray.com list (Braden
11954          * 1993/04/26). 2) That updating only on newer timestamps interferes
11955          * with our earlier PAWS tests, so this check should be solely
11956          * predicated on the sequence space of this segment. 3) That we
11957          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11958          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11959          * SEG.Len, This modified check allows us to overcome RFC1323's
11960          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11961          * p.869. In such cases, we can still calculate the RTT correctly
11962          * when RCV.NXT == Last.ACK.Sent.
11963          */
11964         if ((to->to_flags & TOF_TS) != 0 &&
11965             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11966             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11967             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11968                 tp->ts_recent_age = tcp_ts_getticks();
11969                 tp->ts_recent = to->to_tsval;
11970         }
11971         /*
11972          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11973          * is on (half-synchronized state), then queue data for later
11974          * processing; else drop segment and return.
11975          */
11976         if ((thflags & TH_ACK) == 0) {
11977                 if (tp->t_flags & TF_NEEDSYN) {
11978                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11979                             tiwin, thflags, nxt_pkt));
11980                 } else if (tp->t_flags & TF_ACKNOW) {
11981                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11982                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11983                         return (ret_val);
11984                 } else {
11985                         ctf_do_drop(m, NULL);
11986                         return (0);
11987                 }
11988         }
11989         /*
11990          * Ack processing.
11991          */
11992         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11993                 return (ret_val);
11994         }
11995         if (ourfinisacked) {
11996                 tcp_twstart(tp);
11997                 m_freem(m);
11998                 return (1);
11999         }
12000         if (sbavail(&so->so_snd)) {
12001                 if (ctf_progress_timeout_check(tp, true)) {
12002                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12003                                                 tp, tick, PROGRESS_DROP, __LINE__);
12004                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12005                         return (1);
12006                 }
12007         }
12008         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12009             tiwin, thflags, nxt_pkt));
12010 }
12011
12012 /*
12013  * Return value of 1, the TCB is unlocked and most
12014  * likely gone, return value of 0, the TCP is still
12015  * locked.
12016  */
12017 static int
12018 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12019     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12020     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12021 {
12022         int32_t ret_val = 0;
12023         int32_t ourfinisacked = 0;
12024         struct tcp_rack *rack;
12025
12026         rack = (struct tcp_rack *)tp->t_fb_ptr;
12027         ctf_calc_rwin(so, tp);
12028
12029         if ((thflags & TH_RST) ||
12030             (tp->t_fin_is_rst && (thflags & TH_FIN)))
12031                 return (__ctf_process_rst(m, th, so, tp,
12032                                           &rack->r_ctl.challenge_ack_ts,
12033                                           &rack->r_ctl.challenge_ack_cnt));
12034         /*
12035          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12036          * synchronized state.
12037          */
12038         if (thflags & TH_SYN) {
12039                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12040                 return (ret_val);
12041         }
12042         /*
12043          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12044          * it's less than ts_recent, drop it.
12045          */
12046         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12047             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12048                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12049                         return (ret_val);
12050         }
12051         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12052                               &rack->r_ctl.challenge_ack_ts,
12053                               &rack->r_ctl.challenge_ack_cnt)) {
12054                 return (ret_val);
12055         }
12056         /*
12057          * If new data are received on a connection after the user processes
12058          * are gone, then RST the other end.
12059          */
12060         if ((tp->t_flags & TF_CLOSED) && tlen &&
12061             rack_check_data_after_close(m, tp, &tlen, th, so))
12062                 return (1);
12063         /*
12064          * If last ACK falls within this segment's sequence numbers, record
12065          * its timestamp. NOTE: 1) That the test incorporates suggestions
12066          * from the latest proposal of the tcplw@cray.com list (Braden
12067          * 1993/04/26). 2) That updating only on newer timestamps interferes
12068          * with our earlier PAWS tests, so this check should be solely
12069          * predicated on the sequence space of this segment. 3) That we
12070          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12071          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12072          * SEG.Len, This modified check allows us to overcome RFC1323's
12073          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12074          * p.869. In such cases, we can still calculate the RTT correctly
12075          * when RCV.NXT == Last.ACK.Sent.
12076          */
12077         if ((to->to_flags & TOF_TS) != 0 &&
12078             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12079             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12080             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12081                 tp->ts_recent_age = tcp_ts_getticks();
12082                 tp->ts_recent = to->to_tsval;
12083         }
12084         /*
12085          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12086          * is on (half-synchronized state), then queue data for later
12087          * processing; else drop segment and return.
12088          */
12089         if ((thflags & TH_ACK) == 0) {
12090                 if (tp->t_flags & TF_NEEDSYN) {
12091                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12092                             tiwin, thflags, nxt_pkt));
12093                 } else if (tp->t_flags & TF_ACKNOW) {
12094                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12095                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12096                         return (ret_val);
12097                 } else {
12098                         ctf_do_drop(m, NULL);
12099                         return (0);
12100                 }
12101         }
12102         /*
12103          * case TCPS_LAST_ACK: Ack processing.
12104          */
12105         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12106                 return (ret_val);
12107         }
12108         if (ourfinisacked) {
12109                 tp = tcp_close(tp);
12110                 ctf_do_drop(m, tp);
12111                 return (1);
12112         }
12113         if (sbavail(&so->so_snd)) {
12114                 if (ctf_progress_timeout_check(tp, true)) {
12115                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12116                                                 tp, tick, PROGRESS_DROP, __LINE__);
12117                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12118                         return (1);
12119                 }
12120         }
12121         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12122             tiwin, thflags, nxt_pkt));
12123 }
12124
12125 /*
12126  * Return value of 1, the TCB is unlocked and most
12127  * likely gone, return value of 0, the TCP is still
12128  * locked.
12129  */
12130 static int
12131 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12132     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12133     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12134 {
12135         int32_t ret_val = 0;
12136         int32_t ourfinisacked = 0;
12137         struct tcp_rack *rack;
12138
12139         rack = (struct tcp_rack *)tp->t_fb_ptr;
12140         ctf_calc_rwin(so, tp);
12141
12142         /* Reset receive buffer auto scaling when not in bulk receive mode. */
12143         if ((thflags & TH_RST) ||
12144             (tp->t_fin_is_rst && (thflags & TH_FIN)))
12145                 return (__ctf_process_rst(m, th, so, tp,
12146                                           &rack->r_ctl.challenge_ack_ts,
12147                                           &rack->r_ctl.challenge_ack_cnt));
12148         /*
12149          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12150          * synchronized state.
12151          */
12152         if (thflags & TH_SYN) {
12153                 ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12154                 return (ret_val);
12155         }
12156         /*
12157          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12158          * it's less than ts_recent, drop it.
12159          */
12160         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12161             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12162                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12163                         return (ret_val);
12164         }
12165         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12166                               &rack->r_ctl.challenge_ack_ts,
12167                               &rack->r_ctl.challenge_ack_cnt)) {
12168                 return (ret_val);
12169         }
12170         /*
12171          * If new data are received on a connection after the user processes
12172          * are gone, then RST the other end.
12173          */
12174         if ((tp->t_flags & TF_CLOSED) && tlen &&
12175             rack_check_data_after_close(m, tp, &tlen, th, so))
12176                 return (1);
12177         /*
12178          * If last ACK falls within this segment's sequence numbers, record
12179          * its timestamp. NOTE: 1) That the test incorporates suggestions
12180          * from the latest proposal of the tcplw@cray.com list (Braden
12181          * 1993/04/26). 2) That updating only on newer timestamps interferes
12182          * with our earlier PAWS tests, so this check should be solely
12183          * predicated on the sequence space of this segment. 3) That we
12184          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12185          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12186          * SEG.Len, This modified check allows us to overcome RFC1323's
12187          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12188          * p.869. In such cases, we can still calculate the RTT correctly
12189          * when RCV.NXT == Last.ACK.Sent.
12190          */
12191         if ((to->to_flags & TOF_TS) != 0 &&
12192             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12193             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12194             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12195                 tp->ts_recent_age = tcp_ts_getticks();
12196                 tp->ts_recent = to->to_tsval;
12197         }
12198         /*
12199          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12200          * is on (half-synchronized state), then queue data for later
12201          * processing; else drop segment and return.
12202          */
12203         if ((thflags & TH_ACK) == 0) {
12204                 if (tp->t_flags & TF_NEEDSYN) {
12205                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12206                             tiwin, thflags, nxt_pkt));
12207                 } else if (tp->t_flags & TF_ACKNOW) {
12208                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12209                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12210                         return (ret_val);
12211                 } else {
12212                         ctf_do_drop(m, NULL);
12213                         return (0);
12214                 }
12215         }
12216         /*
12217          * Ack processing.
12218          */
12219         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12220                 return (ret_val);
12221         }
12222         if (sbavail(&so->so_snd)) {
12223                 if (ctf_progress_timeout_check(tp, true)) {
12224                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12225                                                 tp, tick, PROGRESS_DROP, __LINE__);
12226                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12227                         return (1);
12228                 }
12229         }
12230         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12231             tiwin, thflags, nxt_pkt));
12232 }
12233
12234 static void inline
12235 rack_clear_rate_sample(struct tcp_rack *rack)
12236 {
12237         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12238         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12239         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12240 }
12241
12242 static void
12243 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12244 {
12245         uint64_t bw_est, rate_wanted;
12246         int chged = 0;
12247         uint32_t user_max, orig_min, orig_max;
12248
12249         orig_min = rack->r_ctl.rc_pace_min_segs;
12250         orig_max = rack->r_ctl.rc_pace_max_segs;
12251         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12252         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12253                 chged = 1;
12254         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12255         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12256                 if (user_max != rack->r_ctl.rc_pace_max_segs)
12257                         chged = 1;
12258         }
12259         if (rack->rc_force_max_seg) {
12260                 rack->r_ctl.rc_pace_max_segs = user_max;
12261         } else if (rack->use_fixed_rate) {
12262                 bw_est = rack_get_bw(rack);
12263                 if ((rack->r_ctl.crte == NULL) ||
12264                     (bw_est != rack->r_ctl.crte->rate)) {
12265                         rack->r_ctl.rc_pace_max_segs = user_max;
12266                 } else {
12267                         /* We are pacing right at the hardware rate */
12268                         uint32_t segsiz;
12269
12270                         segsiz = min(ctf_fixed_maxseg(tp),
12271                                      rack->r_ctl.rc_pace_min_segs);
12272                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12273                                                            tp, bw_est, segsiz, 0,
12274                                                            rack->r_ctl.crte, NULL);
12275                 }
12276         } else if (rack->rc_always_pace) {
12277                 if (rack->r_ctl.gp_bw ||
12278 #ifdef NETFLIX_PEAKRATE
12279                     rack->rc_tp->t_maxpeakrate ||
12280 #endif
12281                     rack->r_ctl.init_rate) {
12282                         /* We have a rate of some sort set */
12283                         uint32_t  orig;
12284
12285                         bw_est = rack_get_bw(rack);
12286                         orig = rack->r_ctl.rc_pace_max_segs;
12287                         if (fill_override)
12288                                 rate_wanted = *fill_override;
12289                         else
12290                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12291                         if (rate_wanted) {
12292                                 /* We have something */
12293                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12294                                                                                    rate_wanted,
12295                                                                                    ctf_fixed_maxseg(rack->rc_tp));
12296                         } else
12297                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12298                         if (orig != rack->r_ctl.rc_pace_max_segs)
12299                                 chged = 1;
12300                 } else if ((rack->r_ctl.gp_bw == 0) &&
12301                            (rack->r_ctl.rc_pace_max_segs == 0)) {
12302                         /*
12303                          * If we have nothing limit us to bursting
12304                          * out IW sized pieces.
12305                          */
12306                         chged = 1;
12307                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12308                 }
12309         }
12310         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12311                 chged = 1;
12312                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12313         }
12314         if (chged)
12315                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12316 }
12317
12318
12319 static void
12320 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12321 {
12322 #ifdef INET6
12323         struct ip6_hdr *ip6 = NULL;
12324 #endif
12325 #ifdef INET
12326         struct ip *ip = NULL;
12327 #endif
12328         struct udphdr *udp = NULL;
12329
12330         /* Ok lets fill in the fast block, it can only be used with no IP options! */
12331 #ifdef INET6
12332         if (rack->r_is_v6) {
12333                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12334                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12335                 if (tp->t_port) {
12336                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12337                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12338                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12339                         udp->uh_dport = tp->t_port;
12340                         rack->r_ctl.fsb.udp = udp;
12341                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12342                 } else
12343                 {
12344                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12345                         rack->r_ctl.fsb.udp = NULL;
12346                 }
12347                 tcpip_fillheaders(rack->rc_inp,
12348                                   tp->t_port,
12349                                   ip6, rack->r_ctl.fsb.th);
12350         } else
12351 #endif                          /* INET6 */
12352 #ifdef INET
12353         {
12354                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12355                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12356                 if (tp->t_port) {
12357                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12358                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12359                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12360                         udp->uh_dport = tp->t_port;
12361                         rack->r_ctl.fsb.udp = udp;
12362                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12363                 } else
12364                 {
12365                         rack->r_ctl.fsb.udp = NULL;
12366                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12367                 }
12368                 tcpip_fillheaders(rack->rc_inp,
12369                                   tp->t_port,
12370                                   ip, rack->r_ctl.fsb.th);
12371         }
12372 #endif
12373         rack->r_fsb_inited = 1;
12374 }
12375
12376 static int
12377 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12378 {
12379         /*
12380          * Allocate the larger of spaces V6 if available else just
12381          * V4 and include udphdr (overbook)
12382          */
12383 #ifdef INET6
12384         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12385 #else
12386         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12387 #endif
12388         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12389                                             M_TCPFSB, M_NOWAIT|M_ZERO);
12390         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12391                 return (ENOMEM);
12392         }
12393         rack->r_fsb_inited = 0;
12394         return (0);
12395 }
12396
12397 static int
12398 rack_init(struct tcpcb *tp)
12399 {
12400         struct inpcb *inp = tptoinpcb(tp);
12401         struct tcp_rack *rack = NULL;
12402 #ifdef INVARIANTS
12403         struct rack_sendmap *insret;
12404 #endif
12405         uint32_t iwin, snt, us_cts;
12406         int err;
12407
12408         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12409         if (tp->t_fb_ptr == NULL) {
12410                 /*
12411                  * We need to allocate memory but cant. The INP and INP_INFO
12412                  * locks and they are recursive (happens during setup. So a
12413                  * scheme to drop the locks fails :(
12414                  *
12415                  */
12416                 return (ENOMEM);
12417         }
12418         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12419
12420         rack = (struct tcp_rack *)tp->t_fb_ptr;
12421         RB_INIT(&rack->r_ctl.rc_mtree);
12422         TAILQ_INIT(&rack->r_ctl.rc_free);
12423         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12424         rack->rc_tp = tp;
12425         rack->rc_inp = inp;
12426         /* Set the flag */
12427         rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
12428         /* Probably not needed but lets be sure */
12429         rack_clear_rate_sample(rack);
12430         /*
12431          * Save off the default values, socket options will poke
12432          * at these if pacing is not on or we have not yet
12433          * reached where pacing is on (gp_ready/fixed enabled).
12434          * When they get set into the CC module (when gp_ready
12435          * is enabled or we enable fixed) then we will set these
12436          * values into the CC and place in here the old values
12437          * so we have a restoral. Then we will set the flag
12438          * rc_pacing_cc_set. That way whenever we turn off pacing
12439          * or switch off this stack, we will know to go restore
12440          * the saved values.
12441          */
12442         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12443         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12444         /* We want abe like behavior as well */
12445         rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12446         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12447         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12448         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12449         rack->r_ctl.roundends = tp->snd_max;
12450         if (use_rack_rr)
12451                 rack->use_rack_rr = 1;
12452         if (V_tcp_delack_enabled)
12453                 tp->t_delayed_ack = 1;
12454         else
12455                 tp->t_delayed_ack = 0;
12456 #ifdef TCP_ACCOUNTING
12457         if (rack_tcp_accounting) {
12458                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12459         }
12460 #endif
12461         if (rack_enable_shared_cwnd)
12462                 rack->rack_enable_scwnd = 1;
12463         rack->rc_user_set_max_segs = rack_hptsi_segments;
12464         rack->rc_force_max_seg = 0;
12465         if (rack_use_imac_dack)
12466                 rack->rc_dack_mode = 1;
12467         TAILQ_INIT(&rack->r_ctl.opt_list);
12468         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12469         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12470         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12471         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12472         rack->r_ctl.rc_highest_us_rtt = 0;
12473         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12474         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12475         if (rack_use_cmp_acks)
12476                 rack->r_use_cmp_ack = 1;
12477         if (rack_disable_prr)
12478                 rack->rack_no_prr = 1;
12479         if (rack_gp_no_rec_chg)
12480                 rack->rc_gp_no_rec_chg = 1;
12481         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12482                 rack->rc_always_pace = 1;
12483                 if (rack->use_fixed_rate || rack->gp_ready)
12484                         rack_set_cc_pacing(rack);
12485         } else
12486                 rack->rc_always_pace = 0;
12487         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12488                 rack->r_mbuf_queue = 1;
12489         else
12490                 rack->r_mbuf_queue = 0;
12491         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12492                 inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12493         else
12494                 inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12495         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12496         if (rack_limits_scwnd)
12497                 rack->r_limit_scw = 1;
12498         else
12499                 rack->r_limit_scw = 0;
12500         rack->rc_labc = V_tcp_abc_l_var;
12501         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12502         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12503         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12504         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12505         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12506         rack->r_ctl.rc_min_to = rack_min_to;
12507         microuptime(&rack->r_ctl.act_rcv_time);
12508         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12509         rack->rc_init_win = rack_default_init_window;
12510         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12511         if (rack_hw_up_only)
12512                 rack->r_up_only = 1;
12513         if (rack_do_dyn_mul) {
12514                 /* When dynamic adjustment is on CA needs to start at 100% */
12515                 rack->rc_gp_dyn_mul = 1;
12516                 if (rack_do_dyn_mul >= 100)
12517                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12518         } else
12519                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12520         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12521         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12522         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12523         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12524                                 rack_probertt_filter_life);
12525         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12526         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12527         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12528         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12529         rack->r_ctl.rc_time_probertt_starts = 0;
12530         if (rack_dsack_std_based & 0x1) {
12531                 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12532                 rack->rc_rack_tmr_std_based = 1;
12533         }
12534         if (rack_dsack_std_based & 0x2) {
12535                 /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12536                 rack->rc_rack_use_dsack = 1;
12537         }
12538         /* We require at least one measurement, even if the sysctl is 0 */
12539         if (rack_req_measurements)
12540                 rack->r_ctl.req_measurements = rack_req_measurements;
12541         else
12542                 rack->r_ctl.req_measurements = 1;
12543         if (rack_enable_hw_pacing)
12544                 rack->rack_hdw_pace_ena = 1;
12545         if (rack_hw_rate_caps)
12546                 rack->r_rack_hw_rate_caps = 1;
12547         /* Do we force on detection? */
12548 #ifdef NETFLIX_EXP_DETECTION
12549         if (tcp_force_detection)
12550                 rack->do_detection = 1;
12551         else
12552 #endif
12553                 rack->do_detection = 0;
12554         if (rack_non_rxt_use_cr)
12555                 rack->rack_rec_nonrxt_use_cr = 1;
12556         err = rack_init_fsb(tp, rack);
12557         if (err) {
12558                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12559                 tp->t_fb_ptr = NULL;
12560                 return (err);
12561         }
12562         if (tp->snd_una != tp->snd_max) {
12563                 /* Create a send map for the current outstanding data */
12564                 struct rack_sendmap *rsm;
12565
12566                 rsm = rack_alloc(rack);
12567                 if (rsm == NULL) {
12568                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12569                         tp->t_fb_ptr = NULL;
12570                         return (ENOMEM);
12571                 }
12572                 rsm->r_no_rtt_allowed = 1;
12573                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12574                 rsm->r_rtr_cnt = 1;
12575                 rsm->r_rtr_bytes = 0;
12576                 if (tp->t_flags & TF_SENTFIN)
12577                         rsm->r_flags |= RACK_HAS_FIN;
12578                 if ((tp->snd_una == tp->iss) &&
12579                     !TCPS_HAVEESTABLISHED(tp->t_state))
12580                         rsm->r_flags |= RACK_HAS_SYN;
12581                 rsm->r_start = tp->snd_una;
12582                 rsm->r_end = tp->snd_max;
12583                 rsm->r_dupack = 0;
12584                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12585                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12586                         if (rsm->m)
12587                                 rsm->orig_m_len = rsm->m->m_len;
12588                         else
12589                                 rsm->orig_m_len = 0;
12590                 } else {
12591                         /*
12592                          * This can happen if we have a stand-alone FIN or
12593                          *  SYN.
12594                          */
12595                         rsm->m = NULL;
12596                         rsm->orig_m_len = 0;
12597                         rsm->soff = 0;
12598                 }
12599 #ifndef INVARIANTS
12600                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12601 #else
12602                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12603                 if (insret != NULL) {
12604                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12605                               insret, rack, rsm);
12606                 }
12607 #endif
12608                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12609                 rsm->r_in_tmap = 1;
12610         }
12611         /*
12612          * Timers in Rack are kept in microseconds so lets
12613          * convert any initial incoming variables
12614          * from ticks into usecs. Note that we
12615          * also change the values of t_srtt and t_rttvar, if
12616          * they are non-zero. They are kept with a 5
12617          * bit decimal so we have to carefully convert
12618          * these to get the full precision.
12619          */
12620         rack_convert_rtts(tp);
12621         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12622         if (rack_do_hystart) {
12623                 tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
12624                 if (rack_do_hystart > 1)
12625                         tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
12626                 if (rack_do_hystart > 2)
12627                         tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
12628         }
12629         if (rack_def_profile)
12630                 rack_set_profile(rack, rack_def_profile);
12631         /* Cancel the GP measurement in progress */
12632         tp->t_flags &= ~TF_GPUTINPROG;
12633         if (SEQ_GT(tp->snd_max, tp->iss))
12634                 snt = tp->snd_max - tp->iss;
12635         else
12636                 snt = 0;
12637         iwin = rc_init_window(rack);
12638         if (snt < iwin) {
12639                 /* We are not past the initial window
12640                  * so we need to make sure cwnd is
12641                  * correct.
12642                  */
12643                 if (tp->snd_cwnd < iwin)
12644                         tp->snd_cwnd = iwin;
12645                 /*
12646                  * If we are within the initial window
12647                  * we want ssthresh to be unlimited. Setting
12648                  * it to the rwnd (which the default stack does
12649                  * and older racks) is not really a good idea
12650                  * since we want to be in SS and grow both the
12651                  * cwnd and the rwnd (via dynamic rwnd growth). If
12652                  * we set it to the rwnd then as the peer grows its
12653                  * rwnd we will be stuck in CA and never hit SS.
12654                  *
12655                  * Its far better to raise it up high (this takes the
12656                  * risk that there as been a loss already, probably
12657                  * we should have an indicator in all stacks of loss
12658                  * but we don't), but considering the normal use this
12659                  * is a risk worth taking. The consequences of not
12660                  * hitting SS are far worse than going one more time
12661                  * into it early on (before we have sent even a IW).
12662                  * It is highly unlikely that we will have had a loss
12663                  * before getting the IW out.
12664                  */
12665                 tp->snd_ssthresh = 0xffffffff;
12666         }
12667         rack_stop_all_timers(tp);
12668         /* Lets setup the fsb block */
12669         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12670         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12671                              __LINE__, RACK_RTTS_INIT);
12672         return (0);
12673 }
12674
12675 static int
12676 rack_handoff_ok(struct tcpcb *tp)
12677 {
12678         if ((tp->t_state == TCPS_CLOSED) ||
12679             (tp->t_state == TCPS_LISTEN)) {
12680                 /* Sure no problem though it may not stick */
12681                 return (0);
12682         }
12683         if ((tp->t_state == TCPS_SYN_SENT) ||
12684             (tp->t_state == TCPS_SYN_RECEIVED)) {
12685                 /*
12686                  * We really don't know if you support sack,
12687                  * you have to get to ESTAB or beyond to tell.
12688                  */
12689                 return (EAGAIN);
12690         }
12691         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12692                 /*
12693                  * Rack will only send a FIN after all data is acknowledged.
12694                  * So in this case we have more data outstanding. We can't
12695                  * switch stacks until either all data and only the FIN
12696                  * is left (in which case rack_init() now knows how
12697                  * to deal with that) <or> all is acknowledged and we
12698                  * are only left with incoming data, though why you
12699                  * would want to switch to rack after all data is acknowledged
12700                  * I have no idea (rrs)!
12701                  */
12702                 return (EAGAIN);
12703         }
12704         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12705                 return (0);
12706         }
12707         /*
12708          * If we reach here we don't do SACK on this connection so we can
12709          * never do rack.
12710          */
12711         return (EINVAL);
12712 }
12713
12714
12715 static void
12716 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12717 {
12718         struct inpcb *inp = tptoinpcb(tp);
12719
12720         if (tp->t_fb_ptr) {
12721                 struct tcp_rack *rack;
12722                 struct rack_sendmap *rsm, *nrsm;
12723 #ifdef INVARIANTS
12724                 struct rack_sendmap *rm;
12725 #endif
12726
12727                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12728                 if (tp->t_in_pkt) {
12729                         /*
12730                          * It is unsafe to process the packets since a
12731                          * reset may be lurking in them (its rare but it
12732                          * can occur). If we were to find a RST, then we
12733                          * would end up dropping the connection and the
12734                          * INP lock, so when we return the caller (tcp_usrreq)
12735                          * will blow up when it trys to unlock the inp.
12736                          */
12737                         struct mbuf *save, *m;
12738
12739                         m = tp->t_in_pkt;
12740                         tp->t_in_pkt = NULL;
12741                         tp->t_tail_pkt = NULL;
12742                         while (m) {
12743                                 save = m->m_nextpkt;
12744                                 m->m_nextpkt = NULL;
12745                                 m_freem(m);
12746                                 m = save;
12747                         }
12748                 }
12749                 tp->t_flags &= ~TF_FORCEDATA;
12750 #ifdef NETFLIX_SHARED_CWND
12751                 if (rack->r_ctl.rc_scw) {
12752                         uint32_t limit;
12753
12754                         if (rack->r_limit_scw)
12755                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12756                         else
12757                                 limit = 0;
12758                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12759                                                   rack->r_ctl.rc_scw_index,
12760                                                   limit);
12761                         rack->r_ctl.rc_scw = NULL;
12762                 }
12763 #endif
12764                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12765                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12766                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12767                         rack->r_ctl.fsb.th = NULL;
12768                 }
12769                 /* Convert back to ticks, with  */
12770                 if (tp->t_srtt > 1) {
12771                         uint32_t val, frac;
12772
12773                         val = USEC_2_TICKS(tp->t_srtt);
12774                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12775                         tp->t_srtt = val << TCP_RTT_SHIFT;
12776                         /*
12777                          * frac is the fractional part here is left
12778                          * over from converting to hz and shifting.
12779                          * We need to convert this to the 5 bit
12780                          * remainder.
12781                          */
12782                         if (frac) {
12783                                 if (hz == 1000) {
12784                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12785                                 } else {
12786                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12787                                 }
12788                                 tp->t_srtt += frac;
12789                         }
12790                 }
12791                 if (tp->t_rttvar) {
12792                         uint32_t val, frac;
12793
12794                         val = USEC_2_TICKS(tp->t_rttvar);
12795                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12796                         tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12797                         /*
12798                          * frac is the fractional part here is left
12799                          * over from converting to hz and shifting.
12800                          * We need to convert this to the 5 bit
12801                          * remainder.
12802                          */
12803                         if (frac) {
12804                                 if (hz == 1000) {
12805                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12806                                 } else {
12807                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12808                                 }
12809                                 tp->t_rttvar += frac;
12810                         }
12811                 }
12812                 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12813                 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12814                 if (rack->rc_always_pace) {
12815                         tcp_decrement_paced_conn();
12816                         rack_undo_cc_pacing(rack);
12817                         rack->rc_always_pace = 0;
12818                 }
12819                 /* Clean up any options if they were not applied */
12820                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12821                         struct deferred_opt_list *dol;
12822
12823                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12824                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12825                         free(dol, M_TCPDO);
12826                 }
12827                 /* rack does not use force data but other stacks may clear it */
12828                 if (rack->r_ctl.crte != NULL) {
12829                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12830                         rack->rack_hdrw_pacing = 0;
12831                         rack->r_ctl.crte = NULL;
12832                 }
12833 #ifdef TCP_BLACKBOX
12834                 tcp_log_flowend(tp);
12835 #endif
12836                 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12837 #ifndef INVARIANTS
12838                         (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12839 #else
12840                         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12841                         if (rm != rsm) {
12842                                 panic("At fini, rack:%p rsm:%p rm:%p",
12843                                       rack, rsm, rm);
12844                         }
12845 #endif
12846                         uma_zfree(rack_zone, rsm);
12847                 }
12848                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12849                 while (rsm) {
12850                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12851                         uma_zfree(rack_zone, rsm);
12852                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12853                 }
12854                 rack->rc_free_cnt = 0;
12855                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12856                 tp->t_fb_ptr = NULL;
12857         }
12858         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12859         inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12860         inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12861         inp->inp_flags2 &= ~INP_MBUF_ACKCMP;
12862         /* Cancel the GP measurement in progress */
12863         tp->t_flags &= ~TF_GPUTINPROG;
12864         inp->inp_flags2 &= ~INP_MBUF_L_ACKS;
12865         /* Make sure snd_nxt is correctly set */
12866         tp->snd_nxt = tp->snd_max;
12867 }
12868
12869 static void
12870 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12871 {
12872         if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12873                 rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
12874         }
12875         switch (tp->t_state) {
12876         case TCPS_SYN_SENT:
12877                 rack->r_state = TCPS_SYN_SENT;
12878                 rack->r_substate = rack_do_syn_sent;
12879                 break;
12880         case TCPS_SYN_RECEIVED:
12881                 rack->r_state = TCPS_SYN_RECEIVED;
12882                 rack->r_substate = rack_do_syn_recv;
12883                 break;
12884         case TCPS_ESTABLISHED:
12885                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12886                 rack->r_state = TCPS_ESTABLISHED;
12887                 rack->r_substate = rack_do_established;
12888                 break;
12889         case TCPS_CLOSE_WAIT:
12890                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12891                 rack->r_state = TCPS_CLOSE_WAIT;
12892                 rack->r_substate = rack_do_close_wait;
12893                 break;
12894         case TCPS_FIN_WAIT_1:
12895                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12896                 rack->r_state = TCPS_FIN_WAIT_1;
12897                 rack->r_substate = rack_do_fin_wait_1;
12898                 break;
12899         case TCPS_CLOSING:
12900                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12901                 rack->r_state = TCPS_CLOSING;
12902                 rack->r_substate = rack_do_closing;
12903                 break;
12904         case TCPS_LAST_ACK:
12905                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12906                 rack->r_state = TCPS_LAST_ACK;
12907                 rack->r_substate = rack_do_lastack;
12908                 break;
12909         case TCPS_FIN_WAIT_2:
12910                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12911                 rack->r_state = TCPS_FIN_WAIT_2;
12912                 rack->r_substate = rack_do_fin_wait_2;
12913                 break;
12914         case TCPS_LISTEN:
12915         case TCPS_CLOSED:
12916         case TCPS_TIME_WAIT:
12917         default:
12918                 break;
12919         };
12920         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12921                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12922
12923 }
12924
12925 static void
12926 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12927 {
12928         /*
12929          * We received an ack, and then did not
12930          * call send or were bounced out due to the
12931          * hpts was running. Now a timer is up as well, is
12932          * it the right timer?
12933          */
12934         struct rack_sendmap *rsm;
12935         int tmr_up;
12936
12937         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12938         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12939                 return;
12940         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12941         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12942             (tmr_up == PACE_TMR_RXT)) {
12943                 /* Should be an RXT */
12944                 return;
12945         }
12946         if (rsm == NULL) {
12947                 /* Nothing outstanding? */
12948                 if (tp->t_flags & TF_DELACK) {
12949                         if (tmr_up == PACE_TMR_DELACK)
12950                                 /* We are supposed to have delayed ack up and we do */
12951                                 return;
12952                 } else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12953                         /*
12954                          * if we hit enobufs then we would expect the possibility
12955                          * of nothing outstanding and the RXT up (and the hptsi timer).
12956                          */
12957                         return;
12958                 } else if (((V_tcp_always_keepalive ||
12959                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12960                             (tp->t_state <= TCPS_CLOSING)) &&
12961                            (tmr_up == PACE_TMR_KEEP) &&
12962                            (tp->snd_max == tp->snd_una)) {
12963                         /* We should have keep alive up and we do */
12964                         return;
12965                 }
12966         }
12967         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12968                    ((tmr_up == PACE_TMR_TLP) ||
12969                     (tmr_up == PACE_TMR_RACK) ||
12970                     (tmr_up == PACE_TMR_RXT))) {
12971                 /*
12972                  * Either a Rack, TLP or RXT is fine if  we
12973                  * have outstanding data.
12974                  */
12975                 return;
12976         } else if (tmr_up == PACE_TMR_DELACK) {
12977                 /*
12978                  * If the delayed ack was going to go off
12979                  * before the rtx/tlp/rack timer were going to
12980                  * expire, then that would be the timer in control.
12981                  * Note we don't check the time here trusting the
12982                  * code is correct.
12983                  */
12984                 return;
12985         }
12986         /*
12987          * Ok the timer originally started is not what we want now.
12988          * We will force the hpts to be stopped if any, and restart
12989          * with the slot set to what was in the saved slot.
12990          */
12991         if (tcp_in_hpts(rack->rc_inp)) {
12992                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12993                         uint32_t us_cts;
12994
12995                         us_cts = tcp_get_usecs(NULL);
12996                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12997                                 rack->r_early = 1;
12998                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12999                         }
13000                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13001                 }
13002                 tcp_hpts_remove(rack->rc_inp);
13003         }
13004         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13005         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13006 }
13007
13008
13009 static void
13010 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts, uint32_t high_seq)
13011 {
13012         if ((SEQ_LT(tp->snd_wl1, seq) ||
13013             (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
13014             (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
13015                 /* keep track of pure window updates */
13016                 if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
13017                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13018                 tp->snd_wnd = tiwin;
13019                 rack_validate_fo_sendwin_up(tp, rack);
13020                 tp->snd_wl1 = seq;
13021                 tp->snd_wl2 = ack;
13022                 if (tp->snd_wnd > tp->max_sndwnd)
13023                         tp->max_sndwnd = tp->snd_wnd;
13024             rack->r_wanted_output = 1;
13025         } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
13026                 tp->snd_wnd = tiwin;
13027                 rack_validate_fo_sendwin_up(tp, rack);
13028                 tp->snd_wl1 = seq;
13029                 tp->snd_wl2 = ack;
13030         } else {
13031                 /* Not a valid win update */
13032                 return;
13033         }
13034         /* Do we exit persists? */
13035         if ((rack->rc_in_persist != 0) &&
13036             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13037                                 rack->r_ctl.rc_pace_min_segs))) {
13038                 rack_exit_persist(tp, rack, cts);
13039         }
13040         /* Do we enter persists? */
13041         if ((rack->rc_in_persist == 0) &&
13042             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13043             TCPS_HAVEESTABLISHED(tp->t_state) &&
13044             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13045             sbavail(&tptosocket(tp)->so_snd) &&
13046             (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13047                 /*
13048                  * Here the rwnd is less than
13049                  * the pacing size, we are established,
13050                  * nothing is outstanding, and there is
13051                  * data to send. Enter persists.
13052                  */
13053                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13054         }
13055 }
13056
13057 static void
13058 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13059 {
13060
13061         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13062                 struct inpcb *inp = tptoinpcb(tp);
13063                 union tcp_log_stackspecific log;
13064                 struct timeval ltv;
13065                 char tcp_hdr_buf[60];
13066                 struct tcphdr *th;
13067                 struct timespec ts;
13068                 uint32_t orig_snd_una;
13069                 uint8_t xx = 0;
13070
13071 #ifdef NETFLIX_HTTP_LOGGING
13072                 struct http_sendfile_track *http_req;
13073
13074                 if (SEQ_GT(ae->ack, tp->snd_una)) {
13075                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13076                 } else {
13077                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13078                 }
13079 #endif
13080                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13081                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13082                 if (rack->rack_no_prr == 0)
13083                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13084                 else
13085                         log.u_bbr.flex1 = 0;
13086                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13087                 log.u_bbr.use_lt_bw <<= 1;
13088                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13089                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13090                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13091                 log.u_bbr.pkts_out = tp->t_maxseg;
13092                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13093                 log.u_bbr.flex7 = 1;
13094                 log.u_bbr.lost = ae->flags;
13095                 log.u_bbr.cwnd_gain = ackval;
13096                 log.u_bbr.pacing_gain = 0x2;
13097                 if (ae->flags & TSTMP_HDWR) {
13098                         /* Record the hardware timestamp if present */
13099                         log.u_bbr.flex3 = M_TSTMP;
13100                         ts.tv_sec = ae->timestamp / 1000000000;
13101                         ts.tv_nsec = ae->timestamp % 1000000000;
13102                         ltv.tv_sec = ts.tv_sec;
13103                         ltv.tv_usec = ts.tv_nsec / 1000;
13104                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13105                 } else if (ae->flags & TSTMP_LRO) {
13106                         /* Record the LRO the arrival timestamp */
13107                         log.u_bbr.flex3 = M_TSTMP_LRO;
13108                         ts.tv_sec = ae->timestamp / 1000000000;
13109                         ts.tv_nsec = ae->timestamp % 1000000000;
13110                         ltv.tv_sec = ts.tv_sec;
13111                         ltv.tv_usec = ts.tv_nsec / 1000;
13112                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13113                 }
13114                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13115                 /* Log the rcv time */
13116                 log.u_bbr.delRate = ae->timestamp;
13117 #ifdef NETFLIX_HTTP_LOGGING
13118                 log.u_bbr.applimited = tp->t_http_closed;
13119                 log.u_bbr.applimited <<= 8;
13120                 log.u_bbr.applimited |= tp->t_http_open;
13121                 log.u_bbr.applimited <<= 8;
13122                 log.u_bbr.applimited |= tp->t_http_req;
13123                 if (http_req) {
13124                         /* Copy out any client req info */
13125                         /* seconds */
13126                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13127                         /* useconds */
13128                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13129                         log.u_bbr.rttProp = http_req->timestamp;
13130                         log.u_bbr.cur_del_rate = http_req->start;
13131                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13132                                 log.u_bbr.flex8 |= 1;
13133                         } else {
13134                                 log.u_bbr.flex8 |= 2;
13135                                 log.u_bbr.bw_inuse = http_req->end;
13136                         }
13137                         log.u_bbr.flex6 = http_req->start_seq;
13138                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13139                                 log.u_bbr.flex8 |= 4;
13140                                 log.u_bbr.epoch = http_req->end_seq;
13141                         }
13142                 }
13143 #endif
13144                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13145                 th = (struct tcphdr *)tcp_hdr_buf;
13146                 th->th_seq = ae->seq;
13147                 th->th_ack = ae->ack;
13148                 th->th_win = ae->win;
13149                 /* Now fill in the ports */
13150                 th->th_sport = inp->inp_fport;
13151                 th->th_dport = inp->inp_lport;
13152                 tcp_set_flags(th, ae->flags);
13153                 /* Now do we have a timestamp option? */
13154                 if (ae->flags & HAS_TSTMP) {
13155                         u_char *cp;
13156                         uint32_t val;
13157
13158                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13159                         cp = (u_char *)(th + 1);
13160                         *cp = TCPOPT_NOP;
13161                         cp++;
13162                         *cp = TCPOPT_NOP;
13163                         cp++;
13164                         *cp = TCPOPT_TIMESTAMP;
13165                         cp++;
13166                         *cp = TCPOLEN_TIMESTAMP;
13167                         cp++;
13168                         val = htonl(ae->ts_value);
13169                         bcopy((char *)&val,
13170                               (char *)cp, sizeof(uint32_t));
13171                         val = htonl(ae->ts_echo);
13172                         bcopy((char *)&val,
13173                               (char *)(cp + 4), sizeof(uint32_t));
13174                 } else
13175                         th->th_off = (sizeof(struct tcphdr) >> 2);
13176
13177                 /*
13178                  * For sane logging we need to play a little trick.
13179                  * If the ack were fully processed we would have moved
13180                  * snd_una to high_seq, but since compressed acks are
13181                  * processed in two phases, at this point (logging) snd_una
13182                  * won't be advanced. So we would see multiple acks showing
13183                  * the advancement. We can prevent that by "pretending" that
13184                  * snd_una was advanced and then un-advancing it so that the
13185                  * logging code has the right value for tlb_snd_una.
13186                  */
13187                 if (tp->snd_una != high_seq) {
13188                         orig_snd_una = tp->snd_una;
13189                         tp->snd_una = high_seq;
13190                         xx = 1;
13191                 } else
13192                         xx = 0;
13193                 TCP_LOG_EVENTP(tp, th,
13194                                &tptosocket(tp)->so_rcv,
13195                                &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
13196                                0, &log, true, &ltv);
13197                 if (xx) {
13198                         tp->snd_una = orig_snd_una;
13199                 }
13200         }
13201
13202 }
13203
13204 static void
13205 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13206 {
13207         uint32_t us_rtt;
13208         /*
13209          * A persist or keep-alive was forced out, update our
13210          * min rtt time. Note now worry about lost responses.
13211          * When a subsequent keep-alive or persist times out
13212          * and forced_ack is still on, then the last probe
13213          * was not responded to. In such cases we have a
13214          * sysctl that controls the behavior. Either we apply
13215          * the rtt but with reduced confidence (0). Or we just
13216          * plain don't apply the rtt estimate. Having data flow
13217          * will clear the probe_not_answered flag i.e. cum-ack
13218          * move forward <or> exiting and reentering persists.
13219          */
13220
13221         rack->forced_ack = 0;
13222         rack->rc_tp->t_rxtshift = 0;
13223         if ((rack->rc_in_persist &&
13224              (tiwin == rack->rc_tp->snd_wnd)) ||
13225             (rack->rc_in_persist == 0)) {
13226                 /*
13227                  * In persists only apply the RTT update if this is
13228                  * a response to our window probe. And that
13229                  * means the rwnd sent must match the current
13230                  * snd_wnd. If it does not, then we got a
13231                  * window update ack instead. For keepalive
13232                  * we allow the answer no matter what the window.
13233                  *
13234                  * Note that if the probe_not_answered is set then
13235                  * the forced_ack_ts is the oldest one i.e. the first
13236                  * probe sent that might have been lost. This assures
13237                  * us that if we do calculate an RTT it is longer not
13238                  * some short thing.
13239                  */
13240                 if (rack->rc_in_persist)
13241                         counter_u64_add(rack_persists_acks, 1);
13242                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13243                 if (us_rtt == 0)
13244                         us_rtt = 1;
13245                 if (rack->probe_not_answered == 0) {
13246                         rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13247                         tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13248                 } else {
13249                         /* We have a retransmitted probe here too */
13250                         if (rack_apply_rtt_with_reduced_conf) {
13251                                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13252                                 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13253                         }
13254                 }
13255         }
13256 }
13257
13258 static int
13259 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13260 {
13261         /*
13262          * Handle a "special" compressed ack mbuf. Each incoming
13263          * ack has only four possible dispositions:
13264          *
13265          * A) It moves the cum-ack forward
13266          * B) It is behind the cum-ack.
13267          * C) It is a window-update ack.
13268          * D) It is a dup-ack.
13269          *
13270          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13271          * in the incoming mbuf. We also need to still pay attention
13272          * to nxt_pkt since there may be another packet after this
13273          * one.
13274          */
13275 #ifdef TCP_ACCOUNTING
13276         uint64_t ts_val;
13277         uint64_t rdstc;
13278 #endif
13279         int segsiz;
13280         struct timespec ts;
13281         struct tcp_rack *rack;
13282         struct tcp_ackent *ae;
13283         uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13284         int cnt, i, did_out, ourfinisacked = 0;
13285         struct tcpopt to_holder, *to = NULL;
13286 #ifdef TCP_ACCOUNTING
13287         int win_up_req = 0;
13288 #endif
13289         int nsegs = 0;
13290         int under_pacing = 1;
13291         int recovery = 0;
13292 #ifdef TCP_ACCOUNTING
13293         sched_pin();
13294 #endif
13295         rack = (struct tcp_rack *)tp->t_fb_ptr;
13296         if (rack->gp_ready &&
13297             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13298                 under_pacing = 0;
13299         else
13300                 under_pacing = 1;
13301
13302         if (rack->r_state != tp->t_state)
13303                 rack_set_state(tp, rack);
13304         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13305             (tp->t_flags & TF_GPUTINPROG)) {
13306                 /*
13307                  * We have a goodput in progress
13308                  * and we have entered a late state.
13309                  * Do we have enough data in the sb
13310                  * to handle the GPUT request?
13311                  */
13312                 uint32_t bytes;
13313
13314                 bytes = tp->gput_ack - tp->gput_seq;
13315                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
13316                         bytes += tp->gput_seq - tp->snd_una;
13317                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
13318                         /*
13319                          * There are not enough bytes in the socket
13320                          * buffer that have been sent to cover this
13321                          * measurement. Cancel it.
13322                          */
13323                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13324                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
13325                                                    tp->gput_seq,
13326                                                    0, 0, 18, __LINE__, NULL, 0);
13327                         tp->t_flags &= ~TF_GPUTINPROG;
13328                 }
13329         }
13330         to = &to_holder;
13331         to->to_flags = 0;
13332         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13333                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13334         cnt = m->m_len / sizeof(struct tcp_ackent);
13335         counter_u64_add(rack_multi_single_eq, cnt);
13336         high_seq = tp->snd_una;
13337         the_win = tp->snd_wnd;
13338         win_seq = tp->snd_wl1;
13339         win_upd_ack = tp->snd_wl2;
13340         cts = tcp_tv_to_usectick(tv);
13341         ms_cts = tcp_tv_to_mssectick(tv);
13342         rack->r_ctl.rc_rcvtime = cts;
13343         segsiz = ctf_fixed_maxseg(tp);
13344         if ((rack->rc_gp_dyn_mul) &&
13345             (rack->use_fixed_rate == 0) &&
13346             (rack->rc_always_pace)) {
13347                 /* Check in on probertt */
13348                 rack_check_probe_rtt(rack, cts);
13349         }
13350         for (i = 0; i < cnt; i++) {
13351 #ifdef TCP_ACCOUNTING
13352                 ts_val = get_cyclecount();
13353 #endif
13354                 rack_clear_rate_sample(rack);
13355                 ae = ((mtod(m, struct tcp_ackent *)) + i);
13356                 /* Setup the window */
13357                 tiwin = ae->win << tp->snd_scale;
13358                 if (tiwin > rack->r_ctl.rc_high_rwnd)
13359                         rack->r_ctl.rc_high_rwnd = tiwin;
13360                 /* figure out the type of ack */
13361                 if (SEQ_LT(ae->ack, high_seq)) {
13362                         /* Case B*/
13363                         ae->ack_val_set = ACK_BEHIND;
13364                 } else if (SEQ_GT(ae->ack, high_seq)) {
13365                         /* Case A */
13366                         ae->ack_val_set = ACK_CUMACK;
13367                 } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13368                         /* Case D */
13369                         ae->ack_val_set = ACK_DUPACK;
13370                 } else {
13371                         /* Case C */
13372                         ae->ack_val_set = ACK_RWND;
13373                 }
13374                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13375                 /* Validate timestamp */
13376                 if (ae->flags & HAS_TSTMP) {
13377                         /* Setup for a timestamp */
13378                         to->to_flags = TOF_TS;
13379                         ae->ts_echo -= tp->ts_offset;
13380                         to->to_tsecr = ae->ts_echo;
13381                         to->to_tsval = ae->ts_value;
13382                         /*
13383                          * If echoed timestamp is later than the current time, fall back to
13384                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13385                          * were used when this connection was established.
13386                          */
13387                         if (TSTMP_GT(ae->ts_echo, ms_cts))
13388                                 to->to_tsecr = 0;
13389                         if (tp->ts_recent &&
13390                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13391                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13392 #ifdef TCP_ACCOUNTING
13393                                         rdstc = get_cyclecount();
13394                                         if (rdstc > ts_val) {
13395                                                 counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13396                                                                 (rdstc - ts_val));
13397                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13398                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13399                                                 }
13400                                         }
13401 #endif
13402                                         continue;
13403                                 }
13404                         }
13405                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13406                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13407                                 tp->ts_recent_age = tcp_ts_getticks();
13408                                 tp->ts_recent = ae->ts_value;
13409                         }
13410                 } else {
13411                         /* Setup for a no options */
13412                         to->to_flags = 0;
13413                 }
13414                 /* Update the rcv time and perform idle reduction possibly */
13415                 if  (tp->t_idle_reduce &&
13416                      (tp->snd_max == tp->snd_una) &&
13417                      (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13418                         counter_u64_add(rack_input_idle_reduces, 1);
13419                         rack_cc_after_idle(rack, tp);
13420                 }
13421                 tp->t_rcvtime = ticks;
13422                 /* Now what about ECN of a chain of pure ACKs? */
13423                 if (tcp_ecn_input_segment(tp, ae->flags, 0,
13424                         tcp_packets_this_ack(tp, ae->ack),
13425                         ae->codepoint))
13426                         rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
13427 #ifdef TCP_ACCOUNTING
13428                 /* Count for the specific type of ack in */
13429                 counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13430                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13431                         tp->tcp_cnt_counters[ae->ack_val_set]++;
13432                 }
13433 #endif
13434                 /*
13435                  * Note how we could move up these in the determination
13436                  * above, but we don't so that way the timestamp checks (and ECN)
13437                  * is done first before we do any processing on the ACK.
13438                  * The non-compressed path through the code has this
13439                  * weakness (noted by @jtl) that it actually does some
13440                  * processing before verifying the timestamp information.
13441                  * We don't take that path here which is why we set
13442                  * the ack_val_set first, do the timestamp and ecn
13443                  * processing, and then look at what we have setup.
13444                  */
13445                 if (ae->ack_val_set == ACK_BEHIND) {
13446                         /*
13447                          * Case B flag reordering, if window is not closed
13448                          * or it could be a keep-alive or persists
13449                          */
13450                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13451                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13452                         }
13453                 } else if (ae->ack_val_set == ACK_DUPACK) {
13454                         /* Case D */
13455                         rack_strike_dupack(rack);
13456                 } else if (ae->ack_val_set == ACK_RWND) {
13457                         /* Case C */
13458                         if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13459                                 ts.tv_sec = ae->timestamp / 1000000000;
13460                                 ts.tv_nsec = ae->timestamp % 1000000000;
13461                                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13462                                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13463                         } else {
13464                                 rack->r_ctl.act_rcv_time = *tv;
13465                         }
13466                         if (rack->forced_ack) {
13467                                 rack_handle_probe_response(rack, tiwin,
13468                                                            tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13469                         }
13470 #ifdef TCP_ACCOUNTING
13471                         win_up_req = 1;
13472 #endif
13473                         win_upd_ack = ae->ack;
13474                         win_seq = ae->seq;
13475                         the_win = tiwin;
13476                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13477                 } else {
13478                         /* Case A */
13479                         if (SEQ_GT(ae->ack, tp->snd_max)) {
13480                                 /*
13481                                  * We just send an ack since the incoming
13482                                  * ack is beyond the largest seq we sent.
13483                                  */
13484                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
13485                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13486                                         if (tp->t_flags && TF_ACKNOW)
13487                                                 rack->r_wanted_output = 1;
13488                                 }
13489                         } else {
13490                                 nsegs++;
13491                                 /* If the window changed setup to update */
13492                                 if (tiwin != tp->snd_wnd) {
13493                                         win_upd_ack = ae->ack;
13494                                         win_seq = ae->seq;
13495                                         the_win = tiwin;
13496                                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13497                                 }
13498 #ifdef TCP_ACCOUNTING
13499                                 /* Account for the acks */
13500                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13501                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13502                                 }
13503                                 counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13504                                                 (((ae->ack - high_seq) + segsiz - 1) / segsiz));
13505 #endif
13506                                 high_seq = ae->ack;
13507                                 if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
13508                                         union tcp_log_stackspecific log;
13509                                         struct timeval tv;
13510
13511                                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13512                                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13513                                         log.u_bbr.flex1 = high_seq;
13514                                         log.u_bbr.flex2 = rack->r_ctl.roundends;
13515                                         log.u_bbr.flex3 = rack->r_ctl.current_round;
13516                                         log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
13517                                         log.u_bbr.flex8 = 8;
13518                                         tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
13519                                                        0, &log, false, NULL, NULL, 0, &tv);
13520                                 }
13521                                 /*
13522                                  * The draft (v3) calls for us to use SEQ_GEQ, but that
13523                                  * causes issues when we are just going app limited. Lets
13524                                  * instead use SEQ_GT <or> where its equal but more data
13525                                  * is outstanding.
13526                                  */
13527                                 if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) ||
13528                                     ((high_seq == rack->r_ctl.roundends) &&
13529                                      SEQ_GT(tp->snd_max, tp->snd_una))) {
13530                                         rack->r_ctl.current_round++;
13531                                         rack->r_ctl.roundends = tp->snd_max;
13532                                         if (CC_ALGO(tp)->newround != NULL) {
13533                                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
13534                                         }
13535                                 }
13536                                 /* Setup our act_rcv_time */
13537                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13538                                         ts.tv_sec = ae->timestamp / 1000000000;
13539                                         ts.tv_nsec = ae->timestamp % 1000000000;
13540                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13541                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13542                                 } else {
13543                                         rack->r_ctl.act_rcv_time = *tv;
13544                                 }
13545                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13546                                 if (rack->rc_dsack_round_seen) {
13547                                         /* Is the dsack round over? */
13548                                         if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13549                                                 /* Yes it is */
13550                                                 rack->rc_dsack_round_seen = 0;
13551                                                 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13552                                         }
13553                                 }
13554                         }
13555                 }
13556                 /* And lets be sure to commit the rtt measurements for this ack */
13557                 tcp_rack_xmit_timer_commit(rack, tp);
13558 #ifdef TCP_ACCOUNTING
13559                 rdstc = get_cyclecount();
13560                 if (rdstc > ts_val) {
13561                         counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13562                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13563                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13564                                 if (ae->ack_val_set == ACK_CUMACK)
13565                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13566                         }
13567                 }
13568 #endif
13569         }
13570 #ifdef TCP_ACCOUNTING
13571         ts_val = get_cyclecount();
13572 #endif
13573         /* Tend to any collapsed window */
13574         if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
13575                 /* The peer collapsed the window */
13576                 rack_collapsed_window(rack, (tp->snd_max - high_seq), __LINE__);
13577         } else if (rack->rc_has_collapsed)
13578                 rack_un_collapse_window(rack, __LINE__);
13579         if ((rack->r_collapse_point_valid) &&
13580             (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
13581                 rack->r_collapse_point_valid = 0;
13582         acked_amount = acked = (high_seq - tp->snd_una);
13583         if (acked) {
13584                 /*
13585                  * Clear the probe not answered flag
13586                  * since cum-ack moved forward.
13587                  */
13588                 rack->probe_not_answered = 0;
13589                 if (rack->sack_attack_disable == 0)
13590                         rack_do_decay(rack);
13591                 if (acked >= segsiz) {
13592                         /*
13593                          * You only get credit for
13594                          * MSS and greater (and you get extra
13595                          * credit for larger cum-ack moves).
13596                          */
13597                         int ac;
13598
13599                         ac = acked / segsiz;
13600                         rack->r_ctl.ack_count += ac;
13601                         counter_u64_add(rack_ack_total, ac);
13602                 }
13603                 if (rack->r_ctl.ack_count > 0xfff00000) {
13604                         /*
13605                          * reduce the number to keep us under
13606                          * a uint32_t.
13607                          */
13608                         rack->r_ctl.ack_count /= 2;
13609                         rack->r_ctl.sack_count /= 2;
13610                 }
13611                 if (tp->t_flags & TF_NEEDSYN) {
13612                         /*
13613                          * T/TCP: Connection was half-synchronized, and our SYN has
13614                          * been ACK'd (so connection is now fully synchronized).  Go
13615                          * to non-starred state, increment snd_una for ACK of SYN,
13616                          * and check if we can do window scaling.
13617                          */
13618                         tp->t_flags &= ~TF_NEEDSYN;
13619                         tp->snd_una++;
13620                         acked_amount = acked = (high_seq - tp->snd_una);
13621                 }
13622                 if (acked > sbavail(&so->so_snd))
13623                         acked_amount = sbavail(&so->so_snd);
13624 #ifdef NETFLIX_EXP_DETECTION
13625                 /*
13626                  * We only care on a cum-ack move if we are in a sack-disabled
13627                  * state. We have already added in to the ack_count, and we never
13628                  * would disable on a cum-ack move, so we only care to do the
13629                  * detection if it may "undo" it, i.e. we were in disabled already.
13630                  */
13631                 if (rack->sack_attack_disable)
13632                         rack_do_detection(tp, rack, acked_amount, segsiz);
13633 #endif
13634                 if (IN_FASTRECOVERY(tp->t_flags) &&
13635                     (rack->rack_no_prr == 0))
13636                         rack_update_prr(tp, rack, acked_amount, high_seq);
13637                 if (IN_RECOVERY(tp->t_flags)) {
13638                         if (SEQ_LT(high_seq, tp->snd_recover) &&
13639                             (SEQ_LT(high_seq, tp->snd_max))) {
13640                                 tcp_rack_partialack(tp);
13641                         } else {
13642                                 rack_post_recovery(tp, high_seq);
13643                                 recovery = 1;
13644                         }
13645                 }
13646                 /* Handle the rack-log-ack part (sendmap) */
13647                 if ((sbused(&so->so_snd) == 0) &&
13648                     (acked > acked_amount) &&
13649                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
13650                     (tp->t_flags & TF_SENTFIN)) {
13651                         /*
13652                          * We must be sure our fin
13653                          * was sent and acked (we can be
13654                          * in FIN_WAIT_1 without having
13655                          * sent the fin).
13656                          */
13657                         ourfinisacked = 1;
13658                         /*
13659                          * Lets make sure snd_una is updated
13660                          * since most likely acked_amount = 0 (it
13661                          * should be).
13662                          */
13663                         tp->snd_una = high_seq;
13664                 }
13665                 /* Did we make a RTO error? */
13666                 if ((tp->t_flags & TF_PREVVALID) &&
13667                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13668                         tp->t_flags &= ~TF_PREVVALID;
13669                         if (tp->t_rxtshift == 1 &&
13670                             (int)(ticks - tp->t_badrxtwin) < 0)
13671                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
13672                 }
13673                 /* Handle the data in the socket buffer */
13674                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13675                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13676                 if (acked_amount > 0) {
13677                         struct mbuf *mfree;
13678
13679                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13680                         SOCKBUF_LOCK(&so->so_snd);
13681                         mfree = sbcut_locked(&so->so_snd, acked_amount);
13682                         tp->snd_una = high_seq;
13683                         /* Note we want to hold the sb lock through the sendmap adjust */
13684                         rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13685                         /* Wake up the socket if we have room to write more */
13686                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13687                         sowwakeup_locked(so);
13688                         m_freem(mfree);
13689                 }
13690                 /* update progress */
13691                 tp->t_acktime = ticks;
13692                 rack_log_progress_event(rack, tp, tp->t_acktime,
13693                                         PROGRESS_UPDATE, __LINE__);
13694                 /* Clear out shifts and such */
13695                 tp->t_rxtshift = 0;
13696                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13697                                    rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13698                 rack->rc_tlp_in_progress = 0;
13699                 rack->r_ctl.rc_tlp_cnt_out = 0;
13700                 /* Send recover and snd_nxt must be dragged along */
13701                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13702                         tp->snd_recover = tp->snd_una;
13703                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13704                         tp->snd_nxt = tp->snd_una;
13705                 /*
13706                  * If the RXT timer is running we want to
13707                  * stop it, so we can restart a TLP (or new RXT).
13708                  */
13709                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13710                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13711 #ifdef NETFLIX_HTTP_LOGGING
13712                 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13713 #endif
13714                 tp->snd_wl2 = high_seq;
13715                 tp->t_dupacks = 0;
13716                 if (under_pacing &&
13717                     (rack->use_fixed_rate == 0) &&
13718                     (rack->in_probe_rtt == 0) &&
13719                     rack->rc_gp_dyn_mul &&
13720                     rack->rc_always_pace) {
13721                         /* Check if we are dragging bottom */
13722                         rack_check_bottom_drag(tp, rack, so, acked);
13723                 }
13724                 if (tp->snd_una == tp->snd_max) {
13725                         tp->t_flags &= ~TF_PREVVALID;
13726                         rack->r_ctl.retran_during_recovery = 0;
13727                         rack->r_ctl.dsack_byte_cnt = 0;
13728                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13729                         if (rack->r_ctl.rc_went_idle_time == 0)
13730                                 rack->r_ctl.rc_went_idle_time = 1;
13731                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13732                         if (sbavail(&tptosocket(tp)->so_snd) == 0)
13733                                 tp->t_acktime = 0;
13734                         /* Set so we might enter persists... */
13735                         rack->r_wanted_output = 1;
13736                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13737                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13738                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13739                             (sbavail(&so->so_snd) == 0) &&
13740                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13741                                 /*
13742                                  * The socket was gone and the
13743                                  * peer sent data (not now in the past), time to
13744                                  * reset him.
13745                                  */
13746                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13747                                 /* tcp_close will kill the inp pre-log the Reset */
13748                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13749 #ifdef TCP_ACCOUNTING
13750                                 rdstc = get_cyclecount();
13751                                 if (rdstc > ts_val) {
13752                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13753                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13754                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13755                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13756                                         }
13757                                 }
13758 #endif
13759                                 m_freem(m);
13760                                 tp = tcp_close(tp);
13761                                 if (tp == NULL) {
13762 #ifdef TCP_ACCOUNTING
13763                                         sched_unpin();
13764 #endif
13765                                         return (1);
13766                                 }
13767                                 /*
13768                                  * We would normally do drop-with-reset which would
13769                                  * send back a reset. We can't since we don't have
13770                                  * all the needed bits. Instead lets arrange for
13771                                  * a call to tcp_output(). That way since we
13772                                  * are in the closed state we will generate a reset.
13773                                  *
13774                                  * Note if tcp_accounting is on we don't unpin since
13775                                  * we do that after the goto label.
13776                                  */
13777                                 goto send_out_a_rst;
13778                         }
13779                         if ((sbused(&so->so_snd) == 0) &&
13780                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
13781                             (tp->t_flags & TF_SENTFIN)) {
13782                                 /*
13783                                  * If we can't receive any more data, then closing user can
13784                                  * proceed. Starting the timer is contrary to the
13785                                  * specification, but if we don't get a FIN we'll hang
13786                                  * forever.
13787                                  *
13788                                  */
13789                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13790                                         soisdisconnected(so);
13791                                         tcp_timer_activate(tp, TT_2MSL,
13792                                                            (tcp_fast_finwait2_recycle ?
13793                                                             tcp_finwait2_timeout :
13794                                                             TP_MAXIDLE(tp)));
13795                                 }
13796                                 if (ourfinisacked == 0) {
13797                                         /*
13798                                          * We don't change to fin-wait-2 if we have our fin acked
13799                                          * which means we are probably in TCPS_CLOSING.
13800                                          */
13801                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13802                                 }
13803                         }
13804                 }
13805                 /* Wake up the socket if we have room to write more */
13806                 if (sbavail(&so->so_snd)) {
13807                         rack->r_wanted_output = 1;
13808                         if (ctf_progress_timeout_check(tp, true)) {
13809                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13810                                                         tp, tick, PROGRESS_DROP, __LINE__);
13811                                 /*
13812                                  * We cheat here and don't send a RST, we should send one
13813                                  * when the pacer drops the connection.
13814                                  */
13815 #ifdef TCP_ACCOUNTING
13816                                 rdstc = get_cyclecount();
13817                                 if (rdstc > ts_val) {
13818                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13819                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13820                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13821                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13822                                         }
13823                                 }
13824                                 sched_unpin();
13825 #endif
13826                                 (void)tcp_drop(tp, ETIMEDOUT);
13827                                 m_freem(m);
13828                                 return (1);
13829                         }
13830                 }
13831                 if (ourfinisacked) {
13832                         switch(tp->t_state) {
13833                         case TCPS_CLOSING:
13834 #ifdef TCP_ACCOUNTING
13835                                 rdstc = get_cyclecount();
13836                                 if (rdstc > ts_val) {
13837                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13838                                                         (rdstc - ts_val));
13839                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13840                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13841                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13842                                         }
13843                                 }
13844                                 sched_unpin();
13845 #endif
13846                                 tcp_twstart(tp);
13847                                 m_freem(m);
13848                                 return (1);
13849                                 break;
13850                         case TCPS_LAST_ACK:
13851 #ifdef TCP_ACCOUNTING
13852                                 rdstc = get_cyclecount();
13853                                 if (rdstc > ts_val) {
13854                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13855                                                         (rdstc - ts_val));
13856                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13857                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13858                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13859                                         }
13860                                 }
13861                                 sched_unpin();
13862 #endif
13863                                 tp = tcp_close(tp);
13864                                 ctf_do_drop(m, tp);
13865                                 return (1);
13866                                 break;
13867                         case TCPS_FIN_WAIT_1:
13868 #ifdef TCP_ACCOUNTING
13869                                 rdstc = get_cyclecount();
13870                                 if (rdstc > ts_val) {
13871                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13872                                                         (rdstc - ts_val));
13873                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13874                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13875                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13876                                         }
13877                                 }
13878 #endif
13879                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13880                                         soisdisconnected(so);
13881                                         tcp_timer_activate(tp, TT_2MSL,
13882                                                            (tcp_fast_finwait2_recycle ?
13883                                                             tcp_finwait2_timeout :
13884                                                             TP_MAXIDLE(tp)));
13885                                 }
13886                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13887                                 break;
13888                         default:
13889                                 break;
13890                         }
13891                 }
13892                 if (rack->r_fast_output) {
13893                         /*
13894                          * We re doing fast output.. can we expand that?
13895                          */
13896                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13897                 }
13898 #ifdef TCP_ACCOUNTING
13899                 rdstc = get_cyclecount();
13900                 if (rdstc > ts_val) {
13901                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13902                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13903                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13904                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13905                         }
13906                 }
13907
13908         } else if (win_up_req) {
13909                 rdstc = get_cyclecount();
13910                 if (rdstc > ts_val) {
13911                         counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13912                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13913                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13914                         }
13915                 }
13916 #endif
13917         }
13918         /* Now is there a next packet, if so we are done */
13919         m_freem(m);
13920         did_out = 0;
13921         if (nxt_pkt) {
13922 #ifdef TCP_ACCOUNTING
13923                 sched_unpin();
13924 #endif
13925                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13926                 return (0);
13927         }
13928         rack_handle_might_revert(tp, rack);
13929         ctf_calc_rwin(so, tp);
13930         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13931         send_out_a_rst:
13932                 if (tcp_output(tp) < 0) {
13933 #ifdef TCP_ACCOUNTING
13934                         sched_unpin();
13935 #endif
13936                         return (1);
13937                 }
13938                 did_out = 1;
13939         }
13940         rack_free_trim(rack);
13941 #ifdef TCP_ACCOUNTING
13942         sched_unpin();
13943 #endif
13944         rack_timer_audit(tp, rack, &so->so_snd);
13945         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13946         return (0);
13947 }
13948
13949
13950 static int
13951 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13952     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13953     int32_t nxt_pkt, struct timeval *tv)
13954 {
13955         struct inpcb *inp = tptoinpcb(tp);
13956 #ifdef TCP_ACCOUNTING
13957         uint64_t ts_val;
13958 #endif
13959         int32_t thflags, retval, did_out = 0;
13960         int32_t way_out = 0;
13961         /*
13962          * cts - is the current time from tv (caller gets ts) in microseconds.
13963          * ms_cts - is the current time from tv in milliseconds.
13964          * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
13965          */
13966         uint32_t cts, us_cts, ms_cts;
13967         uint32_t tiwin, high_seq;
13968         struct timespec ts;
13969         struct tcpopt to;
13970         struct tcp_rack *rack;
13971         struct rack_sendmap *rsm;
13972         int32_t prev_state = 0;
13973 #ifdef TCP_ACCOUNTING
13974         int ack_val_set = 0xf;
13975 #endif
13976         int nsegs;
13977
13978         NET_EPOCH_ASSERT();
13979         INP_WLOCK_ASSERT(inp);
13980
13981         /*
13982          * tv passed from common code is from either M_TSTMP_LRO or
13983          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13984          */
13985         rack = (struct tcp_rack *)tp->t_fb_ptr;
13986         if (m->m_flags & M_ACKCMP) {
13987                 /*
13988                  * All compressed ack's are ack's by definition so
13989                  * remove any ack required flag and then do the processing.
13990                  */
13991                 rack->rc_ack_required = 0;
13992                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13993         }
13994         if (m->m_flags & M_ACKCMP) {
13995                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13996         }
13997         cts = tcp_tv_to_usectick(tv);
13998         ms_cts =  tcp_tv_to_mssectick(tv);
13999         nsegs = m->m_pkthdr.lro_nsegs;
14000         counter_u64_add(rack_proc_non_comp_ack, 1);
14001         thflags = tcp_get_flags(th);
14002 #ifdef TCP_ACCOUNTING
14003         sched_pin();
14004         if (thflags & TH_ACK)
14005                 ts_val = get_cyclecount();
14006 #endif
14007         if ((m->m_flags & M_TSTMP) ||
14008             (m->m_flags & M_TSTMP_LRO)) {
14009                 mbuf_tstmp2timespec(m, &ts);
14010                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
14011                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
14012         } else
14013                 rack->r_ctl.act_rcv_time = *tv;
14014         kern_prefetch(rack, &prev_state);
14015         prev_state = 0;
14016         /*
14017          * Unscale the window into a 32-bit value. For the SYN_SENT state
14018          * the scale is zero.
14019          */
14020         tiwin = th->th_win << tp->snd_scale;
14021 #ifdef TCP_ACCOUNTING
14022         if (thflags & TH_ACK) {
14023                 /*
14024                  * We have a tradeoff here. We can either do what we are
14025                  * doing i.e. pinning to this CPU and then doing the accounting
14026                  * <or> we could do a critical enter, setup the rdtsc and cpu
14027                  * as in below, and then validate we are on the same CPU on
14028                  * exit. I have choosen to not do the critical enter since
14029                  * that often will gain you a context switch, and instead lock
14030                  * us (line above this if) to the same CPU with sched_pin(). This
14031                  * means we may be context switched out for a higher priority
14032                  * interupt but we won't be moved to another CPU.
14033                  *
14034                  * If this occurs (which it won't very often since we most likely
14035                  * are running this code in interupt context and only a higher
14036                  * priority will bump us ... clock?) we will falsely add in
14037                  * to the time the interupt processing time plus the ack processing
14038                  * time. This is ok since its a rare event.
14039                  */
14040                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
14041                                                     ctf_fixed_maxseg(tp));
14042         }
14043 #endif
14044         /*
14045          * Parse options on any incoming segment.
14046          */
14047         memset(&to, 0, sizeof(to));
14048         tcp_dooptions(&to, (u_char *)(th + 1),
14049             (th->th_off << 2) - sizeof(struct tcphdr),
14050             (thflags & TH_SYN) ? TO_SYN : 0);
14051         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
14052             __func__));
14053         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
14054             __func__));
14055
14056         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
14057             (tp->t_flags & TF_GPUTINPROG)) {
14058                 /*
14059                  * We have a goodput in progress
14060                  * and we have entered a late state.
14061                  * Do we have enough data in the sb
14062                  * to handle the GPUT request?
14063                  */
14064                 uint32_t bytes;
14065
14066                 bytes = tp->gput_ack - tp->gput_seq;
14067                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
14068                         bytes += tp->gput_seq - tp->snd_una;
14069                 if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
14070                         /*
14071                          * There are not enough bytes in the socket
14072                          * buffer that have been sent to cover this
14073                          * measurement. Cancel it.
14074                          */
14075                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14076                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
14077                                                    tp->gput_seq,
14078                                                    0, 0, 18, __LINE__, NULL, 0);
14079                         tp->t_flags &= ~TF_GPUTINPROG;
14080                 }
14081         }
14082         high_seq = th->th_ack;
14083         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14084                 union tcp_log_stackspecific log;
14085                 struct timeval ltv;
14086 #ifdef NETFLIX_HTTP_LOGGING
14087                 struct http_sendfile_track *http_req;
14088
14089                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
14090                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14091                 } else {
14092                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14093                 }
14094 #endif
14095                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14096                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14097                 if (rack->rack_no_prr == 0)
14098                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14099                 else
14100                         log.u_bbr.flex1 = 0;
14101                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14102                 log.u_bbr.use_lt_bw <<= 1;
14103                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
14104                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14105                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14106                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14107                 log.u_bbr.flex3 = m->m_flags;
14108                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14109                 log.u_bbr.lost = thflags;
14110                 log.u_bbr.pacing_gain = 0x1;
14111 #ifdef TCP_ACCOUNTING
14112                 log.u_bbr.cwnd_gain = ack_val_set;
14113 #endif
14114                 log.u_bbr.flex7 = 2;
14115                 if (m->m_flags & M_TSTMP) {
14116                         /* Record the hardware timestamp if present */
14117                         mbuf_tstmp2timespec(m, &ts);
14118                         ltv.tv_sec = ts.tv_sec;
14119                         ltv.tv_usec = ts.tv_nsec / 1000;
14120                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14121                 } else if (m->m_flags & M_TSTMP_LRO) {
14122                         /* Record the LRO the arrival timestamp */
14123                         mbuf_tstmp2timespec(m, &ts);
14124                         ltv.tv_sec = ts.tv_sec;
14125                         ltv.tv_usec = ts.tv_nsec / 1000;
14126                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14127                 }
14128                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14129                 /* Log the rcv time */
14130                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14131 #ifdef NETFLIX_HTTP_LOGGING
14132                 log.u_bbr.applimited = tp->t_http_closed;
14133                 log.u_bbr.applimited <<= 8;
14134                 log.u_bbr.applimited |= tp->t_http_open;
14135                 log.u_bbr.applimited <<= 8;
14136                 log.u_bbr.applimited |= tp->t_http_req;
14137                 if (http_req) {
14138                         /* Copy out any client req info */
14139                         /* seconds */
14140                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14141                         /* useconds */
14142                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14143                         log.u_bbr.rttProp = http_req->timestamp;
14144                         log.u_bbr.cur_del_rate = http_req->start;
14145                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14146                                 log.u_bbr.flex8 |= 1;
14147                         } else {
14148                                 log.u_bbr.flex8 |= 2;
14149                                 log.u_bbr.bw_inuse = http_req->end;
14150                         }
14151                         log.u_bbr.flex6 = http_req->start_seq;
14152                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14153                                 log.u_bbr.flex8 |= 4;
14154                                 log.u_bbr.epoch = http_req->end_seq;
14155                         }
14156                 }
14157 #endif
14158                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14159                     tlen, &log, true, &ltv);
14160         }
14161         /* Remove ack required flag if set, we have one  */
14162         if (thflags & TH_ACK)
14163                 rack->rc_ack_required = 0;
14164         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14165                 way_out = 4;
14166                 retval = 0;
14167                 m_freem(m);
14168                 goto done_with_input;
14169         }
14170         /*
14171          * If a segment with the ACK-bit set arrives in the SYN-SENT state
14172          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14173          */
14174         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14175             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14176                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14177                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14178 #ifdef TCP_ACCOUNTING
14179                 sched_unpin();
14180 #endif
14181                 return (1);
14182         }
14183         /*
14184          * If timestamps were negotiated during SYN/ACK and a
14185          * segment without a timestamp is received, silently drop
14186          * the segment, unless it is a RST segment or missing timestamps are
14187          * tolerated.
14188          * See section 3.2 of RFC 7323.
14189          */
14190         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14191             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14192                 way_out = 5;
14193                 retval = 0;
14194                 m_freem(m);
14195                 goto done_with_input;
14196         }
14197
14198         /*
14199          * Segment received on connection. Reset idle time and keep-alive
14200          * timer. XXX: This should be done after segment validation to
14201          * ignore broken/spoofed segs.
14202          */
14203         if  (tp->t_idle_reduce &&
14204              (tp->snd_max == tp->snd_una) &&
14205              (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14206                 counter_u64_add(rack_input_idle_reduces, 1);
14207                 rack_cc_after_idle(rack, tp);
14208         }
14209         tp->t_rcvtime = ticks;
14210 #ifdef STATS
14211         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14212 #endif
14213         if (tiwin > rack->r_ctl.rc_high_rwnd)
14214                 rack->r_ctl.rc_high_rwnd = tiwin;
14215         /*
14216          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14217          * this to occur after we've validated the segment.
14218          */
14219         if (tcp_ecn_input_segment(tp, thflags, tlen,
14220             tcp_packets_this_ack(tp, th->th_ack),
14221             iptos))
14222                 rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
14223
14224         /*
14225          * If echoed timestamp is later than the current time, fall back to
14226          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14227          * were used when this connection was established.
14228          */
14229         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14230                 to.to_tsecr -= tp->ts_offset;
14231                 if (TSTMP_GT(to.to_tsecr, ms_cts))
14232                         to.to_tsecr = 0;
14233         }
14234
14235         /*
14236          * If its the first time in we need to take care of options and
14237          * verify we can do SACK for rack!
14238          */
14239         if (rack->r_state == 0) {
14240                 /* Should be init'd by rack_init() */
14241                 KASSERT(rack->rc_inp != NULL,
14242                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
14243                 if (rack->rc_inp == NULL) {
14244                         rack->rc_inp = inp;
14245                 }
14246
14247                 /*
14248                  * Process options only when we get SYN/ACK back. The SYN
14249                  * case for incoming connections is handled in tcp_syncache.
14250                  * According to RFC1323 the window field in a SYN (i.e., a
14251                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14252                  * this is traditional behavior, may need to be cleaned up.
14253                  */
14254                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14255                         /* Handle parallel SYN for ECN */
14256                         tcp_ecn_input_parallel_syn(tp, thflags, iptos);
14257                         if ((to.to_flags & TOF_SCALE) &&
14258                             (tp->t_flags & TF_REQ_SCALE)) {
14259                                 tp->t_flags |= TF_RCVD_SCALE;
14260                                 tp->snd_scale = to.to_wscale;
14261                         } else
14262                                 tp->t_flags &= ~TF_REQ_SCALE;
14263                         /*
14264                          * Initial send window.  It will be updated with the
14265                          * next incoming segment to the scaled value.
14266                          */
14267                         tp->snd_wnd = th->th_win;
14268                         rack_validate_fo_sendwin_up(tp, rack);
14269                         if ((to.to_flags & TOF_TS) &&
14270                             (tp->t_flags & TF_REQ_TSTMP)) {
14271                                 tp->t_flags |= TF_RCVD_TSTMP;
14272                                 tp->ts_recent = to.to_tsval;
14273                                 tp->ts_recent_age = cts;
14274                         } else
14275                                 tp->t_flags &= ~TF_REQ_TSTMP;
14276                         if (to.to_flags & TOF_MSS) {
14277                                 tcp_mss(tp, to.to_mss);
14278                         }
14279                         if ((tp->t_flags & TF_SACK_PERMIT) &&
14280                             (to.to_flags & TOF_SACKPERM) == 0)
14281                                 tp->t_flags &= ~TF_SACK_PERMIT;
14282                         if (IS_FASTOPEN(tp->t_flags)) {
14283                                 if (to.to_flags & TOF_FASTOPEN) {
14284                                         uint16_t mss;
14285
14286                                         if (to.to_flags & TOF_MSS)
14287                                                 mss = to.to_mss;
14288                                         else
14289                                                 if ((inp->inp_vflag & INP_IPV6) != 0)
14290                                                         mss = TCP6_MSS;
14291                                                 else
14292                                                         mss = TCP_MSS;
14293                                         tcp_fastopen_update_cache(tp, mss,
14294                                             to.to_tfo_len, to.to_tfo_cookie);
14295                                 } else
14296                                         tcp_fastopen_disable_path(tp);
14297                         }
14298                 }
14299                 /*
14300                  * At this point we are at the initial call. Here we decide
14301                  * if we are doing RACK or not. We do this by seeing if
14302                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
14303                  * The code now does do dup-ack counting so if you don't
14304                  * switch back you won't get rack & TLP, but you will still
14305                  * get this stack.
14306                  */
14307
14308                 if ((rack_sack_not_required == 0) &&
14309                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14310                         tcp_switch_back_to_default(tp);
14311                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14312                             tlen, iptos);
14313 #ifdef TCP_ACCOUNTING
14314                         sched_unpin();
14315 #endif
14316                         return (1);
14317                 }
14318                 tcp_set_hpts(inp);
14319                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14320         }
14321         if (thflags & TH_FIN)
14322                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14323         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14324         if ((rack->rc_gp_dyn_mul) &&
14325             (rack->use_fixed_rate == 0) &&
14326             (rack->rc_always_pace)) {
14327                 /* Check in on probertt */
14328                 rack_check_probe_rtt(rack, us_cts);
14329         }
14330         rack_clear_rate_sample(rack);
14331         if ((rack->forced_ack) &&
14332             ((tcp_get_flags(th) & TH_RST) == 0)) {
14333                 rack_handle_probe_response(rack, tiwin, us_cts);
14334         }
14335         /*
14336          * This is the one exception case where we set the rack state
14337          * always. All other times (timers etc) we must have a rack-state
14338          * set (so we assure we have done the checks above for SACK).
14339          */
14340         rack->r_ctl.rc_rcvtime = cts;
14341         if (rack->r_state != tp->t_state)
14342                 rack_set_state(tp, rack);
14343         if (SEQ_GT(th->th_ack, tp->snd_una) &&
14344             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14345                 kern_prefetch(rsm, &prev_state);
14346         prev_state = rack->r_state;
14347         retval = (*rack->r_substate) (m, th, so,
14348             tp, &to, drop_hdrlen,
14349             tlen, tiwin, thflags, nxt_pkt, iptos);
14350         if (retval == 0) {
14351                 /*
14352                  * If retval is 1 the tcb is unlocked and most likely the tp
14353                  * is gone.
14354                  */
14355                 INP_WLOCK_ASSERT(inp);
14356                 if ((rack->rc_gp_dyn_mul) &&
14357                     (rack->rc_always_pace) &&
14358                     (rack->use_fixed_rate == 0) &&
14359                     rack->in_probe_rtt &&
14360                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
14361                         /*
14362                          * If we are going for target, lets recheck before
14363                          * we output.
14364                          */
14365                         rack_check_probe_rtt(rack, us_cts);
14366                 }
14367                 if (rack->set_pacing_done_a_iw == 0) {
14368                         /* How much has been acked? */
14369                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14370                                 /* We have enough to set in the pacing segment size */
14371                                 rack->set_pacing_done_a_iw = 1;
14372                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
14373                         }
14374                 }
14375                 tcp_rack_xmit_timer_commit(rack, tp);
14376 #ifdef TCP_ACCOUNTING
14377                 /*
14378                  * If we set the ack_val_se to what ack processing we are doing
14379                  * we also want to track how many cycles we burned. Note
14380                  * the bits after tcp_output we let be "free". This is because
14381                  * we are also tracking the tcp_output times as well. Note the
14382                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
14383                  * 0xf cannot be returned and is what we initialize it too to
14384                  * indicate we are not doing the tabulations.
14385                  */
14386                 if (ack_val_set != 0xf) {
14387                         uint64_t crtsc;
14388
14389                         crtsc = get_cyclecount();
14390                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14391                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14392                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14393                         }
14394                 }
14395 #endif
14396                 if (nxt_pkt == 0) {
14397                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14398 do_output_now:
14399                                 if (tcp_output(tp) < 0)
14400                                         return (1);
14401                                 did_out = 1;
14402                         }
14403                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14404                         rack_free_trim(rack);
14405                 }
14406                 /* Update any rounds needed */
14407                 if (rack_verbose_logging &&  (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
14408                         union tcp_log_stackspecific log;
14409                         struct timeval tv;
14410
14411                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14412                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14413                         log.u_bbr.flex1 = high_seq;
14414                         log.u_bbr.flex2 = rack->r_ctl.roundends;
14415                         log.u_bbr.flex3 = rack->r_ctl.current_round;
14416                         log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
14417                         log.u_bbr.flex8 = 9;
14418                         tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
14419                                        0, &log, false, NULL, NULL, 0, &tv);
14420                 }
14421                 /*
14422                  * The draft (v3) calls for us to use SEQ_GEQ, but that
14423                  * causes issues when we are just going app limited. Lets
14424                  * instead use SEQ_GT <or> where its equal but more data
14425                  * is outstanding.
14426                  */
14427                 if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) ||
14428                     ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) {
14429                         rack->r_ctl.current_round++;
14430                         rack->r_ctl.roundends = tp->snd_max;
14431                         if (CC_ALGO(tp)->newround != NULL) {
14432                                 CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
14433                         }
14434                 }
14435                 if ((nxt_pkt == 0) &&
14436                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14437                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
14438                      (tp->t_flags & TF_DELACK) ||
14439                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14440                       (tp->t_state <= TCPS_CLOSING)))) {
14441                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
14442                         if ((tp->snd_max == tp->snd_una) &&
14443                             ((tp->t_flags & TF_DELACK) == 0) &&
14444                             (tcp_in_hpts(rack->rc_inp)) &&
14445                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14446                                 /* keep alive not needed if we are hptsi output yet */
14447                                 ;
14448                         } else {
14449                                 int late = 0;
14450                                 if (tcp_in_hpts(inp)) {
14451                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14452                                                 us_cts = tcp_get_usecs(NULL);
14453                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14454                                                         rack->r_early = 1;
14455                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14456                                                 } else
14457                                                         late = 1;
14458                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14459                                         }
14460                                         tcp_hpts_remove(inp);
14461                                 }
14462                                 if (late && (did_out == 0)) {
14463                                         /*
14464                                          * We are late in the sending
14465                                          * and we did not call the output
14466                                          * (this probably should not happen).
14467                                          */
14468                                         goto do_output_now;
14469                                 }
14470                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14471                         }
14472                         way_out = 1;
14473                 } else if (nxt_pkt == 0) {
14474                         /* Do we have the correct timer running? */
14475                         rack_timer_audit(tp, rack, &so->so_snd);
14476                         way_out = 2;
14477                 }
14478         done_with_input:
14479                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14480                 if (did_out)
14481                         rack->r_wanted_output = 0;
14482 #ifdef TCP_ACCOUNTING
14483         } else {
14484                 /*
14485                  * Track the time (see above).
14486                  */
14487                 if (ack_val_set != 0xf) {
14488                         uint64_t crtsc;
14489
14490                         crtsc = get_cyclecount();
14491                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14492                         /*
14493                          * Note we *DO NOT* increment the per-tcb counters since
14494                          * in the else the TP may be gone!!
14495                          */
14496                 }
14497 #endif
14498         }
14499 #ifdef TCP_ACCOUNTING
14500         sched_unpin();
14501 #endif
14502         return (retval);
14503 }
14504
14505 void
14506 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14507     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14508 {
14509         struct timeval tv;
14510
14511         /* First lets see if we have old packets */
14512         if (tp->t_in_pkt) {
14513                 if (ctf_do_queued_segments(so, tp, 1)) {
14514                         m_freem(m);
14515                         return;
14516                 }
14517         }
14518         if (m->m_flags & M_TSTMP_LRO) {
14519                 mbuf_tstmp2timeval(m, &tv);
14520         } else {
14521                 /* Should not be should we kassert instead? */
14522                 tcp_get_usecs(&tv);
14523         }
14524         if (rack_do_segment_nounlock(m, th, so, tp,
14525                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14526                 INP_WUNLOCK(tptoinpcb(tp));
14527         }
14528 }
14529
14530 struct rack_sendmap *
14531 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14532 {
14533         struct rack_sendmap *rsm = NULL;
14534         int32_t idx;
14535         uint32_t srtt = 0, thresh = 0, ts_low = 0;
14536
14537         /* Return the next guy to be re-transmitted */
14538         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14539                 return (NULL);
14540         }
14541         if (tp->t_flags & TF_SENTFIN) {
14542                 /* retran the end FIN? */
14543                 return (NULL);
14544         }
14545         /* ok lets look at this one */
14546         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14547         if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
14548                 return (rsm);
14549         }
14550         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14551                 goto check_it;
14552         }
14553         rsm = rack_find_lowest_rsm(rack);
14554         if (rsm == NULL) {
14555                 return (NULL);
14556         }
14557 check_it:
14558         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14559             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14560                 /*
14561                  * No sack so we automatically do the 3 strikes and
14562                  * retransmit (no rack timer would be started).
14563                  */
14564
14565                 return (rsm);
14566         }
14567         if (rsm->r_flags & RACK_ACKED) {
14568                 return (NULL);
14569         }
14570         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14571             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14572                 /* Its not yet ready */
14573                 return (NULL);
14574         }
14575         srtt = rack_grab_rtt(tp, rack);
14576         idx = rsm->r_rtr_cnt - 1;
14577         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14578         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14579         if ((tsused == ts_low) ||
14580             (TSTMP_LT(tsused, ts_low))) {
14581                 /* No time since sending */
14582                 return (NULL);
14583         }
14584         if ((tsused - ts_low) < thresh) {
14585                 /* It has not been long enough yet */
14586                 return (NULL);
14587         }
14588         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14589             ((rsm->r_flags & RACK_SACK_PASSED) &&
14590              (rack->sack_attack_disable == 0))) {
14591                 /*
14592                  * We have passed the dup-ack threshold <or>
14593                  * a SACK has indicated this is missing.
14594                  * Note that if you are a declared attacker
14595                  * it is only the dup-ack threshold that
14596                  * will cause retransmits.
14597                  */
14598                 /* log retransmit reason */
14599                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14600                 rack->r_fast_output = 0;
14601                 return (rsm);
14602         }
14603         return (NULL);
14604 }
14605
14606 static void
14607 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14608                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14609                            int line, struct rack_sendmap *rsm, uint8_t quality)
14610 {
14611         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14612                 union tcp_log_stackspecific log;
14613                 struct timeval tv;
14614
14615                 memset(&log, 0, sizeof(log));
14616                 log.u_bbr.flex1 = slot;
14617                 log.u_bbr.flex2 = len;
14618                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14619                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14620                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14621                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14622                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14623                 log.u_bbr.use_lt_bw <<= 1;
14624                 log.u_bbr.use_lt_bw |= rack->r_late;
14625                 log.u_bbr.use_lt_bw <<= 1;
14626                 log.u_bbr.use_lt_bw |= rack->r_early;
14627                 log.u_bbr.use_lt_bw <<= 1;
14628                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14629                 log.u_bbr.use_lt_bw <<= 1;
14630                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14631                 log.u_bbr.use_lt_bw <<= 1;
14632                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14633                 log.u_bbr.use_lt_bw <<= 1;
14634                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14635                 log.u_bbr.use_lt_bw <<= 1;
14636                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14637                 log.u_bbr.pkt_epoch = line;
14638                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14639                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14640                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14641                 log.u_bbr.bw_inuse = bw_est;
14642                 log.u_bbr.delRate = bw;
14643                 if (rack->r_ctl.gp_bw == 0)
14644                         log.u_bbr.cur_del_rate = 0;
14645                 else
14646                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14647                 log.u_bbr.rttProp = len_time;
14648                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14649                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14650                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14651                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14652                         /* We are in slow start */
14653                         log.u_bbr.flex7 = 1;
14654                 } else {
14655                         /* we are on congestion avoidance */
14656                         log.u_bbr.flex7 = 0;
14657                 }
14658                 log.u_bbr.flex8 = method;
14659                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14660                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14661                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14662                 log.u_bbr.cwnd_gain <<= 1;
14663                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14664                 log.u_bbr.cwnd_gain <<= 1;
14665                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14666                 log.u_bbr.bbr_substate = quality;
14667                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14668                     &rack->rc_inp->inp_socket->so_rcv,
14669                     &rack->rc_inp->inp_socket->so_snd,
14670                     BBR_LOG_HPTSI_CALC, 0,
14671                     0, &log, false, &tv);
14672         }
14673 }
14674
14675 static uint32_t
14676 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14677 {
14678         uint32_t new_tso, user_max;
14679
14680         user_max = rack->rc_user_set_max_segs * mss;
14681         if (rack->rc_force_max_seg) {
14682                 return (user_max);
14683         }
14684         if (rack->use_fixed_rate &&
14685             ((rack->r_ctl.crte == NULL) ||
14686              (bw != rack->r_ctl.crte->rate))) {
14687                 /* Use the user mss since we are not exactly matched */
14688                 return (user_max);
14689         }
14690         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14691         if (new_tso > user_max)
14692                 new_tso = user_max;
14693         return (new_tso);
14694 }
14695
14696 static int32_t
14697 pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced)
14698 {
14699         uint64_t lentim, fill_bw;
14700
14701         /* Lets first see if we are full, if so continue with normal rate */
14702         rack->r_via_fill_cw = 0;
14703         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14704                 return (slot);
14705         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14706                 return (slot);
14707         if (rack->r_ctl.rc_last_us_rtt == 0)
14708                 return (slot);
14709         if (rack->rc_pace_fill_if_rttin_range &&
14710             (rack->r_ctl.rc_last_us_rtt >=
14711              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14712                 /* The rtt is huge, N * smallest, lets not fill */
14713                 return (slot);
14714         }
14715         /*
14716          * first lets calculate the b/w based on the last us-rtt
14717          * and the sndwnd.
14718          */
14719         fill_bw = rack->r_ctl.cwnd_to_use;
14720         /* Take the rwnd if its smaller */
14721         if (fill_bw > rack->rc_tp->snd_wnd)
14722                 fill_bw = rack->rc_tp->snd_wnd;
14723         if (rack->r_fill_less_agg) {
14724                 /*
14725                  * Now take away the inflight (this will reduce our
14726                  * aggressiveness and yeah, if we get that much out in 1RTT
14727                  * we will have had acks come back and still be behind).
14728                  */
14729                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14730         }
14731         /* Now lets make it into a b/w */
14732         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14733         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14734         /* We are below the min b/w */
14735         if (non_paced)
14736                 *rate_wanted = fill_bw;
14737         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14738                 return (slot);
14739         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14740                 fill_bw = rack->r_ctl.bw_rate_cap;
14741         rack->r_via_fill_cw = 1;
14742         if (rack->r_rack_hw_rate_caps &&
14743             (rack->r_ctl.crte != NULL)) {
14744                 uint64_t high_rate;
14745
14746                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14747                 if (fill_bw > high_rate) {
14748                         /* We are capping bw at the highest rate table entry */
14749                         if (*rate_wanted > high_rate) {
14750                                 /* The original rate was also capped */
14751                                 rack->r_via_fill_cw = 0;
14752                         }
14753                         rack_log_hdwr_pacing(rack,
14754                                              fill_bw, high_rate, __LINE__,
14755                                              0, 3);
14756                         fill_bw = high_rate;
14757                         if (capped)
14758                                 *capped = 1;
14759                 }
14760         } else if ((rack->r_ctl.crte == NULL) &&
14761                    (rack->rack_hdrw_pacing == 0) &&
14762                    (rack->rack_hdw_pace_ena) &&
14763                    rack->r_rack_hw_rate_caps &&
14764                    (rack->rack_attempt_hdwr_pace == 0) &&
14765                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14766                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14767                 /*
14768                  * Ok we may have a first attempt that is greater than our top rate
14769                  * lets check.
14770                  */
14771                 uint64_t high_rate;
14772
14773                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14774                 if (high_rate) {
14775                         if (fill_bw > high_rate) {
14776                                 fill_bw = high_rate;
14777                                 if (capped)
14778                                         *capped = 1;
14779                         }
14780                 }
14781         }
14782         /*
14783          * Ok fill_bw holds our mythical b/w to fill the cwnd
14784          * in a rtt, what does that time wise equate too?
14785          */
14786         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14787         lentim /= fill_bw;
14788         *rate_wanted = fill_bw;
14789         if (non_paced || (lentim < slot)) {
14790                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14791                                            0, lentim, 12, __LINE__, NULL, 0);
14792                 return ((int32_t)lentim);
14793         } else
14794                 return (slot);
14795 }
14796
14797 static int32_t
14798 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14799 {
14800         uint64_t srtt;
14801         int32_t slot = 0;
14802         int can_start_hw_pacing = 1;
14803         int err;
14804
14805         if (rack->rc_always_pace == 0) {
14806                 /*
14807                  * We use the most optimistic possible cwnd/srtt for
14808                  * sending calculations. This will make our
14809                  * calculation anticipate getting more through
14810                  * quicker then possible. But thats ok we don't want
14811                  * the peer to have a gap in data sending.
14812                  */
14813                 uint64_t cwnd, tr_perms = 0;
14814                 int32_t reduce = 0;
14815
14816         old_method:
14817                 /*
14818                  * We keep no precise pacing with the old method
14819                  * instead we use the pacer to mitigate bursts.
14820                  */
14821                 if (rack->r_ctl.rc_rack_min_rtt)
14822                         srtt = rack->r_ctl.rc_rack_min_rtt;
14823                 else
14824                         srtt = max(tp->t_srtt, 1);
14825                 if (rack->r_ctl.rc_rack_largest_cwnd)
14826                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14827                 else
14828                         cwnd = rack->r_ctl.cwnd_to_use;
14829                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14830                 tr_perms = (cwnd * 1000) / srtt;
14831                 if (tr_perms == 0) {
14832                         tr_perms = ctf_fixed_maxseg(tp);
14833                 }
14834                 /*
14835                  * Calculate how long this will take to drain, if
14836                  * the calculation comes out to zero, thats ok we
14837                  * will use send_a_lot to possibly spin around for
14838                  * more increasing tot_len_this_send to the point
14839                  * that its going to require a pace, or we hit the
14840                  * cwnd. Which in that case we are just waiting for
14841                  * a ACK.
14842                  */
14843                 slot = len / tr_perms;
14844                 /* Now do we reduce the time so we don't run dry? */
14845                 if (slot && rack_slot_reduction) {
14846                         reduce = (slot / rack_slot_reduction);
14847                         if (reduce < slot) {
14848                                 slot -= reduce;
14849                         } else
14850                                 slot = 0;
14851                 }
14852                 slot *= HPTS_USEC_IN_MSEC;
14853                 if (rack->rc_pace_to_cwnd) {
14854                         uint64_t rate_wanted = 0;
14855
14856                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14857                         rack->rc_ack_can_sendout_data = 1;
14858                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14859                 } else
14860                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14861         } else {
14862                 uint64_t bw_est, res, lentim, rate_wanted;
14863                 uint32_t orig_val, segs, oh;
14864                 int capped = 0;
14865                 int prev_fill;
14866
14867                 if ((rack->r_rr_config == 1) && rsm) {
14868                         return (rack->r_ctl.rc_min_to);
14869                 }
14870                 if (rack->use_fixed_rate) {
14871                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14872                 } else if ((rack->r_ctl.init_rate == 0) &&
14873 #ifdef NETFLIX_PEAKRATE
14874                            (rack->rc_tp->t_maxpeakrate == 0) &&
14875 #endif
14876                            (rack->r_ctl.gp_bw == 0)) {
14877                         /* no way to yet do an estimate */
14878                         bw_est = rate_wanted = 0;
14879                 } else {
14880                         bw_est = rack_get_bw(rack);
14881                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14882                 }
14883                 if ((bw_est == 0) || (rate_wanted == 0) ||
14884                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14885                         /*
14886                          * No way yet to make a b/w estimate or
14887                          * our raise is set incorrectly.
14888                          */
14889                         goto old_method;
14890                 }
14891                 /* We need to account for all the overheads */
14892                 segs = (len + segsiz - 1) / segsiz;
14893                 /*
14894                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
14895                  * and how much data we put in each packet. Yes this
14896                  * means we may be off if we are larger than 1500 bytes
14897                  * or smaller. But this just makes us more conservative.
14898                  */
14899                 if (rack_hw_rate_min &&
14900                     (bw_est < rack_hw_rate_min))
14901                         can_start_hw_pacing = 0;
14902                 if (ETHERNET_SEGMENT_SIZE > segsiz)
14903                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
14904                 else
14905                         oh = 0;
14906                 segs *= oh;
14907                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14908                 res = lentim / rate_wanted;
14909                 slot = (uint32_t)res;
14910                 orig_val = rack->r_ctl.rc_pace_max_segs;
14911                 if (rack->r_ctl.crte == NULL) {
14912                         /*
14913                          * Only do this if we are not hardware pacing
14914                          * since if we are doing hw-pacing below we will
14915                          * set make a call after setting up or changing
14916                          * the rate.
14917                          */
14918                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14919                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14920                         /*
14921                          * We lost our rate somehow, this can happen
14922                          * if the interface changed underneath us.
14923                          */
14924                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14925                         rack->r_ctl.crte = NULL;
14926                         /* Lets re-allow attempting to setup pacing */
14927                         rack->rack_hdrw_pacing = 0;
14928                         rack->rack_attempt_hdwr_pace = 0;
14929                         rack_log_hdwr_pacing(rack,
14930                                              rate_wanted, bw_est, __LINE__,
14931                                              0, 6);
14932                 }
14933                 /* Did we change the TSO size, if so log it */
14934                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14935                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
14936                 prev_fill = rack->r_via_fill_cw;
14937                 if ((rack->rc_pace_to_cwnd) &&
14938                     (capped == 0) &&
14939                     (rack->use_fixed_rate == 0) &&
14940                     (rack->in_probe_rtt == 0) &&
14941                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14942                         /*
14943                          * We want to pace at our rate *or* faster to
14944                          * fill the cwnd to the max if its not full.
14945                          */
14946                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14947                 }
14948                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14949                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14950                         if ((rack->rack_hdw_pace_ena) &&
14951                             (can_start_hw_pacing > 0) &&
14952                             (rack->rack_hdrw_pacing == 0) &&
14953                             (rack->rack_attempt_hdwr_pace == 0)) {
14954                                 /*
14955                                  * Lets attempt to turn on hardware pacing
14956                                  * if we can.
14957                                  */
14958                                 rack->rack_attempt_hdwr_pace = 1;
14959                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14960                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
14961                                                                        rate_wanted,
14962                                                                        RS_PACING_GEQ,
14963                                                                        &err, &rack->r_ctl.crte_prev_rate);
14964                                 if (rack->r_ctl.crte) {
14965                                         rack->rack_hdrw_pacing = 1;
14966                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14967                                                                                                  0, rack->r_ctl.crte,
14968                                                                                                  NULL);
14969                                         rack_log_hdwr_pacing(rack,
14970                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14971                                                              err, 0);
14972                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14973                                 } else {
14974                                         counter_u64_add(rack_hw_pace_init_fail, 1);
14975                                 }
14976                         } else if (rack->rack_hdrw_pacing &&
14977                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14978                                 /* Do we need to adjust our rate? */
14979                                 const struct tcp_hwrate_limit_table *nrte;
14980
14981                                 if (rack->r_up_only &&
14982                                     (rate_wanted < rack->r_ctl.crte->rate)) {
14983                                         /**
14984                                          * We have four possible states here
14985                                          * having to do with the previous time
14986                                          * and this time.
14987                                          *   previous  |  this-time
14988                                          * A)     0      |     0   -- fill_cw not in the picture
14989                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
14990                                          * C)     1      |     1   -- all rates from fill_cw
14991                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
14992                                          *
14993                                          * For case A, C and D we don't allow a drop. But for
14994                                          * case B where we now our on our steady rate we do
14995                                          * allow a drop.
14996                                          *
14997                                          */
14998                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14999                                                 goto done_w_hdwr;
15000                                 }
15001                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
15002                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
15003                                         if (rack_hw_rate_to_low &&
15004                                             (bw_est < rack_hw_rate_to_low)) {
15005                                                 /*
15006                                                  * The pacing rate is too low for hardware, but
15007                                                  * do allow hardware pacing to be restarted.
15008                                                  */
15009                                                 rack_log_hdwr_pacing(rack,
15010                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
15011                                                              0, 5);
15012                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15013                                                 rack->r_ctl.crte = NULL;
15014                                                 rack->rack_attempt_hdwr_pace = 0;
15015                                                 rack->rack_hdrw_pacing = 0;
15016                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15017                                                 goto done_w_hdwr;
15018                                         }
15019                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
15020                                                                    rack->rc_tp,
15021                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
15022                                                                    rate_wanted,
15023                                                                    RS_PACING_GEQ,
15024                                                                    &err, &rack->r_ctl.crte_prev_rate);
15025                                         if (nrte == NULL) {
15026                                                 /* Lost the rate */
15027                                                 rack->rack_hdrw_pacing = 0;
15028                                                 rack->r_ctl.crte = NULL;
15029                                                 rack_log_hdwr_pacing(rack,
15030                                                                      rate_wanted, 0, __LINE__,
15031                                                                      err, 1);
15032                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15033                                                 counter_u64_add(rack_hw_pace_lost, 1);
15034                                         } else if (nrte != rack->r_ctl.crte) {
15035                                                 rack->r_ctl.crte = nrte;
15036                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
15037                                                                                                          segsiz, 0,
15038                                                                                                          rack->r_ctl.crte,
15039                                                                                                          NULL);
15040                                                 rack_log_hdwr_pacing(rack,
15041                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15042                                                                      err, 2);
15043                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
15044                                         }
15045                                 } else {
15046                                         /* We just need to adjust the segment size */
15047                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15048                                         rack_log_hdwr_pacing(rack,
15049                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15050                                                              0, 4);
15051                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
15052                                 }
15053                         }
15054                 }
15055                 if ((rack->r_ctl.crte != NULL) &&
15056                     (rack->r_ctl.crte->rate == rate_wanted)) {
15057                         /*
15058                          * We need to add a extra if the rates
15059                          * are exactly matched. The idea is
15060                          * we want the software to make sure the
15061                          * queue is empty before adding more, this
15062                          * gives us N MSS extra pace times where
15063                          * N is our sysctl
15064                          */
15065                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15066                 }
15067 done_w_hdwr:
15068                 if (rack_limit_time_with_srtt &&
15069                     (rack->use_fixed_rate == 0) &&
15070 #ifdef NETFLIX_PEAKRATE
15071                     (rack->rc_tp->t_maxpeakrate == 0) &&
15072 #endif
15073                     (rack->rack_hdrw_pacing == 0)) {
15074                         /*
15075                          * Sanity check, we do not allow the pacing delay
15076                          * to be longer than the SRTT of the path. If it is
15077                          * a slow path, then adding a packet should increase
15078                          * the RTT and compensate for this i.e. the srtt will
15079                          * be greater so the allowed pacing time will be greater.
15080                          *
15081                          * Note this restriction is not for where a peak rate
15082                          * is set, we are doing fixed pacing or hardware pacing.
15083                          */
15084                         if (rack->rc_tp->t_srtt)
15085                                 srtt = rack->rc_tp->t_srtt;
15086                         else
15087                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
15088                         if (srtt < (uint64_t)slot) {
15089                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15090                                 slot = srtt;
15091                         }
15092                 }
15093                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15094         }
15095         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15096                 /*
15097                  * If this rate is seeing enobufs when it
15098                  * goes to send then either the nic is out
15099                  * of gas or we are mis-estimating the time
15100                  * somehow and not letting the queue empty
15101                  * completely. Lets add to the pacing time.
15102                  */
15103                 int hw_boost_delay;
15104
15105                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15106                 if (hw_boost_delay > rack_enobuf_hw_max)
15107                         hw_boost_delay = rack_enobuf_hw_max;
15108                 else if (hw_boost_delay < rack_enobuf_hw_min)
15109                         hw_boost_delay = rack_enobuf_hw_min;
15110                 slot += hw_boost_delay;
15111         }
15112         return (slot);
15113 }
15114
15115 static void
15116 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15117     tcp_seq startseq, uint32_t sb_offset)
15118 {
15119         struct rack_sendmap *my_rsm = NULL;
15120         struct rack_sendmap fe;
15121
15122         if (tp->t_state < TCPS_ESTABLISHED) {
15123                 /*
15124                  * We don't start any measurements if we are
15125                  * not at least established.
15126                  */
15127                 return;
15128         }
15129         if (tp->t_state >= TCPS_FIN_WAIT_1) {
15130                 /*
15131                  * We will get no more data into the SB
15132                  * this means we need to have the data available
15133                  * before we start a measurement.
15134                  */
15135
15136                 if (sbavail(&tptosocket(tp)->so_snd) <
15137                     max(rc_init_window(rack),
15138                         (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15139                         /* Nope not enough data */
15140                         return;
15141                 }
15142         }
15143         tp->t_flags |= TF_GPUTINPROG;
15144         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15145         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15146         tp->gput_seq = startseq;
15147         rack->app_limited_needs_set = 0;
15148         if (rack->in_probe_rtt)
15149                 rack->measure_saw_probe_rtt = 1;
15150         else if ((rack->measure_saw_probe_rtt) &&
15151                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15152                 rack->measure_saw_probe_rtt = 0;
15153         if (rack->rc_gp_filled)
15154                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15155         else {
15156                 /* Special case initial measurement */
15157                 struct timeval tv;
15158
15159                 tp->gput_ts = tcp_get_usecs(&tv);
15160                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15161         }
15162         /*
15163          * We take a guess out into the future,
15164          * if we have no measurement and no
15165          * initial rate, we measure the first
15166          * initial-windows worth of data to
15167          * speed up getting some GP measurement and
15168          * thus start pacing.
15169          */
15170         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15171                 rack->app_limited_needs_set = 1;
15172                 tp->gput_ack = startseq + max(rc_init_window(rack),
15173                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15174                 rack_log_pacing_delay_calc(rack,
15175                                            tp->gput_seq,
15176                                            tp->gput_ack,
15177                                            0,
15178                                            tp->gput_ts,
15179                                            rack->r_ctl.rc_app_limited_cnt,
15180                                            9,
15181                                            __LINE__, NULL, 0);
15182                 return;
15183         }
15184         if (sb_offset) {
15185                 /*
15186                  * We are out somewhere in the sb
15187                  * can we use the already outstanding data?
15188                  */
15189                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
15190                         /*
15191                          * Yes first one is good and in this case
15192                          * the tp->gput_ts is correctly set based on
15193                          * the last ack that arrived (no need to
15194                          * set things up when an ack comes in).
15195                          */
15196                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15197                         if ((my_rsm == NULL) ||
15198                             (my_rsm->r_rtr_cnt != 1)) {
15199                                 /* retransmission? */
15200                                 goto use_latest;
15201                         }
15202                 } else {
15203                         if (rack->r_ctl.rc_first_appl == NULL) {
15204                                 /*
15205                                  * If rc_first_appl is NULL
15206                                  * then the cnt should be 0.
15207                                  * This is probably an error, maybe
15208                                  * a KASSERT would be approprate.
15209                                  */
15210                                 goto use_latest;
15211                         }
15212                         /*
15213                          * If we have a marker pointer to the last one that is
15214                          * app limited we can use that, but we need to set
15215                          * things up so that when it gets ack'ed we record
15216                          * the ack time (if its not already acked).
15217                          */
15218                         rack->app_limited_needs_set = 1;
15219                         /*
15220                          * We want to get to the rsm that is either
15221                          * next with space i.e. over 1 MSS or the one
15222                          * after that (after the app-limited).
15223                          */
15224                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15225                                          rack->r_ctl.rc_first_appl);
15226                         if (my_rsm) {
15227                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15228                                         /* Have to use the next one */
15229                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15230                                                          my_rsm);
15231                                 else {
15232                                         /* Use after the first MSS of it is acked */
15233                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15234                                         goto start_set;
15235                                 }
15236                         }
15237                         if ((my_rsm == NULL) ||
15238                             (my_rsm->r_rtr_cnt != 1)) {
15239                                 /*
15240                                  * Either its a retransmit or
15241                                  * the last is the app-limited one.
15242                                  */
15243                                 goto use_latest;
15244                         }
15245                 }
15246                 tp->gput_seq = my_rsm->r_start;
15247 start_set:
15248                 if (my_rsm->r_flags & RACK_ACKED) {
15249                         /*
15250                          * This one has been acked use the arrival ack time
15251                          */
15252                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15253                         rack->app_limited_needs_set = 0;
15254                 }
15255                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15256                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15257                 rack_log_pacing_delay_calc(rack,
15258                                            tp->gput_seq,
15259                                            tp->gput_ack,
15260                                            (uint64_t)my_rsm,
15261                                            tp->gput_ts,
15262                                            rack->r_ctl.rc_app_limited_cnt,
15263                                            9,
15264                                            __LINE__, NULL, 0);
15265                 return;
15266         }
15267
15268 use_latest:
15269         /*
15270          * We don't know how long we may have been
15271          * idle or if this is the first-send. Lets
15272          * setup the flag so we will trim off
15273          * the first ack'd data so we get a true
15274          * measurement.
15275          */
15276         rack->app_limited_needs_set = 1;
15277         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15278         /* Find this guy so we can pull the send time */
15279         fe.r_start = startseq;
15280         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15281         if (my_rsm) {
15282                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15283                 if (my_rsm->r_flags & RACK_ACKED) {
15284                         /*
15285                          * Unlikely since its probably what was
15286                          * just transmitted (but I am paranoid).
15287                          */
15288                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15289                         rack->app_limited_needs_set = 0;
15290                 }
15291                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15292                         /* This also is unlikely */
15293                         tp->gput_seq = my_rsm->r_start;
15294                 }
15295         } else {
15296                 /*
15297                  * TSNH unless we have some send-map limit,
15298                  * and even at that it should not be hitting
15299                  * that limit (we should have stopped sending).
15300                  */
15301                 struct timeval tv;
15302
15303                 microuptime(&tv);
15304                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15305         }
15306         rack_log_pacing_delay_calc(rack,
15307                                    tp->gput_seq,
15308                                    tp->gput_ack,
15309                                    (uint64_t)my_rsm,
15310                                    tp->gput_ts,
15311                                    rack->r_ctl.rc_app_limited_cnt,
15312                                    9, __LINE__, NULL, 0);
15313 }
15314
15315 static inline uint32_t
15316 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15317     uint32_t avail, int32_t sb_offset)
15318 {
15319         uint32_t len;
15320         uint32_t sendwin;
15321
15322         if (tp->snd_wnd > cwnd_to_use)
15323                 sendwin = cwnd_to_use;
15324         else
15325                 sendwin = tp->snd_wnd;
15326         if (ctf_outstanding(tp) >= tp->snd_wnd) {
15327                 /* We never want to go over our peers rcv-window */
15328                 len = 0;
15329         } else {
15330                 uint32_t flight;
15331
15332                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15333                 if (flight >= sendwin) {
15334                         /*
15335                          * We have in flight what we are allowed by cwnd (if
15336                          * it was rwnd blocking it would have hit above out
15337                          * >= tp->snd_wnd).
15338                          */
15339                         return (0);
15340                 }
15341                 len = sendwin - flight;
15342                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15343                         /* We would send too much (beyond the rwnd) */
15344                         len = tp->snd_wnd - ctf_outstanding(tp);
15345                 }
15346                 if ((len + sb_offset) > avail) {
15347                         /*
15348                          * We don't have that much in the SB, how much is
15349                          * there?
15350                          */
15351                         len = avail - sb_offset;
15352                 }
15353         }
15354         return (len);
15355 }
15356
15357 static void
15358 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15359              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15360              int rsm_is_null, int optlen, int line, uint16_t mode)
15361 {
15362         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15363                 union tcp_log_stackspecific log;
15364                 struct timeval tv;
15365
15366                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15367                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15368                 log.u_bbr.flex1 = error;
15369                 log.u_bbr.flex2 = flags;
15370                 log.u_bbr.flex3 = rsm_is_null;
15371                 log.u_bbr.flex4 = ipoptlen;
15372                 log.u_bbr.flex5 = tp->rcv_numsacks;
15373                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15374                 log.u_bbr.flex7 = optlen;
15375                 log.u_bbr.flex8 = rack->r_fsb_inited;
15376                 log.u_bbr.applimited = rack->r_fast_output;
15377                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15378                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15379                 log.u_bbr.cwnd_gain = mode;
15380                 log.u_bbr.pkts_out = orig_len;
15381                 log.u_bbr.lt_epoch = len;
15382                 log.u_bbr.delivered = line;
15383                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15384                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15385                 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15386                                len, &log, false, NULL, NULL, 0, &tv);
15387         }
15388 }
15389
15390
15391 static struct mbuf *
15392 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15393                    struct rack_fast_send_blk *fsb,
15394                    int32_t seglimit, int32_t segsize, int hw_tls)
15395 {
15396 #ifdef KERN_TLS
15397         struct ktls_session *tls, *ntls;
15398 #ifdef INVARIANTS
15399         struct mbuf *start;
15400 #endif
15401 #endif
15402         struct mbuf *m, *n, **np, *smb;
15403         struct mbuf *top;
15404         int32_t off, soff;
15405         int32_t len = *plen;
15406         int32_t fragsize;
15407         int32_t len_cp = 0;
15408         uint32_t mlen, frags;
15409
15410         soff = off = the_off;
15411         smb = m = the_m;
15412         np = &top;
15413         top = NULL;
15414 #ifdef KERN_TLS
15415         if (hw_tls && (m->m_flags & M_EXTPG))
15416                 tls = m->m_epg_tls;
15417         else
15418                 tls = NULL;
15419 #ifdef INVARIANTS
15420         start = m;
15421 #endif
15422 #endif
15423         while (len > 0) {
15424                 if (m == NULL) {
15425                         *plen = len_cp;
15426                         break;
15427                 }
15428 #ifdef KERN_TLS
15429                 if (hw_tls) {
15430                         if (m->m_flags & M_EXTPG)
15431                                 ntls = m->m_epg_tls;
15432                         else
15433                                 ntls = NULL;
15434
15435                         /*
15436                          * Avoid mixing TLS records with handshake
15437                          * data or TLS records from different
15438                          * sessions.
15439                          */
15440                         if (tls != ntls) {
15441                                 MPASS(m != start);
15442                                 *plen = len_cp;
15443                                 break;
15444                         }
15445                 }
15446 #endif
15447                 mlen = min(len, m->m_len - off);
15448                 if (seglimit) {
15449                         /*
15450                          * For M_EXTPG mbufs, add 3 segments
15451                          * + 1 in case we are crossing page boundaries
15452                          * + 2 in case the TLS hdr/trailer are used
15453                          * It is cheaper to just add the segments
15454                          * than it is to take the cache miss to look
15455                          * at the mbuf ext_pgs state in detail.
15456                          */
15457                         if (m->m_flags & M_EXTPG) {
15458                                 fragsize = min(segsize, PAGE_SIZE);
15459                                 frags = 3;
15460                         } else {
15461                                 fragsize = segsize;
15462                                 frags = 0;
15463                         }
15464
15465                         /* Break if we really can't fit anymore. */
15466                         if ((frags + 1) >= seglimit) {
15467                                 *plen = len_cp;
15468                                 break;
15469                         }
15470
15471                         /*
15472                          * Reduce size if you can't copy the whole
15473                          * mbuf. If we can't copy the whole mbuf, also
15474                          * adjust len so the loop will end after this
15475                          * mbuf.
15476                          */
15477                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15478                                 mlen = (seglimit - frags - 1) * fragsize;
15479                                 len = mlen;
15480                                 *plen = len_cp + len;
15481                         }
15482                         frags += howmany(mlen, fragsize);
15483                         if (frags == 0)
15484                                 frags++;
15485                         seglimit -= frags;
15486                         KASSERT(seglimit > 0,
15487                             ("%s: seglimit went too low", __func__));
15488                 }
15489                 n = m_get(M_NOWAIT, m->m_type);
15490                 *np = n;
15491                 if (n == NULL)
15492                         goto nospace;
15493                 n->m_len = mlen;
15494                 soff += mlen;
15495                 len_cp += n->m_len;
15496                 if (m->m_flags & (M_EXT|M_EXTPG)) {
15497                         n->m_data = m->m_data + off;
15498                         mb_dupcl(n, m);
15499                 } else {
15500                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15501                             (u_int)n->m_len);
15502                 }
15503                 len -= n->m_len;
15504                 off = 0;
15505                 m = m->m_next;
15506                 np = &n->m_next;
15507                 if (len || (soff == smb->m_len)) {
15508                         /*
15509                          * We have more so we move forward  or
15510                          * we have consumed the entire mbuf and
15511                          * len has fell to 0.
15512                          */
15513                         soff = 0;
15514                         smb = m;
15515                 }
15516
15517         }
15518         if (fsb != NULL) {
15519                 fsb->m = smb;
15520                 fsb->off = soff;
15521                 if (smb) {
15522                         /*
15523                          * Save off the size of the mbuf. We do
15524                          * this so that we can recognize when it
15525                          * has been trimmed by sbcut() as acks
15526                          * come in.
15527                          */
15528                         fsb->o_m_len = smb->m_len;
15529                 } else {
15530                         /*
15531                          * This is the case where the next mbuf went to NULL. This
15532                          * means with this copy we have sent everything in the sb.
15533                          * In theory we could clear the fast_output flag, but lets
15534                          * not since its possible that we could get more added
15535                          * and acks that call the extend function which would let
15536                          * us send more.
15537                          */
15538                         fsb->o_m_len = 0;
15539                 }
15540         }
15541         return (top);
15542 nospace:
15543         if (top)
15544                 m_freem(top);
15545         return (NULL);
15546
15547 }
15548
15549 /*
15550  * This is a copy of m_copym(), taking the TSO segment size/limit
15551  * constraints into account, and advancing the sndptr as it goes.
15552  */
15553 static struct mbuf *
15554 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15555                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15556 {
15557         struct mbuf *m, *n;
15558         int32_t soff;
15559
15560         soff = rack->r_ctl.fsb.off;
15561         m = rack->r_ctl.fsb.m;
15562         if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15563                 /*
15564                  * The mbuf had the front of it chopped off by an ack
15565                  * we need to adjust the soff/off by that difference.
15566                  */
15567                 uint32_t delta;
15568
15569                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15570                 soff -= delta;
15571         } else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15572                 /*
15573                  * The mbuf was expanded probably by
15574                  * a m_compress. Just update o_m_len.
15575                  */
15576                 rack->r_ctl.fsb.o_m_len = m->m_len;
15577         }
15578         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15579         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15580         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15581                                  __FUNCTION__,
15582                                  rack, *plen, m, m->m_len));
15583         /* Save off the right location before we copy and advance */
15584         *s_soff = soff;
15585         *s_mb = rack->r_ctl.fsb.m;
15586         n = rack_fo_base_copym(m, soff, plen,
15587                                &rack->r_ctl.fsb,
15588                                seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15589         return (n);
15590 }
15591
15592 static int
15593 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15594                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15595 {
15596         /*
15597          * Enter the fast retransmit path. We are given that a sched_pin is
15598          * in place (if accounting is compliled in) and the cycle count taken
15599          * at the entry is in the ts_val. The concept her is that the rsm
15600          * now holds the mbuf offsets and such so we can directly transmit
15601          * without a lot of overhead, the len field is already set for
15602          * us to prohibit us from sending too much (usually its 1MSS).
15603          */
15604         struct ip *ip = NULL;
15605         struct udphdr *udp = NULL;
15606         struct tcphdr *th = NULL;
15607         struct mbuf *m = NULL;
15608         struct inpcb *inp;
15609         uint8_t *cpto;
15610         struct tcp_log_buffer *lgb;
15611 #ifdef TCP_ACCOUNTING
15612         uint64_t crtsc;
15613         int cnt_thru = 1;
15614 #endif
15615         struct tcpopt to;
15616         u_char opt[TCP_MAXOLEN];
15617         uint32_t hdrlen, optlen;
15618         int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
15619         uint16_t flags;
15620         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15621         uint32_t if_hw_tsomaxsegsize;
15622
15623 #ifdef INET6
15624         struct ip6_hdr *ip6 = NULL;
15625
15626         if (rack->r_is_v6) {
15627                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15628                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15629         } else
15630 #endif                          /* INET6 */
15631         {
15632                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15633                 hdrlen = sizeof(struct tcpiphdr);
15634         }
15635         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15636                 goto failed;
15637         }
15638         if (doing_tlp) {
15639                 /* Its a TLP add the flag, it may already be there but be sure */
15640                 rsm->r_flags |= RACK_TLP;
15641         } else {
15642                 /* If it was a TLP it is not not on this retransmit */
15643                 rsm->r_flags &= ~RACK_TLP;
15644         }
15645         startseq = rsm->r_start;
15646         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15647         inp = rack->rc_inp;
15648         to.to_flags = 0;
15649         flags = tcp_outflags[tp->t_state];
15650         if (flags & (TH_SYN|TH_RST)) {
15651                 goto failed;
15652         }
15653         if (rsm->r_flags & RACK_HAS_FIN) {
15654                 /* We can't send a FIN here */
15655                 goto failed;
15656         }
15657         if (flags & TH_FIN) {
15658                 /* We never send a FIN */
15659                 flags &= ~TH_FIN;
15660         }
15661         if (tp->t_flags & TF_RCVD_TSTMP) {
15662                 to.to_tsval = ms_cts + tp->ts_offset;
15663                 to.to_tsecr = tp->ts_recent;
15664                 to.to_flags = TOF_TS;
15665         }
15666         optlen = tcp_addoptions(&to, opt);
15667         hdrlen += optlen;
15668         udp = rack->r_ctl.fsb.udp;
15669         if (udp)
15670                 hdrlen += sizeof(struct udphdr);
15671         if (rack->r_ctl.rc_pace_max_segs)
15672                 max_val = rack->r_ctl.rc_pace_max_segs;
15673         else if (rack->rc_user_set_max_segs)
15674                 max_val = rack->rc_user_set_max_segs * segsiz;
15675         else
15676                 max_val = len;
15677         if ((tp->t_flags & TF_TSO) &&
15678             V_tcp_do_tso &&
15679             (len > segsiz) &&
15680             (tp->t_port == 0))
15681                 tso = 1;
15682 #ifdef INET6
15683         if (MHLEN < hdrlen + max_linkhdr)
15684                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15685         else
15686 #endif
15687                 m = m_gethdr(M_NOWAIT, MT_DATA);
15688         if (m == NULL)
15689                 goto failed;
15690         m->m_data += max_linkhdr;
15691         m->m_len = hdrlen;
15692         th = rack->r_ctl.fsb.th;
15693         /* Establish the len to send */
15694         if (len > max_val)
15695                 len = max_val;
15696         if ((tso) && (len + optlen > tp->t_maxseg)) {
15697                 uint32_t if_hw_tsomax;
15698                 int32_t max_len;
15699
15700                 /* extract TSO information */
15701                 if_hw_tsomax = tp->t_tsomax;
15702                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15703                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15704                 /*
15705                  * Check if we should limit by maximum payload
15706                  * length:
15707                  */
15708                 if (if_hw_tsomax != 0) {
15709                         /* compute maximum TSO length */
15710                         max_len = (if_hw_tsomax - hdrlen -
15711                                    max_linkhdr);
15712                         if (max_len <= 0) {
15713                                 goto failed;
15714                         } else if (len > max_len) {
15715                                 len = max_len;
15716                         }
15717                 }
15718                 if (len <= segsiz) {
15719                         /*
15720                          * In case there are too many small fragments don't
15721                          * use TSO:
15722                          */
15723                         tso = 0;
15724                 }
15725         } else {
15726                 tso = 0;
15727         }
15728         if ((tso == 0) && (len > segsiz))
15729                 len = segsiz;
15730         if ((len == 0) ||
15731             (len <= MHLEN - hdrlen - max_linkhdr)) {
15732                 goto failed;
15733         }
15734         th->th_seq = htonl(rsm->r_start);
15735         th->th_ack = htonl(tp->rcv_nxt);
15736         /*
15737          * The PUSH bit should only be applied
15738          * if the full retransmission is made. If
15739          * we are sending less than this is the
15740          * left hand edge and should not have
15741          * the PUSH bit.
15742          */
15743         if ((rsm->r_flags & RACK_HAD_PUSH) &&
15744             (len == (rsm->r_end - rsm->r_start)))
15745                 flags |= TH_PUSH;
15746         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15747         if (th->th_win == 0) {
15748                 tp->t_sndzerowin++;
15749                 tp->t_flags |= TF_RXWIN0SENT;
15750         } else
15751                 tp->t_flags &= ~TF_RXWIN0SENT;
15752         if (rsm->r_flags & RACK_TLP) {
15753                 /*
15754                  * TLP should not count in retran count, but
15755                  * in its own bin
15756                  */
15757                 counter_u64_add(rack_tlp_retran, 1);
15758                 counter_u64_add(rack_tlp_retran_bytes, len);
15759         } else {
15760                 tp->t_sndrexmitpack++;
15761                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15762                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15763         }
15764 #ifdef STATS
15765         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15766                                  len);
15767 #endif
15768         if (rsm->m == NULL)
15769                 goto failed;
15770         if (rsm->orig_m_len != rsm->m->m_len) {
15771                 /* Fix up the orig_m_len and possibly the mbuf offset */
15772                 rack_adjust_orig_mlen(rsm);
15773         }
15774         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15775         if (len <= segsiz) {
15776                 /*
15777                  * Must have ran out of mbufs for the copy
15778                  * shorten it to no longer need tso. Lets
15779                  * not put on sendalot since we are low on
15780                  * mbufs.
15781                  */
15782                 tso = 0;
15783         }
15784         if ((m->m_next == NULL) || (len <= 0)){
15785                 goto failed;
15786         }
15787         if (udp) {
15788                 if (rack->r_is_v6)
15789                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15790                 else
15791                         ulen = hdrlen + len - sizeof(struct ip);
15792                 udp->uh_ulen = htons(ulen);
15793         }
15794         m->m_pkthdr.rcvif = (struct ifnet *)0;
15795         if (TCPS_HAVERCVDSYN(tp->t_state) &&
15796             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
15797                 int ect = tcp_ecn_output_established(tp, &flags, len, true);
15798                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
15799                     (tp->t_flags2 & TF2_ECN_SND_ECE))
15800                     tp->t_flags2 &= ~TF2_ECN_SND_ECE;
15801 #ifdef INET6
15802                 if (rack->r_is_v6) {
15803                     ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
15804                     ip6->ip6_flow |= htonl(ect << 20);
15805                 }
15806                 else
15807 #endif
15808                 {
15809                     ip->ip_tos &= ~IPTOS_ECN_MASK;
15810                     ip->ip_tos |= ect;
15811                 }
15812         }
15813         tcp_set_flags(th, flags);
15814         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15815 #ifdef INET6
15816         if (rack->r_is_v6) {
15817                 if (tp->t_port) {
15818                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15819                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15820                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15821                         th->th_sum = htons(0);
15822                         UDPSTAT_INC(udps_opackets);
15823                 } else {
15824                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15825                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15826                         th->th_sum = in6_cksum_pseudo(ip6,
15827                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15828                                                       0);
15829                 }
15830         }
15831 #endif
15832 #if defined(INET6) && defined(INET)
15833         else
15834 #endif
15835 #ifdef INET
15836         {
15837                 if (tp->t_port) {
15838                         m->m_pkthdr.csum_flags = CSUM_UDP;
15839                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15840                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15841                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15842                         th->th_sum = htons(0);
15843                         UDPSTAT_INC(udps_opackets);
15844                 } else {
15845                         m->m_pkthdr.csum_flags = CSUM_TCP;
15846                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15847                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15848                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15849                                                                         IPPROTO_TCP + len + optlen));
15850                 }
15851                 /* IP version must be set here for ipv4/ipv6 checking later */
15852                 KASSERT(ip->ip_v == IPVERSION,
15853                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15854         }
15855 #endif
15856         if (tso) {
15857                 KASSERT(len > tp->t_maxseg - optlen,
15858                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15859                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15860                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15861         }
15862 #ifdef INET6
15863         if (rack->r_is_v6) {
15864                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15865                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15866                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15867                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15868                 else
15869                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15870         }
15871 #endif
15872 #if defined(INET) && defined(INET6)
15873         else
15874 #endif
15875 #ifdef INET
15876         {
15877                 ip->ip_len = htons(m->m_pkthdr.len);
15878                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15879                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15880                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15881                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15882                                 ip->ip_off |= htons(IP_DF);
15883                         }
15884                 } else {
15885                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15886                 }
15887         }
15888 #endif
15889         /* Time to copy in our header */
15890         cpto = mtod(m, uint8_t *);
15891         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15892         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15893         if (optlen) {
15894                 bcopy(opt, th + 1, optlen);
15895                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15896         } else {
15897                 th->th_off = sizeof(struct tcphdr) >> 2;
15898         }
15899         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15900                 union tcp_log_stackspecific log;
15901
15902                 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
15903                         rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
15904                         counter_u64_add(rack_collapsed_win_rxt, 1);
15905                         counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
15906                 }
15907                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15908                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15909                 if (rack->rack_no_prr)
15910                         log.u_bbr.flex1 = 0;
15911                 else
15912                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15913                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15914                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15915                 log.u_bbr.flex4 = max_val;
15916                 log.u_bbr.flex5 = 0;
15917                 /* Save off the early/late values */
15918                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15919                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15920                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15921                 if (doing_tlp == 0)
15922                         log.u_bbr.flex8 = 1;
15923                 else
15924                         log.u_bbr.flex8 = 2;
15925                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15926                 log.u_bbr.flex7 = 55;
15927                 log.u_bbr.pkts_out = tp->t_maxseg;
15928                 log.u_bbr.timeStamp = cts;
15929                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15930                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15931                 log.u_bbr.delivered = 0;
15932                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15933                                      len, &log, false, NULL, NULL, 0, tv);
15934         } else
15935                 lgb = NULL;
15936 #ifdef INET6
15937         if (rack->r_is_v6) {
15938                 error = ip6_output(m, NULL,
15939                                    &inp->inp_route6,
15940                                    0, NULL, NULL, inp);
15941         }
15942         else
15943 #endif
15944 #ifdef INET
15945         {
15946                 error = ip_output(m, NULL,
15947                                   &inp->inp_route,
15948                                   0, 0, inp);
15949         }
15950 #endif
15951         m = NULL;
15952         if (lgb) {
15953                 lgb->tlb_errno = error;
15954                 lgb = NULL;
15955         }
15956         if (error) {
15957                 goto failed;
15958         }
15959         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15960                         rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
15961         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15962                 rack->rc_tlp_in_progress = 1;
15963                 rack->r_ctl.rc_tlp_cnt_out++;
15964         }
15965         if (error == 0) {
15966                 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
15967                 if (doing_tlp) {
15968                         rack->rc_last_sent_tlp_past_cumack = 0;
15969                         rack->rc_last_sent_tlp_seq_valid = 1;
15970                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
15971                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
15972                 }
15973         }
15974         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15975         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15976         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15977                 rack->r_ctl.retran_during_recovery += len;
15978         {
15979                 int idx;
15980
15981                 idx = (len / segsiz) + 3;
15982                 if (idx >= TCP_MSS_ACCT_ATIMER)
15983                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15984                 else
15985                         counter_u64_add(rack_out_size[idx], 1);
15986         }
15987         if (tp->t_rtttime == 0) {
15988                 tp->t_rtttime = ticks;
15989                 tp->t_rtseq = startseq;
15990                 KMOD_TCPSTAT_INC(tcps_segstimed);
15991         }
15992         counter_u64_add(rack_fto_rsm_send, 1);
15993         if (error && (error == ENOBUFS)) {
15994                 if (rack->r_ctl.crte != NULL) {
15995                         rack_trace_point(rack, RACK_TP_HWENOBUF);
15996                 } else
15997                         rack_trace_point(rack, RACK_TP_ENOBUF);
15998                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15999                 if (rack->rc_enobuf < 0x7f)
16000                         rack->rc_enobuf++;
16001                 if (slot < (10 * HPTS_USEC_IN_MSEC))
16002                         slot = 10 * HPTS_USEC_IN_MSEC;
16003         } else
16004                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
16005         if ((slot == 0) ||
16006             (rack->rc_always_pace == 0) ||
16007             (rack->r_rr_config == 1)) {
16008                 /*
16009                  * We have no pacing set or we
16010                  * are using old-style rack or
16011                  * we are overridden to use the old 1ms pacing.
16012                  */
16013                 slot = rack->r_ctl.rc_min_to;
16014         }
16015         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
16016 #ifdef TCP_ACCOUNTING
16017         crtsc = get_cyclecount();
16018         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16019                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16020         }
16021         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16022         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16023                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16024         }
16025         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16026         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16027                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
16028         }
16029         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
16030         sched_unpin();
16031 #endif
16032         return (0);
16033 failed:
16034         if (m)
16035                 m_free(m);
16036         return (-1);
16037 }
16038
16039 static void
16040 rack_sndbuf_autoscale(struct tcp_rack *rack)
16041 {
16042         /*
16043          * Automatic sizing of send socket buffer.  Often the send buffer
16044          * size is not optimally adjusted to the actual network conditions
16045          * at hand (delay bandwidth product).  Setting the buffer size too
16046          * small limits throughput on links with high bandwidth and high
16047          * delay (eg. trans-continental/oceanic links).  Setting the
16048          * buffer size too big consumes too much real kernel memory,
16049          * especially with many connections on busy servers.
16050          *
16051          * The criteria to step up the send buffer one notch are:
16052          *  1. receive window of remote host is larger than send buffer
16053          *     (with a fudge factor of 5/4th);
16054          *  2. send buffer is filled to 7/8th with data (so we actually
16055          *     have data to make use of it);
16056          *  3. send buffer fill has not hit maximal automatic size;
16057          *  4. our send window (slow start and cogestion controlled) is
16058          *     larger than sent but unacknowledged data in send buffer.
16059          *
16060          * Note that the rack version moves things much faster since
16061          * we want to avoid hitting cache lines in the rack_fast_output()
16062          * path so this is called much less often and thus moves
16063          * the SB forward by a percentage.
16064          */
16065         struct socket *so;
16066         struct tcpcb *tp;
16067         uint32_t sendwin, scaleup;
16068
16069         tp = rack->rc_tp;
16070         so = rack->rc_inp->inp_socket;
16071         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16072         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16073                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16074                     sbused(&so->so_snd) >=
16075                     (so->so_snd.sb_hiwat / 8 * 7) &&
16076                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16077                     sendwin >= (sbused(&so->so_snd) -
16078                     (tp->snd_nxt - tp->snd_una))) {
16079                         if (rack_autosndbuf_inc)
16080                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16081                         else
16082                                 scaleup = V_tcp_autosndbuf_inc;
16083                         if (scaleup < V_tcp_autosndbuf_inc)
16084                                 scaleup = V_tcp_autosndbuf_inc;
16085                         scaleup += so->so_snd.sb_hiwat;
16086                         if (scaleup > V_tcp_autosndbuf_max)
16087                                 scaleup = V_tcp_autosndbuf_max;
16088                         if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
16089                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16090                 }
16091         }
16092 }
16093
16094 static int
16095 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16096                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16097 {
16098         /*
16099          * Enter to do fast output. We are given that the sched_pin is
16100          * in place (if accounting is compiled in) and the cycle count taken
16101          * at entry is in place in ts_val. The idea here is that
16102          * we know how many more bytes needs to be sent (presumably either
16103          * during pacing or to fill the cwnd and that was greater than
16104          * the max-burst). We have how much to send and all the info we
16105          * need to just send.
16106          */
16107 #ifdef INET
16108         struct ip *ip = NULL;
16109 #endif
16110         struct udphdr *udp = NULL;
16111         struct tcphdr *th = NULL;
16112         struct mbuf *m, *s_mb;
16113         struct inpcb *inp;
16114         uint8_t *cpto;
16115         struct tcp_log_buffer *lgb;
16116 #ifdef TCP_ACCOUNTING
16117         uint64_t crtsc;
16118 #endif
16119         struct tcpopt to;
16120         u_char opt[TCP_MAXOLEN];
16121         uint32_t hdrlen, optlen;
16122 #ifdef TCP_ACCOUNTING
16123         int cnt_thru = 1;
16124 #endif
16125         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
16126         uint16_t flags;
16127         uint32_t s_soff;
16128         uint32_t if_hw_tsomaxsegcount = 0, startseq;
16129         uint32_t if_hw_tsomaxsegsize;
16130         uint16_t add_flag = RACK_SENT_FP;
16131 #ifdef INET6
16132         struct ip6_hdr *ip6 = NULL;
16133
16134         if (rack->r_is_v6) {
16135                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16136                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16137         } else
16138 #endif                          /* INET6 */
16139         {
16140 #ifdef INET
16141                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16142                 hdrlen = sizeof(struct tcpiphdr);
16143 #endif
16144         }
16145         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16146                 m = NULL;
16147                 goto failed;
16148         }
16149         startseq = tp->snd_max;
16150         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16151         inp = rack->rc_inp;
16152         len = rack->r_ctl.fsb.left_to_send;
16153         to.to_flags = 0;
16154         flags = rack->r_ctl.fsb.tcp_flags;
16155         if (tp->t_flags & TF_RCVD_TSTMP) {
16156                 to.to_tsval = ms_cts + tp->ts_offset;
16157                 to.to_tsecr = tp->ts_recent;
16158                 to.to_flags = TOF_TS;
16159         }
16160         optlen = tcp_addoptions(&to, opt);
16161         hdrlen += optlen;
16162         udp = rack->r_ctl.fsb.udp;
16163         if (udp)
16164                 hdrlen += sizeof(struct udphdr);
16165         if (rack->r_ctl.rc_pace_max_segs)
16166                 max_val = rack->r_ctl.rc_pace_max_segs;
16167         else if (rack->rc_user_set_max_segs)
16168                 max_val = rack->rc_user_set_max_segs * segsiz;
16169         else
16170                 max_val = len;
16171         if ((tp->t_flags & TF_TSO) &&
16172             V_tcp_do_tso &&
16173             (len > segsiz) &&
16174             (tp->t_port == 0))
16175                 tso = 1;
16176 again:
16177 #ifdef INET6
16178         if (MHLEN < hdrlen + max_linkhdr)
16179                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16180         else
16181 #endif
16182                 m = m_gethdr(M_NOWAIT, MT_DATA);
16183         if (m == NULL)
16184                 goto failed;
16185         m->m_data += max_linkhdr;
16186         m->m_len = hdrlen;
16187         th = rack->r_ctl.fsb.th;
16188         /* Establish the len to send */
16189         if (len > max_val)
16190                 len = max_val;
16191         if ((tso) && (len + optlen > tp->t_maxseg)) {
16192                 uint32_t if_hw_tsomax;
16193                 int32_t max_len;
16194
16195                 /* extract TSO information */
16196                 if_hw_tsomax = tp->t_tsomax;
16197                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16198                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16199                 /*
16200                  * Check if we should limit by maximum payload
16201                  * length:
16202                  */
16203                 if (if_hw_tsomax != 0) {
16204                         /* compute maximum TSO length */
16205                         max_len = (if_hw_tsomax - hdrlen -
16206                                    max_linkhdr);
16207                         if (max_len <= 0) {
16208                                 goto failed;
16209                         } else if (len > max_len) {
16210                                 len = max_len;
16211                         }
16212                 }
16213                 if (len <= segsiz) {
16214                         /*
16215                          * In case there are too many small fragments don't
16216                          * use TSO:
16217                          */
16218                         tso = 0;
16219                 }
16220         } else {
16221                 tso = 0;
16222         }
16223         if ((tso == 0) && (len > segsiz))
16224                 len = segsiz;
16225         if ((len == 0) ||
16226             (len <= MHLEN - hdrlen - max_linkhdr)) {
16227                 goto failed;
16228         }
16229         sb_offset = tp->snd_max - tp->snd_una;
16230         th->th_seq = htonl(tp->snd_max);
16231         th->th_ack = htonl(tp->rcv_nxt);
16232         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16233         if (th->th_win == 0) {
16234                 tp->t_sndzerowin++;
16235                 tp->t_flags |= TF_RXWIN0SENT;
16236         } else
16237                 tp->t_flags &= ~TF_RXWIN0SENT;
16238         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
16239         KMOD_TCPSTAT_INC(tcps_sndpack);
16240         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16241 #ifdef STATS
16242         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16243                                  len);
16244 #endif
16245         if (rack->r_ctl.fsb.m == NULL)
16246                 goto failed;
16247
16248         /* s_mb and s_soff are saved for rack_log_output */
16249         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16250                                     &s_mb, &s_soff);
16251         if (len <= segsiz) {
16252                 /*
16253                  * Must have ran out of mbufs for the copy
16254                  * shorten it to no longer need tso. Lets
16255                  * not put on sendalot since we are low on
16256                  * mbufs.
16257                  */
16258                 tso = 0;
16259         }
16260         if (rack->r_ctl.fsb.rfo_apply_push &&
16261             (len == rack->r_ctl.fsb.left_to_send)) {
16262                 flags |= TH_PUSH;
16263                 add_flag |= RACK_HAD_PUSH;
16264         }
16265         if ((m->m_next == NULL) || (len <= 0)){
16266                 goto failed;
16267         }
16268         if (udp) {
16269                 if (rack->r_is_v6)
16270                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
16271                 else
16272                         ulen = hdrlen + len - sizeof(struct ip);
16273                 udp->uh_ulen = htons(ulen);
16274         }
16275         m->m_pkthdr.rcvif = (struct ifnet *)0;
16276         if (TCPS_HAVERCVDSYN(tp->t_state) &&
16277             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
16278                 int ect = tcp_ecn_output_established(tp, &flags, len, false);
16279                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
16280                     (tp->t_flags2 & TF2_ECN_SND_ECE))
16281                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
16282 #ifdef INET6
16283                 if (rack->r_is_v6) {
16284                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
16285                         ip6->ip6_flow |= htonl(ect << 20);
16286                 }
16287                 else
16288 #endif
16289                 {
16290 #ifdef INET
16291                         ip->ip_tos &= ~IPTOS_ECN_MASK;
16292                         ip->ip_tos |= ect;
16293 #endif
16294                 }
16295         }
16296         tcp_set_flags(th, flags);
16297         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
16298 #ifdef INET6
16299         if (rack->r_is_v6) {
16300                 if (tp->t_port) {
16301                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16302                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16303                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16304                         th->th_sum = htons(0);
16305                         UDPSTAT_INC(udps_opackets);
16306                 } else {
16307                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16308                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16309                         th->th_sum = in6_cksum_pseudo(ip6,
16310                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16311                                                       0);
16312                 }
16313         }
16314 #endif
16315 #if defined(INET6) && defined(INET)
16316         else
16317 #endif
16318 #ifdef INET
16319         {
16320                 if (tp->t_port) {
16321                         m->m_pkthdr.csum_flags = CSUM_UDP;
16322                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16323                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16324                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16325                         th->th_sum = htons(0);
16326                         UDPSTAT_INC(udps_opackets);
16327                 } else {
16328                         m->m_pkthdr.csum_flags = CSUM_TCP;
16329                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16330                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
16331                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16332                                                                         IPPROTO_TCP + len + optlen));
16333                 }
16334                 /* IP version must be set here for ipv4/ipv6 checking later */
16335                 KASSERT(ip->ip_v == IPVERSION,
16336                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
16337         }
16338 #endif
16339         if (tso) {
16340                 KASSERT(len > tp->t_maxseg - optlen,
16341                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
16342                 m->m_pkthdr.csum_flags |= CSUM_TSO;
16343                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16344         }
16345 #ifdef INET6
16346         if (rack->r_is_v6) {
16347                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16348                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16349                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16350                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16351                 else
16352                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16353         }
16354 #endif
16355 #if defined(INET) && defined(INET6)
16356         else
16357 #endif
16358 #ifdef INET
16359         {
16360                 ip->ip_len = htons(m->m_pkthdr.len);
16361                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16362                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16363                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16364                         if (tp->t_port == 0 || len < V_tcp_minmss) {
16365                                 ip->ip_off |= htons(IP_DF);
16366                         }
16367                 } else {
16368                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16369                 }
16370         }
16371 #endif
16372         /* Time to copy in our header */
16373         cpto = mtod(m, uint8_t *);
16374         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16375         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16376         if (optlen) {
16377                 bcopy(opt, th + 1, optlen);
16378                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16379         } else {
16380                 th->th_off = sizeof(struct tcphdr) >> 2;
16381         }
16382         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16383                 union tcp_log_stackspecific log;
16384
16385                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16386                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16387                 if (rack->rack_no_prr)
16388                         log.u_bbr.flex1 = 0;
16389                 else
16390                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16391                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16392                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16393                 log.u_bbr.flex4 = max_val;
16394                 log.u_bbr.flex5 = 0;
16395                 /* Save off the early/late values */
16396                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16397                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16398                 log.u_bbr.bw_inuse = rack_get_bw(rack);
16399                 log.u_bbr.flex8 = 0;
16400                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16401                 log.u_bbr.flex7 = 44;
16402                 log.u_bbr.pkts_out = tp->t_maxseg;
16403                 log.u_bbr.timeStamp = cts;
16404                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16405                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16406                 log.u_bbr.delivered = 0;
16407                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16408                                      len, &log, false, NULL, NULL, 0, tv);
16409         } else
16410                 lgb = NULL;
16411 #ifdef INET6
16412         if (rack->r_is_v6) {
16413                 error = ip6_output(m, NULL,
16414                                    &inp->inp_route6,
16415                                    0, NULL, NULL, inp);
16416         }
16417 #endif
16418 #if defined(INET) && defined(INET6)
16419         else
16420 #endif
16421 #ifdef INET
16422         {
16423                 error = ip_output(m, NULL,
16424                                   &inp->inp_route,
16425                                   0, 0, inp);
16426         }
16427 #endif
16428         if (lgb) {
16429                 lgb->tlb_errno = error;
16430                 lgb = NULL;
16431         }
16432         if (error) {
16433                 *send_err = error;
16434                 m = NULL;
16435                 goto failed;
16436         }
16437         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16438                         NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16439         m = NULL;
16440         if (tp->snd_una == tp->snd_max) {
16441                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
16442                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16443                 tp->t_acktime = ticks;
16444         }
16445         if (error == 0)
16446                 tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16447
16448         rack->forced_ack = 0;   /* If we send something zap the FA flag */
16449         tot_len += len;
16450         if ((tp->t_flags & TF_GPUTINPROG) == 0)
16451                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16452         tp->snd_max += len;
16453         tp->snd_nxt = tp->snd_max;
16454         {
16455                 int idx;
16456
16457                 idx = (len / segsiz) + 3;
16458                 if (idx >= TCP_MSS_ACCT_ATIMER)
16459                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16460                 else
16461                         counter_u64_add(rack_out_size[idx], 1);
16462         }
16463         if (len <= rack->r_ctl.fsb.left_to_send)
16464                 rack->r_ctl.fsb.left_to_send -= len;
16465         else
16466                 rack->r_ctl.fsb.left_to_send = 0;
16467         if (rack->r_ctl.fsb.left_to_send < segsiz) {
16468                 rack->r_fast_output = 0;
16469                 rack->r_ctl.fsb.left_to_send = 0;
16470                 /* At the end of fast_output scale up the sb */
16471                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16472                 rack_sndbuf_autoscale(rack);
16473                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16474         }
16475         if (tp->t_rtttime == 0) {
16476                 tp->t_rtttime = ticks;
16477                 tp->t_rtseq = startseq;
16478                 KMOD_TCPSTAT_INC(tcps_segstimed);
16479         }
16480         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16481             (max_val > len) &&
16482             (tso == 0)) {
16483                 max_val -= len;
16484                 len = segsiz;
16485                 th = rack->r_ctl.fsb.th;
16486 #ifdef TCP_ACCOUNTING
16487                 cnt_thru++;
16488 #endif
16489                 goto again;
16490         }
16491         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16492         counter_u64_add(rack_fto_send, 1);
16493         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16494         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16495 #ifdef TCP_ACCOUNTING
16496         crtsc = get_cyclecount();
16497         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16498                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16499         }
16500         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16501         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16502                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16503         }
16504         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16505         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16506                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16507         }
16508         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16509         sched_unpin();
16510 #endif
16511         return (0);
16512 failed:
16513         if (m)
16514                 m_free(m);
16515         rack->r_fast_output = 0;
16516         return (-1);
16517 }
16518
16519 static struct rack_sendmap *
16520 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
16521 {
16522         struct rack_sendmap *rsm = NULL;
16523         struct rack_sendmap fe;
16524         int thresh;
16525
16526 restart:
16527         fe.r_start = rack->r_ctl.last_collapse_point;
16528         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
16529         if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
16530                 /* Nothing, strange turn off validity  */
16531                 rack->r_collapse_point_valid = 0;
16532                 return (NULL);
16533         }
16534         /* Can we send it yet? */
16535         if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
16536                 /*
16537                  * Receiver window has not grown enough for
16538                  * the segment to be put on the wire.
16539                  */
16540                 return (NULL);
16541         }
16542         if (rsm->r_flags & RACK_ACKED) {
16543                 /*
16544                  * It has been sacked, lets move to the
16545                  * next one if possible.
16546                  */
16547                 rack->r_ctl.last_collapse_point = rsm->r_end;
16548                 /* Are we done? */
16549                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16550                             rack->r_ctl.high_collapse_point)) {
16551                         rack->r_collapse_point_valid = 0;
16552                         return (NULL);
16553                 }
16554                 goto restart;
16555         }
16556         /* Now has it been long enough ? */
16557         thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
16558         if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
16559                 rack_log_collapse(rack, rsm->r_start,
16560                                   (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16561                                   thresh, __LINE__, 6, rsm->r_flags, rsm);
16562                 return (rsm);
16563         }
16564         /* Not enough time */
16565         rack_log_collapse(rack, rsm->r_start,
16566                           (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16567                           thresh, __LINE__, 7, rsm->r_flags, rsm);
16568         return (NULL);
16569 }
16570
16571 static int
16572 rack_output(struct tcpcb *tp)
16573 {
16574         struct socket *so;
16575         uint32_t recwin;
16576         uint32_t sb_offset, s_moff = 0;
16577         int32_t len, error = 0;
16578         uint16_t flags;
16579         struct mbuf *m, *s_mb = NULL;
16580         struct mbuf *mb;
16581         uint32_t if_hw_tsomaxsegcount = 0;
16582         uint32_t if_hw_tsomaxsegsize;
16583         int32_t segsiz, minseg;
16584         long tot_len_this_send = 0;
16585 #ifdef INET
16586         struct ip *ip = NULL;
16587 #endif
16588         struct udphdr *udp = NULL;
16589         struct tcp_rack *rack;
16590         struct tcphdr *th;
16591         uint8_t pass = 0;
16592         uint8_t mark = 0;
16593         uint8_t wanted_cookie = 0;
16594         u_char opt[TCP_MAXOLEN];
16595         unsigned ipoptlen, optlen, hdrlen, ulen=0;
16596         uint32_t rack_seq;
16597
16598 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16599         unsigned ipsec_optlen = 0;
16600
16601 #endif
16602         int32_t idle, sendalot;
16603         int32_t sub_from_prr = 0;
16604         volatile int32_t sack_rxmit;
16605         struct rack_sendmap *rsm = NULL;
16606         int32_t tso, mtu;
16607         struct tcpopt to;
16608         int32_t slot = 0;
16609         int32_t sup_rack = 0;
16610         uint32_t cts, ms_cts, delayed, early;
16611         uint16_t add_flag = RACK_SENT_SP;
16612         /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16613         uint8_t hpts_calling,  doing_tlp = 0;
16614         uint32_t cwnd_to_use, pace_max_seg;
16615         int32_t do_a_prefetch = 0;
16616         int32_t prefetch_rsm = 0;
16617         int32_t orig_len = 0;
16618         struct timeval tv;
16619         int32_t prefetch_so_done = 0;
16620         struct tcp_log_buffer *lgb;
16621         struct inpcb *inp = tptoinpcb(tp);
16622         struct sockbuf *sb;
16623         uint64_t ts_val = 0;
16624 #ifdef TCP_ACCOUNTING
16625         uint64_t crtsc;
16626 #endif
16627 #ifdef INET6
16628         struct ip6_hdr *ip6 = NULL;
16629         int32_t isipv6;
16630 #endif
16631         bool hw_tls = false;
16632
16633         NET_EPOCH_ASSERT();
16634         INP_WLOCK_ASSERT(inp);
16635
16636         /* setup and take the cache hits here */
16637         rack = (struct tcp_rack *)tp->t_fb_ptr;
16638 #ifdef TCP_ACCOUNTING
16639         sched_pin();
16640         ts_val = get_cyclecount();
16641 #endif
16642         hpts_calling = inp->inp_hpts_calls;
16643 #ifdef TCP_OFFLOAD
16644         if (tp->t_flags & TF_TOE) {
16645 #ifdef TCP_ACCOUNTING
16646                 sched_unpin();
16647 #endif
16648                 return (tcp_offload_output(tp));
16649         }
16650 #endif
16651         /*
16652          * For TFO connections in SYN_RECEIVED, only allow the initial
16653          * SYN|ACK and those sent by the retransmit timer.
16654          */
16655         if (IS_FASTOPEN(tp->t_flags) &&
16656             (tp->t_state == TCPS_SYN_RECEIVED) &&
16657             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16658             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16659 #ifdef TCP_ACCOUNTING
16660                 sched_unpin();
16661 #endif
16662                 return (0);
16663         }
16664 #ifdef INET6
16665         if (rack->r_state) {
16666                 /* Use the cache line loaded if possible */
16667                 isipv6 = rack->r_is_v6;
16668         } else {
16669                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16670         }
16671 #endif
16672         early = 0;
16673         cts = tcp_get_usecs(&tv);
16674         ms_cts = tcp_tv_to_mssectick(&tv);
16675         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16676             tcp_in_hpts(rack->rc_inp)) {
16677                 /*
16678                  * We are on the hpts for some timer but not hptsi output.
16679                  * Remove from the hpts unconditionally.
16680                  */
16681                 rack_timer_cancel(tp, rack, cts, __LINE__);
16682         }
16683         /* Are we pacing and late? */
16684         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16685             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16686                 /* We are delayed */
16687                 delayed = cts - rack->r_ctl.rc_last_output_to;
16688         } else {
16689                 delayed = 0;
16690         }
16691         /* Do the timers, which may override the pacer */
16692         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16693                 int retval;
16694
16695                 retval = rack_process_timers(tp, rack, cts, hpts_calling,
16696                     &doing_tlp);
16697                 if (retval != 0) {
16698                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16699 #ifdef TCP_ACCOUNTING
16700                         sched_unpin();
16701 #endif
16702                         /*
16703                          * If timers want tcp_drop(), then pass error out,
16704                          * otherwise suppress it.
16705                          */
16706                         return (retval < 0 ? retval : 0);
16707                 }
16708         }
16709         if (rack->rc_in_persist) {
16710                 if (tcp_in_hpts(rack->rc_inp) == 0) {
16711                         /* Timer is not running */
16712                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16713                 }
16714 #ifdef TCP_ACCOUNTING
16715                 sched_unpin();
16716 #endif
16717                 return (0);
16718         }
16719         if ((rack->rc_ack_required == 1) &&
16720             (rack->r_timer_override == 0)){
16721                 /* A timeout occurred and no ack has arrived */
16722                 if (tcp_in_hpts(rack->rc_inp) == 0) {
16723                         /* Timer is not running */
16724                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16725                 }
16726 #ifdef TCP_ACCOUNTING
16727                 sched_unpin();
16728 #endif
16729                 return (0);
16730         }
16731         if ((rack->r_timer_override) ||
16732             (rack->rc_ack_can_sendout_data) ||
16733             (delayed) ||
16734             (tp->t_state < TCPS_ESTABLISHED)) {
16735                 rack->rc_ack_can_sendout_data = 0;
16736                 if (tcp_in_hpts(rack->rc_inp))
16737                         tcp_hpts_remove(rack->rc_inp);
16738         } else if (tcp_in_hpts(rack->rc_inp)) {
16739                 /*
16740                  * On the hpts you can't pass even if ACKNOW is on, we will
16741                  * when the hpts fires.
16742                  */
16743 #ifdef TCP_ACCOUNTING
16744                 crtsc = get_cyclecount();
16745                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16746                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16747                 }
16748                 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16749                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16750                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16751                 }
16752                 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16753                 sched_unpin();
16754 #endif
16755                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16756                 return (0);
16757         }
16758         rack->rc_inp->inp_hpts_calls = 0;
16759         /* Finish out both pacing early and late accounting */
16760         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16761             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16762                 early = rack->r_ctl.rc_last_output_to - cts;
16763         } else
16764                 early = 0;
16765         if (delayed) {
16766                 rack->r_ctl.rc_agg_delayed += delayed;
16767                 rack->r_late = 1;
16768         } else if (early) {
16769                 rack->r_ctl.rc_agg_early += early;
16770                 rack->r_early = 1;
16771         }
16772         /* Now that early/late accounting is done turn off the flag */
16773         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16774         rack->r_wanted_output = 0;
16775         rack->r_timer_override = 0;
16776         if ((tp->t_state != rack->r_state) &&
16777             TCPS_HAVEESTABLISHED(tp->t_state)) {
16778                 rack_set_state(tp, rack);
16779         }
16780         if ((rack->r_fast_output) &&
16781             (doing_tlp == 0) &&
16782             (tp->rcv_numsacks == 0)) {
16783                 int ret;
16784
16785                 error = 0;
16786                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16787                 if (ret >= 0)
16788                         return(ret);
16789                 else if (error) {
16790                         inp = rack->rc_inp;
16791                         so = inp->inp_socket;
16792                         sb = &so->so_snd;
16793                         goto nomore;
16794                 }
16795         }
16796         inp = rack->rc_inp;
16797         /*
16798          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16799          * only allow the initial SYN or SYN|ACK and those sent
16800          * by the retransmit timer.
16801          */
16802         if (IS_FASTOPEN(tp->t_flags) &&
16803             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16804              (tp->t_state == TCPS_SYN_SENT)) &&
16805             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16806             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16807                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16808                 so = inp->inp_socket;
16809                 sb = &so->so_snd;
16810                 goto just_return_nolock;
16811         }
16812         /*
16813          * Determine length of data that should be transmitted, and flags
16814          * that will be used. If there is some data or critical controls
16815          * (SYN, RST) to send, then transmit; otherwise, investigate
16816          * further.
16817          */
16818         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16819         if (tp->t_idle_reduce) {
16820                 if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16821                         rack_cc_after_idle(rack, tp);
16822         }
16823         tp->t_flags &= ~TF_LASTIDLE;
16824         if (idle) {
16825                 if (tp->t_flags & TF_MORETOCOME) {
16826                         tp->t_flags |= TF_LASTIDLE;
16827                         idle = 0;
16828                 }
16829         }
16830         if ((tp->snd_una == tp->snd_max) &&
16831             rack->r_ctl.rc_went_idle_time &&
16832             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16833                 idle = cts - rack->r_ctl.rc_went_idle_time;
16834                 if (idle > rack_min_probertt_hold) {
16835                         /* Count as a probe rtt */
16836                         if (rack->in_probe_rtt == 0) {
16837                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16838                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16839                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16840                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16841                         } else {
16842                                 rack_exit_probertt(rack, cts);
16843                         }
16844                 }
16845                 idle = 0;
16846         }
16847         if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16848                 rack_init_fsb_block(tp, rack);
16849 again:
16850         /*
16851          * If we've recently taken a timeout, snd_max will be greater than
16852          * snd_nxt.  There may be SACK information that allows us to avoid
16853          * resending already delivered data.  Adjust snd_nxt accordingly.
16854          */
16855         sendalot = 0;
16856         cts = tcp_get_usecs(&tv);
16857         ms_cts = tcp_tv_to_mssectick(&tv);
16858         tso = 0;
16859         mtu = 0;
16860         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16861         minseg = segsiz;
16862         if (rack->r_ctl.rc_pace_max_segs == 0)
16863                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16864         else
16865                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16866         sb_offset = tp->snd_max - tp->snd_una;
16867         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16868         flags = tcp_outflags[tp->t_state];
16869         while (rack->rc_free_cnt < rack_free_cache) {
16870                 rsm = rack_alloc(rack);
16871                 if (rsm == NULL) {
16872                         if (inp->inp_hpts_calls)
16873                                 /* Retry in a ms */
16874                                 slot = (1 * HPTS_USEC_IN_MSEC);
16875                         so = inp->inp_socket;
16876                         sb = &so->so_snd;
16877                         goto just_return_nolock;
16878                 }
16879                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16880                 rack->rc_free_cnt++;
16881                 rsm = NULL;
16882         }
16883         if (inp->inp_hpts_calls)
16884                 inp->inp_hpts_calls = 0;
16885         sack_rxmit = 0;
16886         len = 0;
16887         rsm = NULL;
16888         if (flags & TH_RST) {
16889                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16890                 so = inp->inp_socket;
16891                 sb = &so->so_snd;
16892                 goto send;
16893         }
16894         if (rack->r_ctl.rc_resend) {
16895                 /* Retransmit timer */
16896                 rsm = rack->r_ctl.rc_resend;
16897                 rack->r_ctl.rc_resend = NULL;
16898                 len = rsm->r_end - rsm->r_start;
16899                 sack_rxmit = 1;
16900                 sendalot = 0;
16901                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16902                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16903                          __func__, __LINE__,
16904                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16905                 sb_offset = rsm->r_start - tp->snd_una;
16906                 if (len >= segsiz)
16907                         len = segsiz;
16908         } else if (rack->r_collapse_point_valid &&
16909                    ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
16910                 /*
16911                  * If an RSM is returned then enough time has passed
16912                  * for us to retransmit it. Move up the collapse point,
16913                  * since this rsm has its chance to retransmit now.
16914                  */
16915                 rack_trace_point(rack, RACK_TP_COLLAPSED_RXT);
16916                 rack->r_ctl.last_collapse_point = rsm->r_end;
16917                 /* Are we done? */
16918                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16919                             rack->r_ctl.high_collapse_point))
16920                         rack->r_collapse_point_valid = 0;
16921                 sack_rxmit = 1;
16922                 /* We are not doing a TLP */
16923                 doing_tlp = 0;
16924                 len = rsm->r_end - rsm->r_start;
16925                 sb_offset = rsm->r_start - tp->snd_una;
16926                 sendalot = 0;
16927                 if ((rack->full_size_rxt == 0) &&
16928                     (rack->shape_rxt_to_pacing_min == 0) &&
16929                     (len >= segsiz))
16930                         len = segsiz;
16931         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16932                 /* We have a retransmit that takes precedence */
16933                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16934                     ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
16935                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16936                         /* Enter recovery if not induced by a time-out */
16937                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
16938                 }
16939 #ifdef INVARIANTS
16940                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16941                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16942                               tp, rack, rsm, rsm->r_start, tp->snd_una);
16943                 }
16944 #endif
16945                 len = rsm->r_end - rsm->r_start;
16946                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16947                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16948                          __func__, __LINE__,
16949                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16950                 sb_offset = rsm->r_start - tp->snd_una;
16951                 sendalot = 0;
16952                 if (len >= segsiz)
16953                         len = segsiz;
16954                 if (len > 0) {
16955                         sack_rxmit = 1;
16956                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16957                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16958                             min(len, segsiz));
16959                 }
16960         } else if (rack->r_ctl.rc_tlpsend) {
16961                 /* Tail loss probe */
16962                 long cwin;
16963                 long tlen;
16964
16965                 /*
16966                  * Check if we can do a TLP with a RACK'd packet
16967                  * this can happen if we are not doing the rack
16968                  * cheat and we skipped to a TLP and it
16969                  * went off.
16970                  */
16971                 rsm = rack->r_ctl.rc_tlpsend;
16972                 /* We are doing a TLP make sure the flag is preent */
16973                 rsm->r_flags |= RACK_TLP;
16974                 rack->r_ctl.rc_tlpsend = NULL;
16975                 sack_rxmit = 1;
16976                 tlen = rsm->r_end - rsm->r_start;
16977                 if (tlen > segsiz)
16978                         tlen = segsiz;
16979                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16980                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16981                          __func__, __LINE__,
16982                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16983                 sb_offset = rsm->r_start - tp->snd_una;
16984                 cwin = min(tp->snd_wnd, tlen);
16985                 len = cwin;
16986         }
16987         if (rack->r_must_retran &&
16988             (doing_tlp == 0) &&
16989             (SEQ_GT(tp->snd_max, tp->snd_una)) &&
16990             (rsm == NULL)) {
16991                 /*
16992                  * There are two different ways that we
16993                  * can get into this block:
16994                  * a) This is a non-sack connection, we had a time-out
16995                  *    and thus r_must_retran was set and everything
16996                  *    left outstanding as been marked for retransmit.
16997                  * b) The MTU of the path shrank, so that everything
16998                  *    was marked to be retransmitted with the smaller
16999                  *    mtu and r_must_retran was set.
17000                  *
17001                  * This means that we expect the sendmap (outstanding)
17002                  * to all be marked must. We can use the tmap to
17003                  * look at them.
17004                  *
17005                  */
17006                 int sendwin, flight;
17007
17008                 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
17009                 flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
17010                 if (flight >= sendwin) {
17011                         /*
17012                          * We can't send yet.
17013                          */
17014                         so = inp->inp_socket;
17015                         sb = &so->so_snd;
17016                         goto just_return_nolock;
17017                 }
17018                 /*
17019                  * This is the case a/b mentioned above. All
17020                  * outstanding/not-acked should be marked.
17021                  * We can use the tmap to find them.
17022                  */
17023                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17024                 if (rsm == NULL) {
17025                         /* TSNH */
17026                         rack->r_must_retran = 0;
17027                         rack->r_ctl.rc_out_at_rto = 0;
17028                         so = inp->inp_socket;
17029                         sb = &so->so_snd;
17030                         goto just_return_nolock;
17031                 }
17032                 if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
17033                         /*
17034                          * The first one does not have the flag, did we collapse
17035                          * further up in our list?
17036                          */
17037                         rack->r_must_retran = 0;
17038                         rack->r_ctl.rc_out_at_rto = 0;
17039                         rsm = NULL;
17040                         sack_rxmit = 0;
17041                 } else {
17042                         sack_rxmit = 1;
17043                         len = rsm->r_end - rsm->r_start;
17044                         sb_offset = rsm->r_start - tp->snd_una;
17045                         sendalot = 0;
17046                         if ((rack->full_size_rxt == 0) &&
17047                             (rack->shape_rxt_to_pacing_min == 0) &&
17048                             (len >= segsiz))
17049                                 len = segsiz;
17050                         /*
17051                          * Delay removing the flag RACK_MUST_RXT so
17052                          * that the fastpath for retransmit will
17053                          * work with this rsm.
17054                          */
17055                 }
17056         }
17057         /*
17058          * Enforce a connection sendmap count limit if set
17059          * as long as we are not retransmiting.
17060          */
17061         if ((rsm == NULL) &&
17062             (rack->do_detection == 0) &&
17063             (V_tcp_map_entries_limit > 0) &&
17064             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
17065                 counter_u64_add(rack_to_alloc_limited, 1);
17066                 if (!rack->alloc_limit_reported) {
17067                         rack->alloc_limit_reported = 1;
17068                         counter_u64_add(rack_alloc_limited_conns, 1);
17069                 }
17070                 so = inp->inp_socket;
17071                 sb = &so->so_snd;
17072                 goto just_return_nolock;
17073         }
17074         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
17075                 /* we are retransmitting the fin */
17076                 len--;
17077                 if (len) {
17078                         /*
17079                          * When retransmitting data do *not* include the
17080                          * FIN. This could happen from a TLP probe.
17081                          */
17082                         flags &= ~TH_FIN;
17083                 }
17084         }
17085         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
17086             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
17087                 int ret;
17088
17089                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
17090                 if (ret == 0)
17091                         return (0);
17092         }
17093         so = inp->inp_socket;
17094         sb = &so->so_snd;
17095         if (do_a_prefetch == 0) {
17096                 kern_prefetch(sb, &do_a_prefetch);
17097                 do_a_prefetch = 1;
17098         }
17099 #ifdef NETFLIX_SHARED_CWND
17100         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
17101             rack->rack_enable_scwnd) {
17102                 /* We are doing cwnd sharing */
17103                 if (rack->gp_ready &&
17104                     (rack->rack_attempted_scwnd == 0) &&
17105                     (rack->r_ctl.rc_scw == NULL) &&
17106                     tp->t_lib) {
17107                         /* The pcbid is in, lets make an attempt */
17108                         counter_u64_add(rack_try_scwnd, 1);
17109                         rack->rack_attempted_scwnd = 1;
17110                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
17111                                                                    &rack->r_ctl.rc_scw_index,
17112                                                                    segsiz);
17113                 }
17114                 if (rack->r_ctl.rc_scw &&
17115                     (rack->rack_scwnd_is_idle == 1) &&
17116                     sbavail(&so->so_snd)) {
17117                         /* we are no longer out of data */
17118                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17119                         rack->rack_scwnd_is_idle = 0;
17120                 }
17121                 if (rack->r_ctl.rc_scw) {
17122                         /* First lets update and get the cwnd */
17123                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
17124                                                                     rack->r_ctl.rc_scw_index,
17125                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
17126                 }
17127         }
17128 #endif
17129         /*
17130          * Get standard flags, and add SYN or FIN if requested by 'hidden'
17131          * state flags.
17132          */
17133         if (tp->t_flags & TF_NEEDFIN)
17134                 flags |= TH_FIN;
17135         if (tp->t_flags & TF_NEEDSYN)
17136                 flags |= TH_SYN;
17137         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17138                 void *end_rsm;
17139                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17140                 if (end_rsm)
17141                         kern_prefetch(end_rsm, &prefetch_rsm);
17142                 prefetch_rsm = 1;
17143         }
17144         SOCKBUF_LOCK(sb);
17145         /*
17146          * If snd_nxt == snd_max and we have transmitted a FIN, the
17147          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17148          * negative length.  This can also occur when TCP opens up its
17149          * congestion window while receiving additional duplicate acks after
17150          * fast-retransmit because TCP will reset snd_nxt to snd_max after
17151          * the fast-retransmit.
17152          *
17153          * In the normal retransmit-FIN-only case, however, snd_nxt will be
17154          * set to snd_una, the sb_offset will be 0, and the length may wind
17155          * up 0.
17156          *
17157          * If sack_rxmit is true we are retransmitting from the scoreboard
17158          * in which case len is already set.
17159          */
17160         if ((sack_rxmit == 0) &&
17161             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17162                 uint32_t avail;
17163
17164                 avail = sbavail(sb);
17165                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17166                         sb_offset = tp->snd_nxt - tp->snd_una;
17167                 else
17168                         sb_offset = 0;
17169                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17170                         if (rack->r_ctl.rc_tlp_new_data) {
17171                                 /* TLP is forcing out new data */
17172                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17173                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17174                                 }
17175                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17176                                         if (tp->snd_wnd > sb_offset)
17177                                                 len = tp->snd_wnd - sb_offset;
17178                                         else
17179                                                 len = 0;
17180                                 } else {
17181                                         len = rack->r_ctl.rc_tlp_new_data;
17182                                 }
17183                                 rack->r_ctl.rc_tlp_new_data = 0;
17184                         }  else {
17185                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17186                         }
17187                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17188                                 /*
17189                                  * For prr=off, we need to send only 1 MSS
17190                                  * at a time. We do this because another sack could
17191                                  * be arriving that causes us to send retransmits and
17192                                  * we don't want to be on a long pace due to a larger send
17193                                  * that keeps us from sending out the retransmit.
17194                                  */
17195                                 len = segsiz;
17196                         }
17197                 } else {
17198                         uint32_t outstanding;
17199                         /*
17200                          * We are inside of a Fast recovery episode, this
17201                          * is caused by a SACK or 3 dup acks. At this point
17202                          * we have sent all the retransmissions and we rely
17203                          * on PRR to dictate what we will send in the form of
17204                          * new data.
17205                          */
17206
17207                         outstanding = tp->snd_max - tp->snd_una;
17208                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17209                                 if (tp->snd_wnd > outstanding) {
17210                                         len = tp->snd_wnd - outstanding;
17211                                         /* Check to see if we have the data */
17212                                         if ((sb_offset + len) > avail) {
17213                                                 /* It does not all fit */
17214                                                 if (avail > sb_offset)
17215                                                         len = avail - sb_offset;
17216                                                 else
17217                                                         len = 0;
17218                                         }
17219                                 } else {
17220                                         len = 0;
17221                                 }
17222                         } else if (avail > sb_offset) {
17223                                 len = avail - sb_offset;
17224                         } else {
17225                                 len = 0;
17226                         }
17227                         if (len > 0) {
17228                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
17229                                         len = rack->r_ctl.rc_prr_sndcnt;
17230                                 }
17231                                 if (len > 0) {
17232                                         sub_from_prr = 1;
17233                                 }
17234                         }
17235                         if (len > segsiz) {
17236                                 /*
17237                                  * We should never send more than a MSS when
17238                                  * retransmitting or sending new data in prr
17239                                  * mode unless the override flag is on. Most
17240                                  * likely the PRR algorithm is not going to
17241                                  * let us send a lot as well :-)
17242                                  */
17243                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
17244                                         len = segsiz;
17245                                 }
17246                         } else if (len < segsiz) {
17247                                 /*
17248                                  * Do we send any? The idea here is if the
17249                                  * send empty's the socket buffer we want to
17250                                  * do it. However if not then lets just wait
17251                                  * for our prr_sndcnt to get bigger.
17252                                  */
17253                                 long leftinsb;
17254
17255                                 leftinsb = sbavail(sb) - sb_offset;
17256                                 if (leftinsb > len) {
17257                                         /* This send does not empty the sb */
17258                                         len = 0;
17259                                 }
17260                         }
17261                 }
17262         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17263                 /*
17264                  * If you have not established
17265                  * and are not doing FAST OPEN
17266                  * no data please.
17267                  */
17268                 if ((sack_rxmit == 0) &&
17269                     (!IS_FASTOPEN(tp->t_flags))){
17270                         len = 0;
17271                         sb_offset = 0;
17272                 }
17273         }
17274         if (prefetch_so_done == 0) {
17275                 kern_prefetch(so, &prefetch_so_done);
17276                 prefetch_so_done = 1;
17277         }
17278         /*
17279          * Lop off SYN bit if it has already been sent.  However, if this is
17280          * SYN-SENT state and if segment contains data and if we don't know
17281          * that foreign host supports TAO, suppress sending segment.
17282          */
17283         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17284             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17285                 /*
17286                  * When sending additional segments following a TFO SYN|ACK,
17287                  * do not include the SYN bit.
17288                  */
17289                 if (IS_FASTOPEN(tp->t_flags) &&
17290                     (tp->t_state == TCPS_SYN_RECEIVED))
17291                         flags &= ~TH_SYN;
17292         }
17293         /*
17294          * Be careful not to send data and/or FIN on SYN segments. This
17295          * measure is needed to prevent interoperability problems with not
17296          * fully conformant TCP implementations.
17297          */
17298         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17299                 len = 0;
17300                 flags &= ~TH_FIN;
17301         }
17302         /*
17303          * On TFO sockets, ensure no data is sent in the following cases:
17304          *
17305          *  - When retransmitting SYN|ACK on a passively-created socket
17306          *
17307          *  - When retransmitting SYN on an actively created socket
17308          *
17309          *  - When sending a zero-length cookie (cookie request) on an
17310          *    actively created socket
17311          *
17312          *  - When the socket is in the CLOSED state (RST is being sent)
17313          */
17314         if (IS_FASTOPEN(tp->t_flags) &&
17315             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17316              ((tp->t_state == TCPS_SYN_SENT) &&
17317               (tp->t_tfo_client_cookie_len == 0)) ||
17318              (flags & TH_RST))) {
17319                 sack_rxmit = 0;
17320                 len = 0;
17321         }
17322         /* Without fast-open there should never be data sent on a SYN */
17323         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17324                 tp->snd_nxt = tp->iss;
17325                 len = 0;
17326         }
17327         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17328                 /* We only send 1 MSS if we have a DSACK block */
17329                 add_flag |= RACK_SENT_W_DSACK;
17330                 len = segsiz;
17331         }
17332         orig_len = len;
17333         if (len <= 0) {
17334                 /*
17335                  * If FIN has been sent but not acked, but we haven't been
17336                  * called to retransmit, len will be < 0.  Otherwise, window
17337                  * shrank after we sent into it.  If window shrank to 0,
17338                  * cancel pending retransmit, pull snd_nxt back to (closed)
17339                  * window, and set the persist timer if it isn't already
17340                  * going.  If the window didn't close completely, just wait
17341                  * for an ACK.
17342                  *
17343                  * We also do a general check here to ensure that we will
17344                  * set the persist timer when we have data to send, but a
17345                  * 0-byte window. This makes sure the persist timer is set
17346                  * even if the packet hits one of the "goto send" lines
17347                  * below.
17348                  */
17349                 len = 0;
17350                 if ((tp->snd_wnd == 0) &&
17351                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17352                     (tp->snd_una == tp->snd_max) &&
17353                     (sb_offset < (int)sbavail(sb))) {
17354                         rack_enter_persist(tp, rack, cts);
17355                 }
17356         } else if ((rsm == NULL) &&
17357                    (doing_tlp == 0) &&
17358                    (len < pace_max_seg)) {
17359                 /*
17360                  * We are not sending a maximum sized segment for
17361                  * some reason. Should we not send anything (think
17362                  * sws or persists)?
17363                  */
17364                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17365                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17366                     (len < minseg) &&
17367                     (len < (int)(sbavail(sb) - sb_offset))) {
17368                         /*
17369                          * Here the rwnd is less than
17370                          * the minimum pacing size, this is not a retransmit,
17371                          * we are established and
17372                          * the send is not the last in the socket buffer
17373                          * we send nothing, and we may enter persists
17374                          * if nothing is outstanding.
17375                          */
17376                         len = 0;
17377                         if (tp->snd_max == tp->snd_una) {
17378                                 /*
17379                                  * Nothing out we can
17380                                  * go into persists.
17381                                  */
17382                                 rack_enter_persist(tp, rack, cts);
17383                         }
17384                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17385                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17386                            (len < (int)(sbavail(sb) - sb_offset)) &&
17387                            (len < minseg)) {
17388                         /*
17389                          * Here we are not retransmitting, and
17390                          * the cwnd is not so small that we could
17391                          * not send at least a min size (rxt timer
17392                          * not having gone off), We have 2 segments or
17393                          * more already in flight, its not the tail end
17394                          * of the socket buffer  and the cwnd is blocking
17395                          * us from sending out a minimum pacing segment size.
17396                          * Lets not send anything.
17397                          */
17398                         len = 0;
17399                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17400                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17401                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17402                            (len < (int)(sbavail(sb) - sb_offset)) &&
17403                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
17404                         /*
17405                          * Here we have a send window but we have
17406                          * filled it up and we can't send another pacing segment.
17407                          * We also have in flight more than 2 segments
17408                          * and we are not completing the sb i.e. we allow
17409                          * the last bytes of the sb to go out even if
17410                          * its not a full pacing segment.
17411                          */
17412                         len = 0;
17413                 } else if ((rack->r_ctl.crte != NULL) &&
17414                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17415                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17416                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17417                            (len < (int)(sbavail(sb) - sb_offset))) {
17418                         /*
17419                          * Here we are doing hardware pacing, this is not a TLP,
17420                          * we are not sending a pace max segment size, there is rwnd
17421                          * room to send at least N pace_max_seg, the cwnd is greater
17422                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
17423                          * more segments in flight and its not the tail of the socket buffer.
17424                          *
17425                          * We don't want to send instead we need to get more ack's in to
17426                          * allow us to send a full pacing segment. Normally, if we are pacing
17427                          * about the right speed, we should have finished our pacing
17428                          * send as most of the acks have come back if we are at the
17429                          * right rate. This is a bit fuzzy since return path delay
17430                          * can delay the acks, which is why we want to make sure we
17431                          * have cwnd space to have a bit more than a max pace segments in flight.
17432                          *
17433                          * If we have not gotten our acks back we are pacing at too high a
17434                          * rate delaying will not hurt and will bring our GP estimate down by
17435                          * injecting the delay. If we don't do this we will send
17436                          * 2 MSS out in response to the acks being clocked in which
17437                          * defeats the point of hw-pacing (i.e. to help us get
17438                          * larger TSO's out).
17439                          */
17440                         len = 0;
17441
17442                 }
17443
17444         }
17445         /* len will be >= 0 after this point. */
17446         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17447         rack_sndbuf_autoscale(rack);
17448         /*
17449          * Decide if we can use TCP Segmentation Offloading (if supported by
17450          * hardware).
17451          *
17452          * TSO may only be used if we are in a pure bulk sending state.  The
17453          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17454          * options prevent using TSO.  With TSO the TCP header is the same
17455          * (except for the sequence number) for all generated packets.  This
17456          * makes it impossible to transmit any options which vary per
17457          * generated segment or packet.
17458          *
17459          * IPv4 handling has a clear separation of ip options and ip header
17460          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17461          * the right thing below to provide length of just ip options and thus
17462          * checking for ipoptlen is enough to decide if ip options are present.
17463          */
17464         ipoptlen = 0;
17465 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17466         /*
17467          * Pre-calculate here as we save another lookup into the darknesses
17468          * of IPsec that way and can actually decide if TSO is ok.
17469          */
17470 #ifdef INET6
17471         if (isipv6 && IPSEC_ENABLED(ipv6))
17472                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
17473 #ifdef INET
17474         else
17475 #endif
17476 #endif                          /* INET6 */
17477 #ifdef INET
17478                 if (IPSEC_ENABLED(ipv4))
17479                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
17480 #endif                          /* INET */
17481 #endif
17482
17483 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17484         ipoptlen += ipsec_optlen;
17485 #endif
17486         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17487             (tp->t_port == 0) &&
17488             ((tp->t_flags & TF_SIGNATURE) == 0) &&
17489             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17490             ipoptlen == 0)
17491                 tso = 1;
17492         {
17493                 uint32_t outstanding __unused;
17494
17495                 outstanding = tp->snd_max - tp->snd_una;
17496                 if (tp->t_flags & TF_SENTFIN) {
17497                         /*
17498                          * If we sent a fin, snd_max is 1 higher than
17499                          * snd_una
17500                          */
17501                         outstanding--;
17502                 }
17503                 if (sack_rxmit) {
17504                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17505                                 flags &= ~TH_FIN;
17506                 } else {
17507                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17508                                    sbused(sb)))
17509                                 flags &= ~TH_FIN;
17510                 }
17511         }
17512         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17513             (long)TCP_MAXWIN << tp->rcv_scale);
17514
17515         /*
17516          * Sender silly window avoidance.   We transmit under the following
17517          * conditions when len is non-zero:
17518          *
17519          * - We have a full segment (or more with TSO) - This is the last
17520          * buffer in a write()/send() and we are either idle or running
17521          * NODELAY - we've timed out (e.g. persist timer) - we have more
17522          * then 1/2 the maximum send window's worth of data (receiver may be
17523          * limited the window size) - we need to retransmit
17524          */
17525         if (len) {
17526                 if (len >= segsiz) {
17527                         goto send;
17528                 }
17529                 /*
17530                  * NOTE! on localhost connections an 'ack' from the remote
17531                  * end may occur synchronously with the output and cause us
17532                  * to flush a buffer queued with moretocome.  XXX
17533                  *
17534                  */
17535                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
17536                     (idle || (tp->t_flags & TF_NODELAY)) &&
17537                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17538                     (tp->t_flags & TF_NOPUSH) == 0) {
17539                         pass = 2;
17540                         goto send;
17541                 }
17542                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
17543                         pass = 22;
17544                         goto send;
17545                 }
17546                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17547                         pass = 4;
17548                         goto send;
17549                 }
17550                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
17551                         pass = 5;
17552                         goto send;
17553                 }
17554                 if (sack_rxmit) {
17555                         pass = 6;
17556                         goto send;
17557                 }
17558                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17559                     (ctf_outstanding(tp) < (segsiz * 2))) {
17560                         /*
17561                          * We have less than two MSS outstanding (delayed ack)
17562                          * and our rwnd will not let us send a full sized
17563                          * MSS. Lets go ahead and let this small segment
17564                          * out because we want to try to have at least two
17565                          * packets inflight to not be caught by delayed ack.
17566                          */
17567                         pass = 12;
17568                         goto send;
17569                 }
17570         }
17571         /*
17572          * Sending of standalone window updates.
17573          *
17574          * Window updates are important when we close our window due to a
17575          * full socket buffer and are opening it again after the application
17576          * reads data from it.  Once the window has opened again and the
17577          * remote end starts to send again the ACK clock takes over and
17578          * provides the most current window information.
17579          *
17580          * We must avoid the silly window syndrome whereas every read from
17581          * the receive buffer, no matter how small, causes a window update
17582          * to be sent.  We also should avoid sending a flurry of window
17583          * updates when the socket buffer had queued a lot of data and the
17584          * application is doing small reads.
17585          *
17586          * Prevent a flurry of pointless window updates by only sending an
17587          * update when we can increase the advertized window by more than
17588          * 1/4th of the socket buffer capacity.  When the buffer is getting
17589          * full or is very small be more aggressive and send an update
17590          * whenever we can increase by two mss sized segments. In all other
17591          * situations the ACK's to new incoming data will carry further
17592          * window increases.
17593          *
17594          * Don't send an independent window update if a delayed ACK is
17595          * pending (it will get piggy-backed on it) or the remote side
17596          * already has done a half-close and won't send more data.  Skip
17597          * this if the connection is in T/TCP half-open state.
17598          */
17599         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17600             !(tp->t_flags & TF_DELACK) &&
17601             !TCPS_HAVERCVDFIN(tp->t_state)) {
17602                 /*
17603                  * "adv" is the amount we could increase the window, taking
17604                  * into account that we are limited by TCP_MAXWIN <<
17605                  * tp->rcv_scale.
17606                  */
17607                 int32_t adv;
17608                 int oldwin;
17609
17610                 adv = recwin;
17611                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17612                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
17613                         if (adv > oldwin)
17614                             adv -= oldwin;
17615                         else {
17616                                 /* We can't increase the window */
17617                                 adv = 0;
17618                         }
17619                 } else
17620                         oldwin = 0;
17621
17622                 /*
17623                  * If the new window size ends up being the same as or less
17624                  * than the old size when it is scaled, then don't force
17625                  * a window update.
17626                  */
17627                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17628                         goto dontupdate;
17629
17630                 if (adv >= (int32_t)(2 * segsiz) &&
17631                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17632                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17633                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17634                         pass = 7;
17635                         goto send;
17636                 }
17637                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17638                         pass = 23;
17639                         goto send;
17640                 }
17641         }
17642 dontupdate:
17643
17644         /*
17645          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17646          * is also a catch-all for the retransmit timer timeout case.
17647          */
17648         if (tp->t_flags & TF_ACKNOW) {
17649                 pass = 8;
17650                 goto send;
17651         }
17652         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17653                 pass = 9;
17654                 goto send;
17655         }
17656         /*
17657          * If our state indicates that FIN should be sent and we have not
17658          * yet done so, then we need to send.
17659          */
17660         if ((flags & TH_FIN) &&
17661             (tp->snd_nxt == tp->snd_una)) {
17662                 pass = 11;
17663                 goto send;
17664         }
17665         /*
17666          * No reason to send a segment, just return.
17667          */
17668 just_return:
17669         SOCKBUF_UNLOCK(sb);
17670 just_return_nolock:
17671         {
17672                 int app_limited = CTF_JR_SENT_DATA;
17673
17674                 if (tot_len_this_send > 0) {
17675                         /* Make sure snd_nxt is up to max */
17676                         rack->r_ctl.fsb.recwin = recwin;
17677                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17678                         if ((error == 0) &&
17679                             rack_use_rfo &&
17680                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
17681                             (ipoptlen == 0) &&
17682                             (tp->snd_nxt == tp->snd_max) &&
17683                             (tp->rcv_numsacks == 0) &&
17684                             rack->r_fsb_inited &&
17685                             TCPS_HAVEESTABLISHED(tp->t_state) &&
17686                             (rack->r_must_retran == 0) &&
17687                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
17688                             (len > 0) && (orig_len > 0) &&
17689                             (orig_len > len) &&
17690                             ((orig_len - len) >= segsiz) &&
17691                             ((optlen == 0) ||
17692                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17693                                 /* We can send at least one more MSS using our fsb */
17694
17695                                 rack->r_fast_output = 1;
17696                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17697                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17698                                 rack->r_ctl.fsb.tcp_flags = flags;
17699                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
17700                                 if (hw_tls)
17701                                         rack->r_ctl.fsb.hw_tls = 1;
17702                                 else
17703                                         rack->r_ctl.fsb.hw_tls = 0;
17704                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17705                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
17706                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17707                                          (tp->snd_max - tp->snd_una)));
17708                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
17709                                         rack->r_fast_output = 0;
17710                                 else {
17711                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17712                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
17713                                         else
17714                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
17715                                 }
17716                         } else
17717                                 rack->r_fast_output = 0;
17718
17719
17720                         rack_log_fsb(rack, tp, so, flags,
17721                                      ipoptlen, orig_len, len, 0,
17722                                      1, optlen, __LINE__, 1);
17723                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17724                                 tp->snd_nxt = tp->snd_max;
17725                 } else {
17726                         int end_window = 0;
17727                         uint32_t seq = tp->gput_ack;
17728
17729                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17730                         if (rsm) {
17731                                 /*
17732                                  * Mark the last sent that we just-returned (hinting
17733                                  * that delayed ack may play a role in any rtt measurement).
17734                                  */
17735                                 rsm->r_just_ret = 1;
17736                         }
17737                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17738                         rack->r_ctl.rc_agg_delayed = 0;
17739                         rack->r_early = 0;
17740                         rack->r_late = 0;
17741                         rack->r_ctl.rc_agg_early = 0;
17742                         if ((ctf_outstanding(tp) +
17743                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17744                                  minseg)) >= tp->snd_wnd) {
17745                                 /* We are limited by the rwnd */
17746                                 app_limited = CTF_JR_RWND_LIMITED;
17747                                 if (IN_FASTRECOVERY(tp->t_flags))
17748                                     rack->r_ctl.rc_prr_sndcnt = 0;
17749                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17750                                 /* We are limited by whats available -- app limited */
17751                                 app_limited = CTF_JR_APP_LIMITED;
17752                                 if (IN_FASTRECOVERY(tp->t_flags))
17753                                     rack->r_ctl.rc_prr_sndcnt = 0;
17754                         } else if ((idle == 0) &&
17755                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17756                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17757                                    (len < segsiz)) {
17758                                 /*
17759                                  * No delay is not on and the
17760                                  * user is sending less than 1MSS. This
17761                                  * brings out SWS avoidance so we
17762                                  * don't send. Another app-limited case.
17763                                  */
17764                                 app_limited = CTF_JR_APP_LIMITED;
17765                         } else if (tp->t_flags & TF_NOPUSH) {
17766                                 /*
17767                                  * The user has requested no push of
17768                                  * the last segment and we are
17769                                  * at the last segment. Another app
17770                                  * limited case.
17771                                  */
17772                                 app_limited = CTF_JR_APP_LIMITED;
17773                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17774                                 /* Its the cwnd */
17775                                 app_limited = CTF_JR_CWND_LIMITED;
17776                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17777                                    (rack->rack_no_prr == 0) &&
17778                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17779                                 app_limited = CTF_JR_PRR;
17780                         } else {
17781                                 /* Now why here are we not sending? */
17782 #ifdef NOW
17783 #ifdef INVARIANTS
17784                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17785 #endif
17786 #endif
17787                                 app_limited = CTF_JR_ASSESSING;
17788                         }
17789                         /*
17790                          * App limited in some fashion, for our pacing GP
17791                          * measurements we don't want any gap (even cwnd).
17792                          * Close  down the measurement window.
17793                          */
17794                         if (rack_cwnd_block_ends_measure &&
17795                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17796                              (app_limited == CTF_JR_PRR))) {
17797                                 /*
17798                                  * The reason we are not sending is
17799                                  * the cwnd (or prr). We have been configured
17800                                  * to end the measurement window in
17801                                  * this case.
17802                                  */
17803                                 end_window = 1;
17804                         } else if (rack_rwnd_block_ends_measure &&
17805                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17806                                 /*
17807                                  * We are rwnd limited and have been
17808                                  * configured to end the measurement
17809                                  * window in this case.
17810                                  */
17811                                 end_window = 1;
17812                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17813                                 /*
17814                                  * A true application limited period, we have
17815                                  * ran out of data.
17816                                  */
17817                                 end_window = 1;
17818                         } else if (app_limited == CTF_JR_ASSESSING) {
17819                                 /*
17820                                  * In the assessing case we hit the end of
17821                                  * the if/else and had no known reason
17822                                  * This will panic us under invariants..
17823                                  *
17824                                  * If we get this out in logs we need to
17825                                  * investagate which reason we missed.
17826                                  */
17827                                 end_window = 1;
17828                         }
17829                         if (end_window) {
17830                                 uint8_t log = 0;
17831
17832                                 /* Adjust the Gput measurement */
17833                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17834                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17835                                         tp->gput_ack = tp->snd_max;
17836                                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17837                                                 /*
17838                                                  * There is not enough to measure.
17839                                                  */
17840                                                 tp->t_flags &= ~TF_GPUTINPROG;
17841                                                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17842                                                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
17843                                                                            tp->gput_seq,
17844                                                                            0, 0, 18, __LINE__, NULL, 0);
17845                                         } else
17846                                                 log = 1;
17847                                 }
17848                                 /* Mark the last packet has app limited */
17849                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17850                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17851                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17852                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17853                                         else {
17854                                                 /*
17855                                                  * Go out to the end app limited and mark
17856                                                  * this new one as next and move the end_appl up
17857                                                  * to this guy.
17858                                                  */
17859                                                 if (rack->r_ctl.rc_end_appl)
17860                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17861                                                 rack->r_ctl.rc_end_appl = rsm;
17862                                         }
17863                                         rsm->r_flags |= RACK_APP_LIMITED;
17864                                         rack->r_ctl.rc_app_limited_cnt++;
17865                                 }
17866                                 if (log)
17867                                         rack_log_pacing_delay_calc(rack,
17868                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17869                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17870                         }
17871                 }
17872                 /* Check if we need to go into persists or not */
17873                 if ((tp->snd_max == tp->snd_una) &&
17874                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17875                     sbavail(sb) &&
17876                     (sbavail(sb) > tp->snd_wnd) &&
17877                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17878                         /* Yes lets make sure to move to persist before timer-start */
17879                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17880                 }
17881                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17882                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17883         }
17884 #ifdef NETFLIX_SHARED_CWND
17885         if ((sbavail(sb) == 0) &&
17886             rack->r_ctl.rc_scw) {
17887                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17888                 rack->rack_scwnd_is_idle = 1;
17889         }
17890 #endif
17891 #ifdef TCP_ACCOUNTING
17892         if (tot_len_this_send > 0) {
17893                 crtsc = get_cyclecount();
17894                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17895                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
17896                 }
17897                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17898                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17899                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17900                 }
17901                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17902                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17903                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17904                 }
17905                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17906         } else {
17907                 crtsc = get_cyclecount();
17908                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17909                         tp->tcp_cnt_counters[SND_LIMITED]++;
17910                 }
17911                 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17912                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17913                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17914                 }
17915                 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17916         }
17917         sched_unpin();
17918 #endif
17919         return (0);
17920
17921 send:
17922         if (rsm || sack_rxmit)
17923                 counter_u64_add(rack_nfto_resend, 1);
17924         else
17925                 counter_u64_add(rack_non_fto_send, 1);
17926         if ((flags & TH_FIN) &&
17927             sbavail(sb)) {
17928                 /*
17929                  * We do not transmit a FIN
17930                  * with data outstanding. We
17931                  * need to make it so all data
17932                  * is acked first.
17933                  */
17934                 flags &= ~TH_FIN;
17935         }
17936         /* Enforce stack imposed max seg size if we have one */
17937         if (rack->r_ctl.rc_pace_max_segs &&
17938             (len > rack->r_ctl.rc_pace_max_segs)) {
17939                 mark = 1;
17940                 len = rack->r_ctl.rc_pace_max_segs;
17941         }
17942         SOCKBUF_LOCK_ASSERT(sb);
17943         if (len > 0) {
17944                 if (len >= segsiz)
17945                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17946                 else
17947                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17948         }
17949         /*
17950          * Before ESTABLISHED, force sending of initial options unless TCP
17951          * set not to do any options. NOTE: we assume that the IP/TCP header
17952          * plus TCP options always fit in a single mbuf, leaving room for a
17953          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17954          * + optlen <= MCLBYTES
17955          */
17956         optlen = 0;
17957 #ifdef INET6
17958         if (isipv6)
17959                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17960         else
17961 #endif
17962                 hdrlen = sizeof(struct tcpiphdr);
17963
17964         /*
17965          * Compute options for segment. We only have to care about SYN and
17966          * established connection segments.  Options for SYN-ACK segments
17967          * are handled in TCP syncache.
17968          */
17969         to.to_flags = 0;
17970         if ((tp->t_flags & TF_NOOPT) == 0) {
17971                 /* Maximum segment size. */
17972                 if (flags & TH_SYN) {
17973                         tp->snd_nxt = tp->iss;
17974                         to.to_mss = tcp_mssopt(&inp->inp_inc);
17975                         if (tp->t_port)
17976                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
17977                         to.to_flags |= TOF_MSS;
17978
17979                         /*
17980                          * On SYN or SYN|ACK transmits on TFO connections,
17981                          * only include the TFO option if it is not a
17982                          * retransmit, as the presence of the TFO option may
17983                          * have caused the original SYN or SYN|ACK to have
17984                          * been dropped by a middlebox.
17985                          */
17986                         if (IS_FASTOPEN(tp->t_flags) &&
17987                             (tp->t_rxtshift == 0)) {
17988                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
17989                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17990                                         to.to_tfo_cookie =
17991                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
17992                                         to.to_flags |= TOF_FASTOPEN;
17993                                         wanted_cookie = 1;
17994                                 } else if (tp->t_state == TCPS_SYN_SENT) {
17995                                         to.to_tfo_len =
17996                                                 tp->t_tfo_client_cookie_len;
17997                                         to.to_tfo_cookie =
17998                                                 tp->t_tfo_cookie.client;
17999                                         to.to_flags |= TOF_FASTOPEN;
18000                                         wanted_cookie = 1;
18001                                         /*
18002                                          * If we wind up having more data to
18003                                          * send with the SYN than can fit in
18004                                          * one segment, don't send any more
18005                                          * until the SYN|ACK comes back from
18006                                          * the other end.
18007                                          */
18008                                         sendalot = 0;
18009                                 }
18010                         }
18011                 }
18012                 /* Window scaling. */
18013                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
18014                         to.to_wscale = tp->request_r_scale;
18015                         to.to_flags |= TOF_SCALE;
18016                 }
18017                 /* Timestamps. */
18018                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
18019                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
18020                         to.to_tsval = ms_cts + tp->ts_offset;
18021                         to.to_tsecr = tp->ts_recent;
18022                         to.to_flags |= TOF_TS;
18023                 }
18024                 /* Set receive buffer autosizing timestamp. */
18025                 if (tp->rfbuf_ts == 0 &&
18026                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
18027                         tp->rfbuf_ts = tcp_ts_getticks();
18028                 /* Selective ACK's. */
18029                 if (tp->t_flags & TF_SACK_PERMIT) {
18030                         if (flags & TH_SYN)
18031                                 to.to_flags |= TOF_SACKPERM;
18032                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18033                                  tp->rcv_numsacks > 0) {
18034                                 to.to_flags |= TOF_SACK;
18035                                 to.to_nsacks = tp->rcv_numsacks;
18036                                 to.to_sacks = (u_char *)tp->sackblks;
18037                         }
18038                 }
18039 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18040                 /* TCP-MD5 (RFC2385). */
18041                 if (tp->t_flags & TF_SIGNATURE)
18042                         to.to_flags |= TOF_SIGNATURE;
18043 #endif                          /* TCP_SIGNATURE */
18044
18045                 /* Processing the options. */
18046                 hdrlen += optlen = tcp_addoptions(&to, opt);
18047                 /*
18048                  * If we wanted a TFO option to be added, but it was unable
18049                  * to fit, ensure no data is sent.
18050                  */
18051                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
18052                     !(to.to_flags & TOF_FASTOPEN))
18053                         len = 0;
18054         }
18055         if (tp->t_port) {
18056                 if (V_tcp_udp_tunneling_port == 0) {
18057                         /* The port was removed?? */
18058                         SOCKBUF_UNLOCK(&so->so_snd);
18059 #ifdef TCP_ACCOUNTING
18060                         crtsc = get_cyclecount();
18061                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18062                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18063                         }
18064                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18065                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18066                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18067                         }
18068                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18069                         sched_unpin();
18070 #endif
18071                         return (EHOSTUNREACH);
18072                 }
18073                 hdrlen += sizeof(struct udphdr);
18074         }
18075 #ifdef INET6
18076         if (isipv6)
18077                 ipoptlen = ip6_optlen(inp);
18078         else
18079 #endif
18080                 if (inp->inp_options)
18081                         ipoptlen = inp->inp_options->m_len -
18082                                 offsetof(struct ipoption, ipopt_list);
18083                 else
18084                         ipoptlen = 0;
18085 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18086         ipoptlen += ipsec_optlen;
18087 #endif
18088
18089         /*
18090          * Adjust data length if insertion of options will bump the packet
18091          * length beyond the t_maxseg length. Clear the FIN bit because we
18092          * cut off the tail of the segment.
18093          */
18094         if (len + optlen + ipoptlen > tp->t_maxseg) {
18095                 if (tso) {
18096                         uint32_t if_hw_tsomax;
18097                         uint32_t moff;
18098                         int32_t max_len;
18099
18100                         /* extract TSO information */
18101                         if_hw_tsomax = tp->t_tsomax;
18102                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18103                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18104                         KASSERT(ipoptlen == 0,
18105                                 ("%s: TSO can't do IP options", __func__));
18106
18107                         /*
18108                          * Check if we should limit by maximum payload
18109                          * length:
18110                          */
18111                         if (if_hw_tsomax != 0) {
18112                                 /* compute maximum TSO length */
18113                                 max_len = (if_hw_tsomax - hdrlen -
18114                                            max_linkhdr);
18115                                 if (max_len <= 0) {
18116                                         len = 0;
18117                                 } else if (len > max_len) {
18118                                         sendalot = 1;
18119                                         len = max_len;
18120                                         mark = 2;
18121                                 }
18122                         }
18123                         /*
18124                          * Prevent the last segment from being fractional
18125                          * unless the send sockbuf can be emptied:
18126                          */
18127                         max_len = (tp->t_maxseg - optlen);
18128                         if ((sb_offset + len) < sbavail(sb)) {
18129                                 moff = len % (u_int)max_len;
18130                                 if (moff != 0) {
18131                                         mark = 3;
18132                                         len -= moff;
18133                                 }
18134                         }
18135                         /*
18136                          * In case there are too many small fragments don't
18137                          * use TSO:
18138                          */
18139                         if (len <= segsiz) {
18140                                 mark = 4;
18141                                 tso = 0;
18142                         }
18143                         /*
18144                          * Send the FIN in a separate segment after the bulk
18145                          * sending is done. We don't trust the TSO
18146                          * implementations to clear the FIN flag on all but
18147                          * the last segment.
18148                          */
18149                         if (tp->t_flags & TF_NEEDFIN) {
18150                                 sendalot = 4;
18151                         }
18152                 } else {
18153                         mark = 5;
18154                         if (optlen + ipoptlen >= tp->t_maxseg) {
18155                                 /*
18156                                  * Since we don't have enough space to put
18157                                  * the IP header chain and the TCP header in
18158                                  * one packet as required by RFC 7112, don't
18159                                  * send it. Also ensure that at least one
18160                                  * byte of the payload can be put into the
18161                                  * TCP segment.
18162                                  */
18163                                 SOCKBUF_UNLOCK(&so->so_snd);
18164                                 error = EMSGSIZE;
18165                                 sack_rxmit = 0;
18166                                 goto out;
18167                         }
18168                         len = tp->t_maxseg - optlen - ipoptlen;
18169                         sendalot = 5;
18170                 }
18171         } else {
18172                 tso = 0;
18173                 mark = 6;
18174         }
18175         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18176                 ("%s: len > IP_MAXPACKET", __func__));
18177 #ifdef DIAGNOSTIC
18178 #ifdef INET6
18179         if (max_linkhdr + hdrlen > MCLBYTES)
18180 #else
18181                 if (max_linkhdr + hdrlen > MHLEN)
18182 #endif
18183                         panic("tcphdr too big");
18184 #endif
18185
18186         /*
18187          * This KASSERT is here to catch edge cases at a well defined place.
18188          * Before, those had triggered (random) panic conditions further
18189          * down.
18190          */
18191         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18192         if ((len == 0) &&
18193             (flags & TH_FIN) &&
18194             (sbused(sb))) {
18195                 /*
18196                  * We have outstanding data, don't send a fin by itself!.
18197                  */
18198                 goto just_return;
18199         }
18200         /*
18201          * Grab a header mbuf, attaching a copy of data to be transmitted,
18202          * and initialize the header from the template for sends on this
18203          * connection.
18204          */
18205         hw_tls = tp->t_nic_ktls_xmit != 0;
18206         if (len) {
18207                 uint32_t max_val;
18208                 uint32_t moff;
18209
18210                 if (rack->r_ctl.rc_pace_max_segs)
18211                         max_val = rack->r_ctl.rc_pace_max_segs;
18212                 else if (rack->rc_user_set_max_segs)
18213                         max_val = rack->rc_user_set_max_segs * segsiz;
18214                 else
18215                         max_val = len;
18216                 /*
18217                  * We allow a limit on sending with hptsi.
18218                  */
18219                 if (len > max_val) {
18220                         mark = 7;
18221                         len = max_val;
18222                 }
18223 #ifdef INET6
18224                 if (MHLEN < hdrlen + max_linkhdr)
18225                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18226                 else
18227 #endif
18228                         m = m_gethdr(M_NOWAIT, MT_DATA);
18229
18230                 if (m == NULL) {
18231                         SOCKBUF_UNLOCK(sb);
18232                         error = ENOBUFS;
18233                         sack_rxmit = 0;
18234                         goto out;
18235                 }
18236                 m->m_data += max_linkhdr;
18237                 m->m_len = hdrlen;
18238
18239                 /*
18240                  * Start the m_copy functions from the closest mbuf to the
18241                  * sb_offset in the socket buffer chain.
18242                  */
18243                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
18244                 s_mb = mb;
18245                 s_moff = moff;
18246                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18247                         m_copydata(mb, moff, (int)len,
18248                                    mtod(m, caddr_t)+hdrlen);
18249                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18250                                 sbsndptr_adv(sb, mb, len);
18251                         m->m_len += len;
18252                 } else {
18253                         struct sockbuf *msb;
18254
18255                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18256                                 msb = NULL;
18257                         else
18258                                 msb = sb;
18259                         m->m_next = tcp_m_copym(
18260                                 mb, moff, &len,
18261                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18262                                 ((rsm == NULL) ? hw_tls : 0)
18263 #ifdef NETFLIX_COPY_ARGS
18264                                 , &s_mb, &s_moff
18265 #endif
18266                                 );
18267                         if (len <= (tp->t_maxseg - optlen)) {
18268                                 /*
18269                                  * Must have ran out of mbufs for the copy
18270                                  * shorten it to no longer need tso. Lets
18271                                  * not put on sendalot since we are low on
18272                                  * mbufs.
18273                                  */
18274                                 tso = 0;
18275                         }
18276                         if (m->m_next == NULL) {
18277                                 SOCKBUF_UNLOCK(sb);
18278                                 (void)m_free(m);
18279                                 error = ENOBUFS;
18280                                 sack_rxmit = 0;
18281                                 goto out;
18282                         }
18283                 }
18284                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18285                         if (rsm && (rsm->r_flags & RACK_TLP)) {
18286                                 /*
18287                                  * TLP should not count in retran count, but
18288                                  * in its own bin
18289                                  */
18290                                 counter_u64_add(rack_tlp_retran, 1);
18291                                 counter_u64_add(rack_tlp_retran_bytes, len);
18292                         } else {
18293                                 tp->t_sndrexmitpack++;
18294                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18295                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18296                         }
18297 #ifdef STATS
18298                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18299                                                  len);
18300 #endif
18301                 } else {
18302                         KMOD_TCPSTAT_INC(tcps_sndpack);
18303                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18304 #ifdef STATS
18305                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18306                                                  len);
18307 #endif
18308                 }
18309                 /*
18310                  * If we're sending everything we've got, set PUSH. (This
18311                  * will keep happy those implementations which only give
18312                  * data to the user when a buffer fills or a PUSH comes in.)
18313                  */
18314                 if (sb_offset + len == sbused(sb) &&
18315                     sbused(sb) &&
18316                     !(flags & TH_SYN)) {
18317                         flags |= TH_PUSH;
18318                         add_flag |= RACK_HAD_PUSH;
18319                 }
18320
18321                 SOCKBUF_UNLOCK(sb);
18322         } else {
18323                 SOCKBUF_UNLOCK(sb);
18324                 if (tp->t_flags & TF_ACKNOW)
18325                         KMOD_TCPSTAT_INC(tcps_sndacks);
18326                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
18327                         KMOD_TCPSTAT_INC(tcps_sndctrl);
18328                 else
18329                         KMOD_TCPSTAT_INC(tcps_sndwinup);
18330
18331                 m = m_gethdr(M_NOWAIT, MT_DATA);
18332                 if (m == NULL) {
18333                         error = ENOBUFS;
18334                         sack_rxmit = 0;
18335                         goto out;
18336                 }
18337 #ifdef INET6
18338                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18339                     MHLEN >= hdrlen) {
18340                         M_ALIGN(m, hdrlen);
18341                 } else
18342 #endif
18343                         m->m_data += max_linkhdr;
18344                 m->m_len = hdrlen;
18345         }
18346         SOCKBUF_UNLOCK_ASSERT(sb);
18347         m->m_pkthdr.rcvif = (struct ifnet *)0;
18348 #ifdef MAC
18349         mac_inpcb_create_mbuf(inp, m);
18350 #endif
18351         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18352 #ifdef INET6
18353                 if (isipv6)
18354                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18355                 else
18356 #endif                          /* INET6 */
18357 #ifdef INET
18358                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18359 #endif
18360                 th = rack->r_ctl.fsb.th;
18361                 udp = rack->r_ctl.fsb.udp;
18362                 if (udp) {
18363 #ifdef INET6
18364                         if (isipv6)
18365                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18366                         else
18367 #endif                          /* INET6 */
18368                                 ulen = hdrlen + len - sizeof(struct ip);
18369                         udp->uh_ulen = htons(ulen);
18370                 }
18371         } else {
18372 #ifdef INET6
18373                 if (isipv6) {
18374                         ip6 = mtod(m, struct ip6_hdr *);
18375                         if (tp->t_port) {
18376                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18377                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18378                                 udp->uh_dport = tp->t_port;
18379                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18380                                 udp->uh_ulen = htons(ulen);
18381                                 th = (struct tcphdr *)(udp + 1);
18382                         } else
18383                                 th = (struct tcphdr *)(ip6 + 1);
18384                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
18385                 } else
18386 #endif                          /* INET6 */
18387                 {
18388 #ifdef INET
18389                         ip = mtod(m, struct ip *);
18390                         if (tp->t_port) {
18391                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18392                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18393                                 udp->uh_dport = tp->t_port;
18394                                 ulen = hdrlen + len - sizeof(struct ip);
18395                                 udp->uh_ulen = htons(ulen);
18396                                 th = (struct tcphdr *)(udp + 1);
18397                         } else
18398                                 th = (struct tcphdr *)(ip + 1);
18399                         tcpip_fillheaders(inp, tp->t_port, ip, th);
18400 #endif
18401                 }
18402         }
18403         /*
18404          * Fill in fields, remembering maximum advertised window for use in
18405          * delaying messages about window sizes. If resending a FIN, be sure
18406          * not to use a new sequence number.
18407          */
18408         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18409             tp->snd_nxt == tp->snd_max)
18410                 tp->snd_nxt--;
18411         /*
18412          * If we are starting a connection, send ECN setup SYN packet. If we
18413          * are on a retransmit, we may resend those bits a number of times
18414          * as per RFC 3168.
18415          */
18416         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
18417                 flags |= tcp_ecn_output_syn_sent(tp);
18418         }
18419         /* Also handle parallel SYN for ECN */
18420         if (TCPS_HAVERCVDSYN(tp->t_state) &&
18421             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18422                 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
18423                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18424                     (tp->t_flags2 & TF2_ECN_SND_ECE))
18425                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18426 #ifdef INET6
18427                 if (isipv6) {
18428                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18429                         ip6->ip6_flow |= htonl(ect << 20);
18430                 }
18431                 else
18432 #endif
18433                 {
18434 #ifdef INET
18435                         ip->ip_tos &= ~IPTOS_ECN_MASK;
18436                         ip->ip_tos |= ect;
18437 #endif
18438                 }
18439         }
18440         /*
18441          * If we are doing retransmissions, then snd_nxt will not reflect
18442          * the first unsent octet.  For ACK only packets, we do not want the
18443          * sequence number of the retransmitted packet, we want the sequence
18444          * number of the next unsent octet.  So, if there is no data (and no
18445          * SYN or FIN), use snd_max instead of snd_nxt when filling in
18446          * ti_seq.  But if we are in persist state, snd_max might reflect
18447          * one byte beyond the right edge of the window, so use snd_nxt in
18448          * that case, since we know we aren't doing a retransmission.
18449          * (retransmit and persist are mutually exclusive...)
18450          */
18451         if (sack_rxmit == 0) {
18452                 if (len || (flags & (TH_SYN | TH_FIN))) {
18453                         th->th_seq = htonl(tp->snd_nxt);
18454                         rack_seq = tp->snd_nxt;
18455                 } else {
18456                         th->th_seq = htonl(tp->snd_max);
18457                         rack_seq = tp->snd_max;
18458                 }
18459         } else {
18460                 th->th_seq = htonl(rsm->r_start);
18461                 rack_seq = rsm->r_start;
18462         }
18463         th->th_ack = htonl(tp->rcv_nxt);
18464         tcp_set_flags(th, flags);
18465         /*
18466          * Calculate receive window.  Don't shrink window, but avoid silly
18467          * window syndrome.
18468          * If a RST segment is sent, advertise a window of zero.
18469          */
18470         if (flags & TH_RST) {
18471                 recwin = 0;
18472         } else {
18473                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18474                     recwin < (long)segsiz) {
18475                         recwin = 0;
18476                 }
18477                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18478                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18479                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18480         }
18481
18482         /*
18483          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18484          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18485          * handled in syncache.
18486          */
18487         if (flags & TH_SYN)
18488                 th->th_win = htons((u_short)
18489                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18490         else {
18491                 /* Avoid shrinking window with window scaling. */
18492                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
18493                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18494         }
18495         /*
18496          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18497          * window.  This may cause the remote transmitter to stall.  This
18498          * flag tells soreceive() to disable delayed acknowledgements when
18499          * draining the buffer.  This can occur if the receiver is
18500          * attempting to read more data than can be buffered prior to
18501          * transmitting on the connection.
18502          */
18503         if (th->th_win == 0) {
18504                 tp->t_sndzerowin++;
18505                 tp->t_flags |= TF_RXWIN0SENT;
18506         } else
18507                 tp->t_flags &= ~TF_RXWIN0SENT;
18508         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
18509         /* Now are we using fsb?, if so copy the template data to the mbuf */
18510         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18511                 uint8_t *cpto;
18512
18513                 cpto = mtod(m, uint8_t *);
18514                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18515                 /*
18516                  * We have just copied in:
18517                  * IP/IP6
18518                  * <optional udphdr>
18519                  * tcphdr (no options)
18520                  *
18521                  * We need to grab the correct pointers into the mbuf
18522                  * for both the tcp header, and possibly the udp header (if tunneling).
18523                  * We do this by using the offset in the copy buffer and adding it
18524                  * to the mbuf base pointer (cpto).
18525                  */
18526 #ifdef INET6
18527                 if (isipv6)
18528                         ip6 = mtod(m, struct ip6_hdr *);
18529                 else
18530 #endif                          /* INET6 */
18531 #ifdef INET
18532                         ip = mtod(m, struct ip *);
18533 #endif
18534                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18535                 /* If we have a udp header lets set it into the mbuf as well */
18536                 if (udp)
18537                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18538         }
18539 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18540         if (to.to_flags & TOF_SIGNATURE) {
18541                 /*
18542                  * Calculate MD5 signature and put it into the place
18543                  * determined before.
18544                  * NOTE: since TCP options buffer doesn't point into
18545                  * mbuf's data, calculate offset and use it.
18546                  */
18547                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18548                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18549                         /*
18550                          * Do not send segment if the calculation of MD5
18551                          * digest has failed.
18552                          */
18553                         goto out;
18554                 }
18555         }
18556 #endif
18557         if (optlen) {
18558                 bcopy(opt, th + 1, optlen);
18559                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18560         }
18561         /*
18562          * Put TCP length in extended header, and then checksum extended
18563          * header and data.
18564          */
18565         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
18566 #ifdef INET6
18567         if (isipv6) {
18568                 /*
18569                  * ip6_plen is not need to be filled now, and will be filled
18570                  * in ip6_output.
18571                  */
18572                 if (tp->t_port) {
18573                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18574                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18575                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18576                         th->th_sum = htons(0);
18577                         UDPSTAT_INC(udps_opackets);
18578                 } else {
18579                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18580                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18581                         th->th_sum = in6_cksum_pseudo(ip6,
18582                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18583                                                       0);
18584                 }
18585         }
18586 #endif
18587 #if defined(INET6) && defined(INET)
18588         else
18589 #endif
18590 #ifdef INET
18591         {
18592                 if (tp->t_port) {
18593                         m->m_pkthdr.csum_flags = CSUM_UDP;
18594                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18595                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18596                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18597                         th->th_sum = htons(0);
18598                         UDPSTAT_INC(udps_opackets);
18599                 } else {
18600                         m->m_pkthdr.csum_flags = CSUM_TCP;
18601                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18602                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
18603                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18604                                                                         IPPROTO_TCP + len + optlen));
18605                 }
18606                 /* IP version must be set here for ipv4/ipv6 checking later */
18607                 KASSERT(ip->ip_v == IPVERSION,
18608                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
18609         }
18610 #endif
18611         /*
18612          * Enable TSO and specify the size of the segments. The TCP pseudo
18613          * header checksum is always provided. XXX: Fixme: This is currently
18614          * not the case for IPv6.
18615          */
18616         if (tso) {
18617                 KASSERT(len > tp->t_maxseg - optlen,
18618                         ("%s: len <= tso_segsz", __func__));
18619                 m->m_pkthdr.csum_flags |= CSUM_TSO;
18620                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18621         }
18622         KASSERT(len + hdrlen == m_length(m, NULL),
18623                 ("%s: mbuf chain different than expected: %d + %u != %u",
18624                  __func__, len, hdrlen, m_length(m, NULL)));
18625
18626 #ifdef TCP_HHOOK
18627         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18628         hhook_run_tcp_est_out(tp, th, &to, len, tso);
18629 #endif
18630         /* We're getting ready to send; log now. */
18631         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18632                 union tcp_log_stackspecific log;
18633
18634                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18635                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18636                 if (rack->rack_no_prr)
18637                         log.u_bbr.flex1 = 0;
18638                 else
18639                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18640                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18641                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18642                 log.u_bbr.flex4 = orig_len;
18643                 /* Save off the early/late values */
18644                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18645                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18646                 log.u_bbr.bw_inuse = rack_get_bw(rack);
18647                 log.u_bbr.flex8 = 0;
18648                 if (rsm) {
18649                         if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18650                                 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18651                                 counter_u64_add(rack_collapsed_win_rxt, 1);
18652                                 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18653                         }
18654                         if (doing_tlp)
18655                                 log.u_bbr.flex8 = 2;
18656                         else
18657                                 log.u_bbr.flex8 = 1;
18658                 } else {
18659                         if (doing_tlp)
18660                                 log.u_bbr.flex8 = 3;
18661                         else
18662                                 log.u_bbr.flex8 = 0;
18663                 }
18664                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18665                 log.u_bbr.flex7 = mark;
18666                 log.u_bbr.flex7 <<= 8;
18667                 log.u_bbr.flex7 |= pass;
18668                 log.u_bbr.pkts_out = tp->t_maxseg;
18669                 log.u_bbr.timeStamp = cts;
18670                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18671                 log.u_bbr.lt_epoch = cwnd_to_use;
18672                 log.u_bbr.delivered = sendalot;
18673                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18674                                      len, &log, false, NULL, NULL, 0, &tv);
18675         } else
18676                 lgb = NULL;
18677
18678         /*
18679          * Fill in IP length and desired time to live and send to IP level.
18680          * There should be a better way to handle ttl and tos; we could keep
18681          * them in the template, but need a way to checksum without them.
18682          */
18683         /*
18684          * m->m_pkthdr.len should have been set before cksum calcuration,
18685          * because in6_cksum() need it.
18686          */
18687 #ifdef INET6
18688         if (isipv6) {
18689                 /*
18690                  * we separately set hoplimit for every segment, since the
18691                  * user might want to change the value via setsockopt. Also,
18692                  * desired default hop limit might be changed via Neighbor
18693                  * Discovery.
18694                  */
18695                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18696
18697                 /*
18698                  * Set the packet size here for the benefit of DTrace
18699                  * probes. ip6_output() will set it properly; it's supposed
18700                  * to include the option header lengths as well.
18701                  */
18702                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18703
18704                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18705                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18706                 else
18707                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18708
18709                 if (tp->t_state == TCPS_SYN_SENT)
18710                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18711
18712                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18713                 /* TODO: IPv6 IP6TOS_ECT bit on */
18714                 error = ip6_output(m,
18715 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18716                                    inp->in6p_outputopts,
18717 #else
18718                                    NULL,
18719 #endif
18720                                    &inp->inp_route6,
18721                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18722                                    NULL, NULL, inp);
18723
18724                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18725                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18726         }
18727 #endif                          /* INET6 */
18728 #if defined(INET) && defined(INET6)
18729         else
18730 #endif
18731 #ifdef INET
18732         {
18733                 ip->ip_len = htons(m->m_pkthdr.len);
18734 #ifdef INET6
18735                 if (inp->inp_vflag & INP_IPV6PROTO)
18736                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18737 #endif                          /* INET6 */
18738                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18739                 /*
18740                  * If we do path MTU discovery, then we set DF on every
18741                  * packet. This might not be the best thing to do according
18742                  * to RFC3390 Section 2. However the tcp hostcache migitates
18743                  * the problem so it affects only the first tcp connection
18744                  * with a host.
18745                  *
18746                  * NB: Don't set DF on small MTU/MSS to have a safe
18747                  * fallback.
18748                  */
18749                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18750                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18751                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18752                                 ip->ip_off |= htons(IP_DF);
18753                         }
18754                 } else {
18755                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18756                 }
18757
18758                 if (tp->t_state == TCPS_SYN_SENT)
18759                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18760
18761                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18762
18763                 error = ip_output(m,
18764 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18765                                   inp->inp_options,
18766 #else
18767                                   NULL,
18768 #endif
18769                                   &inp->inp_route,
18770                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18771                                   inp);
18772                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18773                         mtu = inp->inp_route.ro_nh->nh_mtu;
18774         }
18775 #endif                          /* INET */
18776
18777 out:
18778         if (lgb) {
18779                 lgb->tlb_errno = error;
18780                 lgb = NULL;
18781         }
18782         /*
18783          * In transmit state, time the transmission and arrange for the
18784          * retransmit.  In persist state, just set snd_max.
18785          */
18786         if (error == 0) {
18787                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18788                 if (rsm && doing_tlp) {
18789                         rack->rc_last_sent_tlp_past_cumack = 0;
18790                         rack->rc_last_sent_tlp_seq_valid = 1;
18791                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18792                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18793                 }
18794                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18795                 if (rsm && (doing_tlp == 0)) {
18796                         /* Set we retransmitted */
18797                         rack->rc_gp_saw_rec = 1;
18798                 } else {
18799                         if (cwnd_to_use > tp->snd_ssthresh) {
18800                                 /* Set we sent in CA */
18801                                 rack->rc_gp_saw_ca = 1;
18802                         } else {
18803                                 /* Set we sent in SS */
18804                                 rack->rc_gp_saw_ss = 1;
18805                         }
18806                 }
18807                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18808                     (tp->t_flags & TF_SACK_PERMIT) &&
18809                     tp->rcv_numsacks > 0)
18810                         tcp_clean_dsack_blocks(tp);
18811                 tot_len_this_send += len;
18812                 if (len == 0)
18813                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18814                 else if (len == 1) {
18815                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18816                 } else if (len > 1) {
18817                         int idx;
18818
18819                         idx = (len / segsiz) + 3;
18820                         if (idx >= TCP_MSS_ACCT_ATIMER)
18821                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18822                         else
18823                                 counter_u64_add(rack_out_size[idx], 1);
18824                 }
18825         }
18826         if ((rack->rack_no_prr == 0) &&
18827             sub_from_prr &&
18828             (error == 0)) {
18829                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18830                         rack->r_ctl.rc_prr_sndcnt -= len;
18831                 else
18832                         rack->r_ctl.rc_prr_sndcnt = 0;
18833         }
18834         sub_from_prr = 0;
18835         if (doing_tlp) {
18836                 /* Make sure the TLP is added */
18837                 add_flag |= RACK_TLP;
18838         } else if (rsm) {
18839                 /* If its a resend without TLP then it must not have the flag */
18840                 rsm->r_flags &= ~RACK_TLP;
18841         }
18842         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18843                         rack_to_usec_ts(&tv),
18844                         rsm, add_flag, s_mb, s_moff, hw_tls);
18845
18846
18847         if ((error == 0) &&
18848             (len > 0) &&
18849             (tp->snd_una == tp->snd_max))
18850                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18851         {
18852                 tcp_seq startseq = tp->snd_nxt;
18853
18854                 /* Track our lost count */
18855                 if (rsm && (doing_tlp == 0))
18856                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18857                 /*
18858                  * Advance snd_nxt over sequence space of this segment.
18859                  */
18860                 if (error)
18861                         /* We don't log or do anything with errors */
18862                         goto nomore;
18863                 if (doing_tlp == 0) {
18864                         if (rsm == NULL) {
18865                                 /*
18866                                  * Not a retransmission of some
18867                                  * sort, new data is going out so
18868                                  * clear our TLP count and flag.
18869                                  */
18870                                 rack->rc_tlp_in_progress = 0;
18871                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18872                         }
18873                 } else {
18874                         /*
18875                          * We have just sent a TLP, mark that it is true
18876                          * and make sure our in progress is set so we
18877                          * continue to check the count.
18878                          */
18879                         rack->rc_tlp_in_progress = 1;
18880                         rack->r_ctl.rc_tlp_cnt_out++;
18881                 }
18882                 if (flags & (TH_SYN | TH_FIN)) {
18883                         if (flags & TH_SYN)
18884                                 tp->snd_nxt++;
18885                         if (flags & TH_FIN) {
18886                                 tp->snd_nxt++;
18887                                 tp->t_flags |= TF_SENTFIN;
18888                         }
18889                 }
18890                 /* In the ENOBUFS case we do *not* update snd_max */
18891                 if (sack_rxmit)
18892                         goto nomore;
18893
18894                 tp->snd_nxt += len;
18895                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18896                         if (tp->snd_una == tp->snd_max) {
18897                                 /*
18898                                  * Update the time we just added data since
18899                                  * none was outstanding.
18900                                  */
18901                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18902                                 tp->t_acktime = ticks;
18903                         }
18904                         tp->snd_max = tp->snd_nxt;
18905                         /*
18906                          * Time this transmission if not a retransmission and
18907                          * not currently timing anything.
18908                          * This is only relevant in case of switching back to
18909                          * the base stack.
18910                          */
18911                         if (tp->t_rtttime == 0) {
18912                                 tp->t_rtttime = ticks;
18913                                 tp->t_rtseq = startseq;
18914                                 KMOD_TCPSTAT_INC(tcps_segstimed);
18915                         }
18916                         if (len &&
18917                             ((tp->t_flags & TF_GPUTINPROG) == 0))
18918                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18919                 }
18920                 /*
18921                  * If we are doing FO we need to update the mbuf position and subtract
18922                  * this happens when the peer sends us duplicate information and
18923                  * we thus want to send a DSACK.
18924                  *
18925                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18926                  * turned off? If not then we are going to echo multiple DSACK blocks
18927                  * out (with the TSO), which we should not be doing.
18928                  */
18929                 if (rack->r_fast_output && len) {
18930                         if (rack->r_ctl.fsb.left_to_send > len)
18931                                 rack->r_ctl.fsb.left_to_send -= len;
18932                         else
18933                                 rack->r_ctl.fsb.left_to_send = 0;
18934                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18935                                 rack->r_fast_output = 0;
18936                         if (rack->r_fast_output) {
18937                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18938                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18939                         }
18940                 }
18941         }
18942 nomore:
18943         if (error) {
18944                 rack->r_ctl.rc_agg_delayed = 0;
18945                 rack->r_early = 0;
18946                 rack->r_late = 0;
18947                 rack->r_ctl.rc_agg_early = 0;
18948                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
18949                 /*
18950                  * Failures do not advance the seq counter above. For the
18951                  * case of ENOBUFS we will fall out and retry in 1ms with
18952                  * the hpts. Everything else will just have to retransmit
18953                  * with the timer.
18954                  *
18955                  * In any case, we do not want to loop around for another
18956                  * send without a good reason.
18957                  */
18958                 sendalot = 0;
18959                 switch (error) {
18960                 case EPERM:
18961                         tp->t_softerror = error;
18962 #ifdef TCP_ACCOUNTING
18963                         crtsc = get_cyclecount();
18964                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18965                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18966                         }
18967                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18968                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18969                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18970                         }
18971                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18972                         sched_unpin();
18973 #endif
18974                         return (error);
18975                 case ENOBUFS:
18976                         /*
18977                          * Pace us right away to retry in a some
18978                          * time
18979                          */
18980                         if (rack->r_ctl.crte != NULL) {
18981                                 rack_trace_point(rack, RACK_TP_HWENOBUF);
18982                         } else
18983                                 rack_trace_point(rack, RACK_TP_ENOBUF);
18984                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18985                         if (rack->rc_enobuf < 0x7f)
18986                                 rack->rc_enobuf++;
18987                         if (slot < (10 * HPTS_USEC_IN_MSEC))
18988                                 slot = 10 * HPTS_USEC_IN_MSEC;
18989                         if (rack->r_ctl.crte != NULL) {
18990                                 counter_u64_add(rack_saw_enobuf_hw, 1);
18991                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
18992                         }
18993                         counter_u64_add(rack_saw_enobuf, 1);
18994                         goto enobufs;
18995                 case EMSGSIZE:
18996                         /*
18997                          * For some reason the interface we used initially
18998                          * to send segments changed to another or lowered
18999                          * its MTU. If TSO was active we either got an
19000                          * interface without TSO capabilits or TSO was
19001                          * turned off. If we obtained mtu from ip_output()
19002                          * then update it and try again.
19003                          */
19004                         if (tso)
19005                                 tp->t_flags &= ~TF_TSO;
19006                         if (mtu != 0) {
19007                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
19008                                 goto again;
19009                         }
19010                         slot = 10 * HPTS_USEC_IN_MSEC;
19011                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19012 #ifdef TCP_ACCOUNTING
19013                         crtsc = get_cyclecount();
19014                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19015                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19016                         }
19017                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19018                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19019                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19020                         }
19021                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19022                         sched_unpin();
19023 #endif
19024                         return (error);
19025                 case ENETUNREACH:
19026                         counter_u64_add(rack_saw_enetunreach, 1);
19027                 case EHOSTDOWN:
19028                 case EHOSTUNREACH:
19029                 case ENETDOWN:
19030                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
19031                                 tp->t_softerror = error;
19032                         }
19033                         /* FALLTHROUGH */
19034                 default:
19035                         slot = 10 * HPTS_USEC_IN_MSEC;
19036                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19037 #ifdef TCP_ACCOUNTING
19038                         crtsc = get_cyclecount();
19039                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19040                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19041                         }
19042                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19043                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19044                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19045                         }
19046                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19047                         sched_unpin();
19048 #endif
19049                         return (error);
19050                 }
19051         } else {
19052                 rack->rc_enobuf = 0;
19053                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19054                         rack->r_ctl.retran_during_recovery += len;
19055         }
19056         KMOD_TCPSTAT_INC(tcps_sndtotal);
19057
19058         /*
19059          * Data sent (as far as we can tell). If this advertises a larger
19060          * window than any other segment, then remember the size of the
19061          * advertised window. Any pending ACK has now been sent.
19062          */
19063         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
19064                 tp->rcv_adv = tp->rcv_nxt + recwin;
19065
19066         tp->last_ack_sent = tp->rcv_nxt;
19067         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19068 enobufs:
19069         if (sendalot) {
19070                 /* Do we need to turn off sendalot? */
19071                 if (rack->r_ctl.rc_pace_max_segs &&
19072                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
19073                         /* We hit our max. */
19074                         sendalot = 0;
19075                 } else if ((rack->rc_user_set_max_segs) &&
19076                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
19077                         /* We hit the user defined max */
19078                         sendalot = 0;
19079                 }
19080         }
19081         if ((error == 0) && (flags & TH_FIN))
19082                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
19083         if (flags & TH_RST) {
19084                 /*
19085                  * We don't send again after sending a RST.
19086                  */
19087                 slot = 0;
19088                 sendalot = 0;
19089                 if (error == 0)
19090                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
19091         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
19092                 /*
19093                  * Get our pacing rate, if an error
19094                  * occurred in sending (ENOBUF) we would
19095                  * hit the else if with slot preset. Other
19096                  * errors return.
19097                  */
19098                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
19099         }
19100         if (rsm &&
19101             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
19102             rack->use_rack_rr) {
19103                 /* Its a retransmit and we use the rack cheat? */
19104                 if ((slot == 0) ||
19105                     (rack->rc_always_pace == 0) ||
19106                     (rack->r_rr_config == 1)) {
19107                         /*
19108                          * We have no pacing set or we
19109                          * are using old-style rack or
19110                          * we are overridden to use the old 1ms pacing.
19111                          */
19112                         slot = rack->r_ctl.rc_min_to;
19113                 }
19114         }
19115         /* We have sent clear the flag */
19116         rack->r_ent_rec_ns = 0;
19117         if (rack->r_must_retran) {
19118                 if (rsm) {
19119                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
19120                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
19121                                 /*
19122                                  * We have retransmitted all.
19123                                  */
19124                                 rack->r_must_retran = 0;
19125                                 rack->r_ctl.rc_out_at_rto = 0;
19126                         }
19127                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19128                         /*
19129                          * Sending new data will also kill
19130                          * the loop.
19131                          */
19132                         rack->r_must_retran = 0;
19133                         rack->r_ctl.rc_out_at_rto = 0;
19134                 }
19135         }
19136         rack->r_ctl.fsb.recwin = recwin;
19137         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19138             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19139                 /*
19140                  * We hit an RTO and now have past snd_max at the RTO
19141                  * clear all the WAS flags.
19142                  */
19143                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19144         }
19145         if (slot) {
19146                 /* set the rack tcb into the slot N */
19147                 if ((error == 0) &&
19148                     rack_use_rfo &&
19149                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
19150                     (rsm == NULL) &&
19151                     (tp->snd_nxt == tp->snd_max) &&
19152                     (ipoptlen == 0) &&
19153                     (tp->rcv_numsacks == 0) &&
19154                     rack->r_fsb_inited &&
19155                     TCPS_HAVEESTABLISHED(tp->t_state) &&
19156                     (rack->r_must_retran == 0) &&
19157                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
19158                     (len > 0) && (orig_len > 0) &&
19159                     (orig_len > len) &&
19160                     ((orig_len - len) >= segsiz) &&
19161                     ((optlen == 0) ||
19162                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19163                         /* We can send at least one more MSS using our fsb */
19164
19165                         rack->r_fast_output = 1;
19166                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19167                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19168                         rack->r_ctl.fsb.tcp_flags = flags;
19169                         rack->r_ctl.fsb.left_to_send = orig_len - len;
19170                         if (hw_tls)
19171                                 rack->r_ctl.fsb.hw_tls = 1;
19172                         else
19173                                 rack->r_ctl.fsb.hw_tls = 0;
19174                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19175                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19176                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19177                                  (tp->snd_max - tp->snd_una)));
19178                         if (rack->r_ctl.fsb.left_to_send < segsiz)
19179                                 rack->r_fast_output = 0;
19180                         else {
19181                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19182                                         rack->r_ctl.fsb.rfo_apply_push = 1;
19183                                 else
19184                                         rack->r_ctl.fsb.rfo_apply_push = 0;
19185                         }
19186                 } else
19187                         rack->r_fast_output = 0;
19188                 rack_log_fsb(rack, tp, so, flags,
19189                              ipoptlen, orig_len, len, error,
19190                              (rsm == NULL), optlen, __LINE__, 2);
19191         } else if (sendalot) {
19192                 int ret;
19193
19194                 sack_rxmit = 0;
19195                 if ((error == 0) &&
19196                     rack_use_rfo &&
19197                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
19198                     (rsm == NULL) &&
19199                     (ipoptlen == 0) &&
19200                     (tp->rcv_numsacks == 0) &&
19201                     (tp->snd_nxt == tp->snd_max) &&
19202                     (rack->r_must_retran == 0) &&
19203                     rack->r_fsb_inited &&
19204                     TCPS_HAVEESTABLISHED(tp->t_state) &&
19205                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
19206                     (len > 0) && (orig_len > 0) &&
19207                     (orig_len > len) &&
19208                     ((orig_len - len) >= segsiz) &&
19209                     ((optlen == 0) ||
19210                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19211                         /* we can use fast_output for more */
19212
19213                         rack->r_fast_output = 1;
19214                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19215                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19216                         rack->r_ctl.fsb.tcp_flags = flags;
19217                         rack->r_ctl.fsb.left_to_send = orig_len - len;
19218                         if (hw_tls)
19219                                 rack->r_ctl.fsb.hw_tls = 1;
19220                         else
19221                                 rack->r_ctl.fsb.hw_tls = 0;
19222                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19223                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19224                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19225                                  (tp->snd_max - tp->snd_una)));
19226                         if (rack->r_ctl.fsb.left_to_send < segsiz) {
19227                                 rack->r_fast_output = 0;
19228                         }
19229                         if (rack->r_fast_output) {
19230                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19231                                         rack->r_ctl.fsb.rfo_apply_push = 1;
19232                                 else
19233                                         rack->r_ctl.fsb.rfo_apply_push = 0;
19234                                 rack_log_fsb(rack, tp, so, flags,
19235                                              ipoptlen, orig_len, len, error,
19236                                              (rsm == NULL), optlen, __LINE__, 3);
19237                                 error = 0;
19238                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19239                                 if (ret >= 0)
19240                                         return (ret);
19241                                 else if (error)
19242                                         goto nomore;
19243
19244                         }
19245                 }
19246                 goto again;
19247         }
19248         /* Assure when we leave that snd_nxt will point to top */
19249         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19250                 tp->snd_nxt = tp->snd_max;
19251         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19252 #ifdef TCP_ACCOUNTING
19253         crtsc = get_cyclecount() - ts_val;
19254         if (tot_len_this_send) {
19255                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19256                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
19257                 }
19258                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19259                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19260                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19261                 }
19262                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19263                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19264                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19265                 }
19266                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19267         } else {
19268                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19269                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
19270                 }
19271                 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19272                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19273                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19274                 }
19275                 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19276         }
19277         sched_unpin();
19278 #endif
19279         if (error == ENOBUFS)
19280                 error = 0;
19281         return (error);
19282 }
19283
19284 static void
19285 rack_update_seg(struct tcp_rack *rack)
19286 {
19287         uint32_t orig_val;
19288
19289         orig_val = rack->r_ctl.rc_pace_max_segs;
19290         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19291         if (orig_val != rack->r_ctl.rc_pace_max_segs)
19292                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19293 }
19294
19295 static void
19296 rack_mtu_change(struct tcpcb *tp)
19297 {
19298         /*
19299          * The MSS may have changed
19300          */
19301         struct tcp_rack *rack;
19302         struct rack_sendmap *rsm;
19303
19304         rack = (struct tcp_rack *)tp->t_fb_ptr;
19305         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19306                 /*
19307                  * The MTU has changed we need to resend everything
19308                  * since all we have sent is lost. We first fix
19309                  * up the mtu though.
19310                  */
19311                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19312                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
19313                 rack_remxt_tmr(tp);
19314                 rack->r_fast_output = 0;
19315                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19316                                                 rack->r_ctl.rc_sacked);
19317                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19318                 rack->r_must_retran = 1;
19319                 /* Mark all inflight to needing to be rxt'd */
19320                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19321                         rsm->r_flags |= RACK_MUST_RXT;
19322                 }
19323         }
19324         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19325         /* We don't use snd_nxt to retransmit */
19326         tp->snd_nxt = tp->snd_max;
19327 }
19328
19329 static int
19330 rack_set_profile(struct tcp_rack *rack, int prof)
19331 {
19332         int err = EINVAL;
19333         if (prof == 1) {
19334                 /* pace_always=1 */
19335                 if (rack->rc_always_pace == 0) {
19336                         if (tcp_can_enable_pacing() == 0)
19337                                 return (EBUSY);
19338                 }
19339                 rack->rc_always_pace = 1;
19340                 if (rack->use_fixed_rate || rack->gp_ready)
19341                         rack_set_cc_pacing(rack);
19342                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19343                 rack->rack_attempt_hdwr_pace = 0;
19344                 /* cmpack=1 */
19345                 if (rack_use_cmp_acks)
19346                         rack->r_use_cmp_ack = 1;
19347                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19348                     rack->r_use_cmp_ack)
19349                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19350                 /* scwnd=1 */
19351                 rack->rack_enable_scwnd = 1;
19352                 /* dynamic=100 */
19353                 rack->rc_gp_dyn_mul = 1;
19354                 /* gp_inc_ca */
19355                 rack->r_ctl.rack_per_of_gp_ca = 100;
19356                 /* rrr_conf=3 */
19357                 rack->r_rr_config = 3;
19358                 /* npush=2 */
19359                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19360                 /* fillcw=1 */
19361                 rack->rc_pace_to_cwnd = 1;
19362                 rack->rc_pace_fill_if_rttin_range = 0;
19363                 rack->rtt_limit_mul = 0;
19364                 /* noprr=1 */
19365                 rack->rack_no_prr = 1;
19366                 /* lscwnd=1 */
19367                 rack->r_limit_scw = 1;
19368                 /* gp_inc_rec */
19369                 rack->r_ctl.rack_per_of_gp_rec = 90;
19370                 err = 0;
19371
19372         } else if (prof == 3) {
19373                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19374                 /* pace_always=1 */
19375                 if (rack->rc_always_pace == 0) {
19376                         if (tcp_can_enable_pacing() == 0)
19377                                 return (EBUSY);
19378                 }
19379                 rack->rc_always_pace = 1;
19380                 if (rack->use_fixed_rate || rack->gp_ready)
19381                         rack_set_cc_pacing(rack);
19382                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19383                 rack->rack_attempt_hdwr_pace = 0;
19384                 /* cmpack=1 */
19385                 if (rack_use_cmp_acks)
19386                         rack->r_use_cmp_ack = 1;
19387                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19388                     rack->r_use_cmp_ack)
19389                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19390                 /* scwnd=1 */
19391                 rack->rack_enable_scwnd = 1;
19392                 /* dynamic=100 */
19393                 rack->rc_gp_dyn_mul = 1;
19394                 /* gp_inc_ca */
19395                 rack->r_ctl.rack_per_of_gp_ca = 100;
19396                 /* rrr_conf=3 */
19397                 rack->r_rr_config = 3;
19398                 /* npush=2 */
19399                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19400                 /* fillcw=2 */
19401                 rack->rc_pace_to_cwnd = 1;
19402                 rack->r_fill_less_agg = 1;
19403                 rack->rc_pace_fill_if_rttin_range = 0;
19404                 rack->rtt_limit_mul = 0;
19405                 /* noprr=1 */
19406                 rack->rack_no_prr = 1;
19407                 /* lscwnd=1 */
19408                 rack->r_limit_scw = 1;
19409                 /* gp_inc_rec */
19410                 rack->r_ctl.rack_per_of_gp_rec = 90;
19411                 err = 0;
19412
19413
19414         } else if (prof == 2) {
19415                 /* cmpack=1 */
19416                 if (rack->rc_always_pace == 0) {
19417                         if (tcp_can_enable_pacing() == 0)
19418                                 return (EBUSY);
19419                 }
19420                 rack->rc_always_pace = 1;
19421                 if (rack->use_fixed_rate || rack->gp_ready)
19422                         rack_set_cc_pacing(rack);
19423                 rack->r_use_cmp_ack = 1;
19424                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19425                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19426                 /* pace_always=1 */
19427                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19428                 /* scwnd=1 */
19429                 rack->rack_enable_scwnd = 1;
19430                 /* dynamic=100 */
19431                 rack->rc_gp_dyn_mul = 1;
19432                 rack->r_ctl.rack_per_of_gp_ca = 100;
19433                 /* rrr_conf=3 */
19434                 rack->r_rr_config = 3;
19435                 /* npush=2 */
19436                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19437                 /* fillcw=1 */
19438                 rack->rc_pace_to_cwnd = 1;
19439                 rack->rc_pace_fill_if_rttin_range = 0;
19440                 rack->rtt_limit_mul = 0;
19441                 /* noprr=1 */
19442                 rack->rack_no_prr = 1;
19443                 /* lscwnd=0 */
19444                 rack->r_limit_scw = 0;
19445                 err = 0;
19446         } else if (prof == 0) {
19447                 /* This changes things back to the default settings */
19448                 err = 0;
19449                 if (rack->rc_always_pace) {
19450                         tcp_decrement_paced_conn();
19451                         rack_undo_cc_pacing(rack);
19452                         rack->rc_always_pace = 0;
19453                 }
19454                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19455                         rack->rc_always_pace = 1;
19456                         if (rack->use_fixed_rate || rack->gp_ready)
19457                                 rack_set_cc_pacing(rack);
19458                 } else
19459                         rack->rc_always_pace = 0;
19460                 if (rack_dsack_std_based & 0x1) {
19461                         /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19462                         rack->rc_rack_tmr_std_based = 1;
19463                 }
19464                 if (rack_dsack_std_based & 0x2) {
19465                         /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19466                         rack->rc_rack_use_dsack = 1;
19467                 }
19468                 if (rack_use_cmp_acks)
19469                         rack->r_use_cmp_ack = 1;
19470                 else
19471                         rack->r_use_cmp_ack = 0;
19472                 if (rack_disable_prr)
19473                         rack->rack_no_prr = 1;
19474                 else
19475                         rack->rack_no_prr = 0;
19476                 if (rack_gp_no_rec_chg)
19477                         rack->rc_gp_no_rec_chg = 1;
19478                 else
19479                         rack->rc_gp_no_rec_chg = 0;
19480                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19481                         rack->r_mbuf_queue = 1;
19482                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19483                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19484                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19485                 } else {
19486                         rack->r_mbuf_queue = 0;
19487                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19488                 }
19489                 if (rack_enable_shared_cwnd)
19490                         rack->rack_enable_scwnd = 1;
19491                 else
19492                         rack->rack_enable_scwnd = 0;
19493                 if (rack_do_dyn_mul) {
19494                         /* When dynamic adjustment is on CA needs to start at 100% */
19495                         rack->rc_gp_dyn_mul = 1;
19496                         if (rack_do_dyn_mul >= 100)
19497                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19498                 } else {
19499                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19500                         rack->rc_gp_dyn_mul = 0;
19501                 }
19502                 rack->r_rr_config = 0;
19503                 rack->r_ctl.rc_no_push_at_mrtt = 0;
19504                 rack->rc_pace_to_cwnd = 0;
19505                 rack->rc_pace_fill_if_rttin_range = 0;
19506                 rack->rtt_limit_mul = 0;
19507
19508                 if (rack_enable_hw_pacing)
19509                         rack->rack_hdw_pace_ena = 1;
19510                 else
19511                         rack->rack_hdw_pace_ena = 0;
19512                 if (rack_disable_prr)
19513                         rack->rack_no_prr = 1;
19514                 else
19515                         rack->rack_no_prr = 0;
19516                 if (rack_limits_scwnd)
19517                         rack->r_limit_scw  = 1;
19518                 else
19519                         rack->r_limit_scw  = 0;
19520                 err = 0;
19521         }
19522         return (err);
19523 }
19524
19525 static int
19526 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19527 {
19528         struct deferred_opt_list *dol;
19529
19530         dol = malloc(sizeof(struct deferred_opt_list),
19531                      M_TCPFSB, M_NOWAIT|M_ZERO);
19532         if (dol == NULL) {
19533                 /*
19534                  * No space yikes -- fail out..
19535                  */
19536                 return (0);
19537         }
19538         dol->optname = sopt_name;
19539         dol->optval = loptval;
19540         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19541         return (1);
19542 }
19543
19544 static int
19545 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19546                     uint32_t optval, uint64_t loptval)
19547 {
19548         struct epoch_tracker et;
19549         struct sockopt sopt;
19550         struct cc_newreno_opts opt;
19551         struct inpcb *inp = tptoinpcb(tp);
19552         uint64_t val;
19553         int error = 0;
19554         uint16_t ca, ss;
19555
19556         switch (sopt_name) {
19557
19558         case TCP_RACK_DSACK_OPT:
19559                 RACK_OPTS_INC(tcp_rack_dsack_opt);
19560                 if (optval & 0x1) {
19561                         rack->rc_rack_tmr_std_based = 1;
19562                 } else {
19563                         rack->rc_rack_tmr_std_based = 0;
19564                 }
19565                 if (optval & 0x2) {
19566                         rack->rc_rack_use_dsack = 1;
19567                 } else {
19568                         rack->rc_rack_use_dsack = 0;
19569                 }
19570                 rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19571                 break;
19572         case TCP_RACK_PACING_BETA:
19573                 RACK_OPTS_INC(tcp_rack_beta);
19574                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19575                         /* This only works for newreno. */
19576                         error = EINVAL;
19577                         break;
19578                 }
19579                 if (rack->rc_pacing_cc_set) {
19580                         /*
19581                          * Set them into the real CC module
19582                          * whats in the rack pcb is the old values
19583                          * to be used on restoral/
19584                          */
19585                         sopt.sopt_dir = SOPT_SET;
19586                         opt.name = CC_NEWRENO_BETA;
19587                         opt.val = optval;
19588                         if (CC_ALGO(tp)->ctl_output != NULL)
19589                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19590                         else {
19591                                 error = ENOENT;
19592                                 break;
19593                         }
19594                 } else {
19595                         /*
19596                          * Not pacing yet so set it into our local
19597                          * rack pcb storage.
19598                          */
19599                         rack->r_ctl.rc_saved_beta.beta = optval;
19600                 }
19601                 break;
19602         case TCP_RACK_TIMER_SLOP:
19603                 RACK_OPTS_INC(tcp_rack_timer_slop);
19604                 rack->r_ctl.timer_slop = optval;
19605                 if (rack->rc_tp->t_srtt) {
19606                         /*
19607                          * If we have an SRTT lets update t_rxtcur
19608                          * to have the new slop.
19609                          */
19610                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19611                                            rack_rto_min, rack_rto_max,
19612                                            rack->r_ctl.timer_slop);
19613                 }
19614                 break;
19615         case TCP_RACK_PACING_BETA_ECN:
19616                 RACK_OPTS_INC(tcp_rack_beta_ecn);
19617                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19618                         /* This only works for newreno. */
19619                         error = EINVAL;
19620                         break;
19621                 }
19622                 if (rack->rc_pacing_cc_set) {
19623                         /*
19624                          * Set them into the real CC module
19625                          * whats in the rack pcb is the old values
19626                          * to be used on restoral/
19627                          */
19628                         sopt.sopt_dir = SOPT_SET;
19629                         opt.name = CC_NEWRENO_BETA_ECN;
19630                         opt.val = optval;
19631                         if (CC_ALGO(tp)->ctl_output != NULL)
19632                                 error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19633                         else
19634                                 error = ENOENT;
19635                 } else {
19636                         /*
19637                          * Not pacing yet so set it into our local
19638                          * rack pcb storage.
19639                          */
19640                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19641                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19642                 }
19643                 break;
19644         case TCP_DEFER_OPTIONS:
19645                 RACK_OPTS_INC(tcp_defer_opt);
19646                 if (optval) {
19647                         if (rack->gp_ready) {
19648                                 /* Too late */
19649                                 error = EINVAL;
19650                                 break;
19651                         }
19652                         rack->defer_options = 1;
19653                 } else
19654                         rack->defer_options = 0;
19655                 break;
19656         case TCP_RACK_MEASURE_CNT:
19657                 RACK_OPTS_INC(tcp_rack_measure_cnt);
19658                 if (optval && (optval <= 0xff)) {
19659                         rack->r_ctl.req_measurements = optval;
19660                 } else
19661                         error = EINVAL;
19662                 break;
19663         case TCP_REC_ABC_VAL:
19664                 RACK_OPTS_INC(tcp_rec_abc_val);
19665                 if (optval > 0)
19666                         rack->r_use_labc_for_rec = 1;
19667                 else
19668                         rack->r_use_labc_for_rec = 0;
19669                 break;
19670         case TCP_RACK_ABC_VAL:
19671                 RACK_OPTS_INC(tcp_rack_abc_val);
19672                 if ((optval > 0) && (optval < 255))
19673                         rack->rc_labc = optval;
19674                 else
19675                         error = EINVAL;
19676                 break;
19677         case TCP_HDWR_UP_ONLY:
19678                 RACK_OPTS_INC(tcp_pacing_up_only);
19679                 if (optval)
19680                         rack->r_up_only = 1;
19681                 else
19682                         rack->r_up_only = 0;
19683                 break;
19684         case TCP_PACING_RATE_CAP:
19685                 RACK_OPTS_INC(tcp_pacing_rate_cap);
19686                 rack->r_ctl.bw_rate_cap = loptval;
19687                 break;
19688         case TCP_RACK_PROFILE:
19689                 RACK_OPTS_INC(tcp_profile);
19690                 error = rack_set_profile(rack, optval);
19691                 break;
19692         case TCP_USE_CMP_ACKS:
19693                 RACK_OPTS_INC(tcp_use_cmp_acks);
19694                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19695                         /* You can't turn it off once its on! */
19696                         error = EINVAL;
19697                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19698                         rack->r_use_cmp_ack = 1;
19699                         rack->r_mbuf_queue = 1;
19700                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19701                 }
19702                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19703                         inp->inp_flags2 |= INP_MBUF_ACKCMP;
19704                 break;
19705         case TCP_SHARED_CWND_TIME_LIMIT:
19706                 RACK_OPTS_INC(tcp_lscwnd);
19707                 if (optval)
19708                         rack->r_limit_scw = 1;
19709                 else
19710                         rack->r_limit_scw = 0;
19711                 break;
19712         case TCP_RACK_PACE_TO_FILL:
19713                 RACK_OPTS_INC(tcp_fillcw);
19714                 if (optval == 0)
19715                         rack->rc_pace_to_cwnd = 0;
19716                 else {
19717                         rack->rc_pace_to_cwnd = 1;
19718                         if (optval > 1)
19719                                 rack->r_fill_less_agg = 1;
19720                 }
19721                 if ((optval >= rack_gp_rtt_maxmul) &&
19722                     rack_gp_rtt_maxmul &&
19723                     (optval < 0xf)) {
19724                         rack->rc_pace_fill_if_rttin_range = 1;
19725                         rack->rtt_limit_mul = optval;
19726                 } else {
19727                         rack->rc_pace_fill_if_rttin_range = 0;
19728                         rack->rtt_limit_mul = 0;
19729                 }
19730                 break;
19731         case TCP_RACK_NO_PUSH_AT_MAX:
19732                 RACK_OPTS_INC(tcp_npush);
19733                 if (optval == 0)
19734                         rack->r_ctl.rc_no_push_at_mrtt = 0;
19735                 else if (optval < 0xff)
19736                         rack->r_ctl.rc_no_push_at_mrtt = optval;
19737                 else
19738                         error = EINVAL;
19739                 break;
19740         case TCP_SHARED_CWND_ENABLE:
19741                 RACK_OPTS_INC(tcp_rack_scwnd);
19742                 if (optval == 0)
19743                         rack->rack_enable_scwnd = 0;
19744                 else
19745                         rack->rack_enable_scwnd = 1;
19746                 break;
19747         case TCP_RACK_MBUF_QUEUE:
19748                 /* Now do we use the LRO mbuf-queue feature */
19749                 RACK_OPTS_INC(tcp_rack_mbufq);
19750                 if (optval || rack->r_use_cmp_ack)
19751                         rack->r_mbuf_queue = 1;
19752                 else
19753                         rack->r_mbuf_queue = 0;
19754                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19755                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19756                 else
19757                         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19758                 break;
19759         case TCP_RACK_NONRXT_CFG_RATE:
19760                 RACK_OPTS_INC(tcp_rack_cfg_rate);
19761                 if (optval == 0)
19762                         rack->rack_rec_nonrxt_use_cr = 0;
19763                 else
19764                         rack->rack_rec_nonrxt_use_cr = 1;
19765                 break;
19766         case TCP_NO_PRR:
19767                 RACK_OPTS_INC(tcp_rack_noprr);
19768                 if (optval == 0)
19769                         rack->rack_no_prr = 0;
19770                 else if (optval == 1)
19771                         rack->rack_no_prr = 1;
19772                 else if (optval == 2)
19773                         rack->no_prr_addback = 1;
19774                 else
19775                         error = EINVAL;
19776                 break;
19777         case TCP_TIMELY_DYN_ADJ:
19778                 RACK_OPTS_INC(tcp_timely_dyn);
19779                 if (optval == 0)
19780                         rack->rc_gp_dyn_mul = 0;
19781                 else {
19782                         rack->rc_gp_dyn_mul = 1;
19783                         if (optval >= 100) {
19784                                 /*
19785                                  * If the user sets something 100 or more
19786                                  * its the gp_ca value.
19787                                  */
19788                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19789                         }
19790                 }
19791                 break;
19792         case TCP_RACK_DO_DETECTION:
19793                 RACK_OPTS_INC(tcp_rack_do_detection);
19794                 if (optval == 0)
19795                         rack->do_detection = 0;
19796                 else
19797                         rack->do_detection = 1;
19798                 break;
19799         case TCP_RACK_TLP_USE:
19800                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19801                         error = EINVAL;
19802                         break;
19803                 }
19804                 RACK_OPTS_INC(tcp_tlp_use);
19805                 rack->rack_tlp_threshold_use = optval;
19806                 break;
19807         case TCP_RACK_TLP_REDUCE:
19808                 /* RACK TLP cwnd reduction (bool) */
19809                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19810                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19811                 break;
19812         /*  Pacing related ones */
19813         case TCP_RACK_PACE_ALWAYS:
19814                 /*
19815                  * zero is old rack method, 1 is new
19816                  * method using a pacing rate.
19817                  */
19818                 RACK_OPTS_INC(tcp_rack_pace_always);
19819                 if (optval > 0) {
19820                         if (rack->rc_always_pace) {
19821                                 error = EALREADY;
19822                                 break;
19823                         } else if (tcp_can_enable_pacing()) {
19824                                 rack->rc_always_pace = 1;
19825                                 if (rack->use_fixed_rate || rack->gp_ready)
19826                                         rack_set_cc_pacing(rack);
19827                         }
19828                         else {
19829                                 error = ENOSPC;
19830                                 break;
19831                         }
19832                 } else {
19833                         if (rack->rc_always_pace) {
19834                                 tcp_decrement_paced_conn();
19835                                 rack->rc_always_pace = 0;
19836                                 rack_undo_cc_pacing(rack);
19837                         }
19838                 }
19839                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19840                         inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19841                 else
19842                         inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19843                 /* A rate may be set irate or other, if so set seg size */
19844                 rack_update_seg(rack);
19845                 break;
19846         case TCP_BBR_RACK_INIT_RATE:
19847                 RACK_OPTS_INC(tcp_initial_rate);
19848                 val = optval;
19849                 /* Change from kbits per second to bytes per second */
19850                 val *= 1000;
19851                 val /= 8;
19852                 rack->r_ctl.init_rate = val;
19853                 if (rack->rc_init_win != rack_default_init_window) {
19854                         uint32_t win, snt;
19855
19856                         /*
19857                          * Options don't always get applied
19858                          * in the order you think. So in order
19859                          * to assure we update a cwnd we need
19860                          * to check and see if we are still
19861                          * where we should raise the cwnd.
19862                          */
19863                         win = rc_init_window(rack);
19864                         if (SEQ_GT(tp->snd_max, tp->iss))
19865                                 snt = tp->snd_max - tp->iss;
19866                         else
19867                                 snt = 0;
19868                         if ((snt < win) &&
19869                             (tp->snd_cwnd < win))
19870                                 tp->snd_cwnd = win;
19871                 }
19872                 if (rack->rc_always_pace)
19873                         rack_update_seg(rack);
19874                 break;
19875         case TCP_BBR_IWINTSO:
19876                 RACK_OPTS_INC(tcp_initial_win);
19877                 if (optval && (optval <= 0xff)) {
19878                         uint32_t win, snt;
19879
19880                         rack->rc_init_win = optval;
19881                         win = rc_init_window(rack);
19882                         if (SEQ_GT(tp->snd_max, tp->iss))
19883                                 snt = tp->snd_max - tp->iss;
19884                         else
19885                                 snt = 0;
19886                         if ((snt < win) &&
19887                             (tp->t_srtt |
19888 #ifdef NETFLIX_PEAKRATE
19889                              tp->t_maxpeakrate |
19890 #endif
19891                              rack->r_ctl.init_rate)) {
19892                                 /*
19893                                  * We are not past the initial window
19894                                  * and we have some bases for pacing,
19895                                  * so we need to possibly adjust up
19896                                  * the cwnd. Note even if we don't set
19897                                  * the cwnd, its still ok to raise the rc_init_win
19898                                  * which can be used coming out of idle when we
19899                                  * would have a rate.
19900                                  */
19901                                 if (tp->snd_cwnd < win)
19902                                         tp->snd_cwnd = win;
19903                         }
19904                         if (rack->rc_always_pace)
19905                                 rack_update_seg(rack);
19906                 } else
19907                         error = EINVAL;
19908                 break;
19909         case TCP_RACK_FORCE_MSEG:
19910                 RACK_OPTS_INC(tcp_rack_force_max_seg);
19911                 if (optval)
19912                         rack->rc_force_max_seg = 1;
19913                 else
19914                         rack->rc_force_max_seg = 0;
19915                 break;
19916         case TCP_RACK_PACE_MAX_SEG:
19917                 /* Max segments size in a pace in bytes */
19918                 RACK_OPTS_INC(tcp_rack_max_seg);
19919                 rack->rc_user_set_max_segs = optval;
19920                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19921                 break;
19922         case TCP_RACK_PACE_RATE_REC:
19923                 /* Set the fixed pacing rate in Bytes per second ca */
19924                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19925                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19926                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19927                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19928                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19929                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19930                 rack->use_fixed_rate = 1;
19931                 if (rack->rc_always_pace)
19932                         rack_set_cc_pacing(rack);
19933                 rack_log_pacing_delay_calc(rack,
19934                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19935                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19936                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19937                                            __LINE__, NULL,0);
19938                 break;
19939
19940         case TCP_RACK_PACE_RATE_SS:
19941                 /* Set the fixed pacing rate in Bytes per second ca */
19942                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19943                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19944                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19945                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19946                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19947                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19948                 rack->use_fixed_rate = 1;
19949                 if (rack->rc_always_pace)
19950                         rack_set_cc_pacing(rack);
19951                 rack_log_pacing_delay_calc(rack,
19952                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19953                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19954                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19955                                            __LINE__, NULL, 0);
19956                 break;
19957
19958         case TCP_RACK_PACE_RATE_CA:
19959                 /* Set the fixed pacing rate in Bytes per second ca */
19960                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19961                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19962                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19963                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19964                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19965                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19966                 rack->use_fixed_rate = 1;
19967                 if (rack->rc_always_pace)
19968                         rack_set_cc_pacing(rack);
19969                 rack_log_pacing_delay_calc(rack,
19970                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19971                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19972                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19973                                            __LINE__, NULL, 0);
19974                 break;
19975         case TCP_RACK_GP_INCREASE_REC:
19976                 RACK_OPTS_INC(tcp_gp_inc_rec);
19977                 rack->r_ctl.rack_per_of_gp_rec = optval;
19978                 rack_log_pacing_delay_calc(rack,
19979                                            rack->r_ctl.rack_per_of_gp_ss,
19980                                            rack->r_ctl.rack_per_of_gp_ca,
19981                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19982                                            __LINE__, NULL, 0);
19983                 break;
19984         case TCP_RACK_GP_INCREASE_CA:
19985                 RACK_OPTS_INC(tcp_gp_inc_ca);
19986                 ca = optval;
19987                 if (ca < 100) {
19988                         /*
19989                          * We don't allow any reduction
19990                          * over the GP b/w.
19991                          */
19992                         error = EINVAL;
19993                         break;
19994                 }
19995                 rack->r_ctl.rack_per_of_gp_ca = ca;
19996                 rack_log_pacing_delay_calc(rack,
19997                                            rack->r_ctl.rack_per_of_gp_ss,
19998                                            rack->r_ctl.rack_per_of_gp_ca,
19999                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20000                                            __LINE__, NULL, 0);
20001                 break;
20002         case TCP_RACK_GP_INCREASE_SS:
20003                 RACK_OPTS_INC(tcp_gp_inc_ss);
20004                 ss = optval;
20005                 if (ss < 100) {
20006                         /*
20007                          * We don't allow any reduction
20008                          * over the GP b/w.
20009                          */
20010                         error = EINVAL;
20011                         break;
20012                 }
20013                 rack->r_ctl.rack_per_of_gp_ss = ss;
20014                 rack_log_pacing_delay_calc(rack,
20015                                            rack->r_ctl.rack_per_of_gp_ss,
20016                                            rack->r_ctl.rack_per_of_gp_ca,
20017                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20018                                            __LINE__, NULL, 0);
20019                 break;
20020         case TCP_RACK_RR_CONF:
20021                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
20022                 if (optval && optval <= 3)
20023                         rack->r_rr_config = optval;
20024                 else
20025                         rack->r_rr_config = 0;
20026                 break;
20027         case TCP_HDWR_RATE_CAP:
20028                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
20029                 if (optval) {
20030                         if (rack->r_rack_hw_rate_caps == 0)
20031                                 rack->r_rack_hw_rate_caps = 1;
20032                         else
20033                                 error = EALREADY;
20034                 } else {
20035                         rack->r_rack_hw_rate_caps = 0;
20036                 }
20037                 break;
20038         case TCP_BBR_HDWR_PACE:
20039                 RACK_OPTS_INC(tcp_hdwr_pacing);
20040                 if (optval){
20041                         if (rack->rack_hdrw_pacing == 0) {
20042                                 rack->rack_hdw_pace_ena = 1;
20043                                 rack->rack_attempt_hdwr_pace = 0;
20044                         } else
20045                                 error = EALREADY;
20046                 } else {
20047                         rack->rack_hdw_pace_ena = 0;
20048 #ifdef RATELIMIT
20049                         if (rack->r_ctl.crte != NULL) {
20050                                 rack->rack_hdrw_pacing = 0;
20051                                 rack->rack_attempt_hdwr_pace = 0;
20052                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
20053                                 rack->r_ctl.crte = NULL;
20054                         }
20055 #endif
20056                 }
20057                 break;
20058         /*  End Pacing related ones */
20059         case TCP_RACK_PRR_SENDALOT:
20060                 /* Allow PRR to send more than one seg */
20061                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
20062                 rack->r_ctl.rc_prr_sendalot = optval;
20063                 break;
20064         case TCP_RACK_MIN_TO:
20065                 /* Minimum time between rack t-o's in ms */
20066                 RACK_OPTS_INC(tcp_rack_min_to);
20067                 rack->r_ctl.rc_min_to = optval;
20068                 break;
20069         case TCP_RACK_EARLY_SEG:
20070                 /* If early recovery max segments */
20071                 RACK_OPTS_INC(tcp_rack_early_seg);
20072                 rack->r_ctl.rc_early_recovery_segs = optval;
20073                 break;
20074         case TCP_RACK_ENABLE_HYSTART:
20075         {
20076                 if (optval) {
20077                         tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
20078                         if (rack_do_hystart > RACK_HYSTART_ON)
20079                                 tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
20080                         if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
20081                                 tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
20082                 } else {
20083                         tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
20084                 }
20085         }
20086         break;
20087         case TCP_RACK_REORD_THRESH:
20088                 /* RACK reorder threshold (shift amount) */
20089                 RACK_OPTS_INC(tcp_rack_reord_thresh);
20090                 if ((optval > 0) && (optval < 31))
20091                         rack->r_ctl.rc_reorder_shift = optval;
20092                 else
20093                         error = EINVAL;
20094                 break;
20095         case TCP_RACK_REORD_FADE:
20096                 /* Does reordering fade after ms time */
20097                 RACK_OPTS_INC(tcp_rack_reord_fade);
20098                 rack->r_ctl.rc_reorder_fade = optval;
20099                 break;
20100         case TCP_RACK_TLP_THRESH:
20101                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
20102                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
20103                 if (optval)
20104                         rack->r_ctl.rc_tlp_threshold = optval;
20105                 else
20106                         error = EINVAL;
20107                 break;
20108         case TCP_BBR_USE_RACK_RR:
20109                 RACK_OPTS_INC(tcp_rack_rr);
20110                 if (optval)
20111                         rack->use_rack_rr = 1;
20112                 else
20113                         rack->use_rack_rr = 0;
20114                 break;
20115         case TCP_FAST_RSM_HACK:
20116                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
20117                 if (optval)
20118                         rack->fast_rsm_hack = 1;
20119                 else
20120                         rack->fast_rsm_hack = 0;
20121                 break;
20122         case TCP_RACK_PKT_DELAY:
20123                 /* RACK added ms i.e. rack-rtt + reord + N */
20124                 RACK_OPTS_INC(tcp_rack_pkt_delay);
20125                 rack->r_ctl.rc_pkt_delay = optval;
20126                 break;
20127         case TCP_DELACK:
20128                 RACK_OPTS_INC(tcp_rack_delayed_ack);
20129                 if (optval == 0)
20130                         tp->t_delayed_ack = 0;
20131                 else
20132                         tp->t_delayed_ack = 1;
20133                 if (tp->t_flags & TF_DELACK) {
20134                         tp->t_flags &= ~TF_DELACK;
20135                         tp->t_flags |= TF_ACKNOW;
20136                         NET_EPOCH_ENTER(et);
20137                         rack_output(tp);
20138                         NET_EPOCH_EXIT(et);
20139                 }
20140                 break;
20141
20142         case TCP_BBR_RACK_RTT_USE:
20143                 RACK_OPTS_INC(tcp_rack_rtt_use);
20144                 if ((optval != USE_RTT_HIGH) &&
20145                     (optval != USE_RTT_LOW) &&
20146                     (optval != USE_RTT_AVG))
20147                         error = EINVAL;
20148                 else
20149                         rack->r_ctl.rc_rate_sample_method = optval;
20150                 break;
20151         case TCP_DATA_AFTER_CLOSE:
20152                 RACK_OPTS_INC(tcp_data_after_close);
20153                 if (optval)
20154                         rack->rc_allow_data_af_clo = 1;
20155                 else
20156                         rack->rc_allow_data_af_clo = 0;
20157                 break;
20158         default:
20159                 break;
20160         }
20161 #ifdef NETFLIX_STATS
20162         tcp_log_socket_option(tp, sopt_name, optval, error);
20163 #endif
20164         return (error);
20165 }
20166
20167
20168 static void
20169 rack_apply_deferred_options(struct tcp_rack *rack)
20170 {
20171         struct deferred_opt_list *dol, *sdol;
20172         uint32_t s_optval;
20173
20174         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20175                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20176                 /* Disadvantage of deferal is you loose the error return */
20177                 s_optval = (uint32_t)dol->optval;
20178                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20179                 free(dol, M_TCPDO);
20180         }
20181 }
20182
20183 static void
20184 rack_hw_tls_change(struct tcpcb *tp, int chg)
20185 {
20186         /* Update HW tls state */
20187         struct tcp_rack *rack;
20188
20189         rack = (struct tcp_rack *)tp->t_fb_ptr;
20190         if (chg)
20191                 rack->r_ctl.fsb.hw_tls = 1;
20192         else
20193                 rack->r_ctl.fsb.hw_tls = 0;
20194 }
20195
20196 static int
20197 rack_pru_options(struct tcpcb *tp, int flags)
20198 {
20199         if (flags & PRUS_OOB)
20200                 return (EOPNOTSUPP);
20201         return (0);
20202 }
20203
20204 static struct tcp_function_block __tcp_rack = {
20205         .tfb_tcp_block_name = __XSTRING(STACKNAME),
20206         .tfb_tcp_output = rack_output,
20207         .tfb_do_queued_segments = ctf_do_queued_segments,
20208         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
20209         .tfb_tcp_do_segment = rack_do_segment,
20210         .tfb_tcp_ctloutput = rack_ctloutput,
20211         .tfb_tcp_fb_init = rack_init,
20212         .tfb_tcp_fb_fini = rack_fini,
20213         .tfb_tcp_timer_stop_all = rack_stopall,
20214         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20215         .tfb_tcp_handoff_ok = rack_handoff_ok,
20216         .tfb_tcp_mtu_chg = rack_mtu_change,
20217         .tfb_pru_options = rack_pru_options,
20218         .tfb_hwtls_change = rack_hw_tls_change,
20219         .tfb_compute_pipe = rack_compute_pipe,
20220         .tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20221 };
20222
20223 /*
20224  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20225  * socket option arguments.  When it re-acquires the lock after the copy, it
20226  * has to revalidate that the connection is still valid for the socket
20227  * option.
20228  */
20229 static int
20230 rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
20231 {
20232 #ifdef INET6
20233         struct ip6_hdr *ip6;
20234 #endif
20235 #ifdef INET
20236         struct ip *ip;
20237 #endif
20238         struct tcpcb *tp;
20239         struct tcp_rack *rack;
20240         uint64_t loptval;
20241         int32_t error = 0, optval;
20242
20243         tp = intotcpcb(inp);
20244         rack = (struct tcp_rack *)tp->t_fb_ptr;
20245         if (rack == NULL) {
20246                 INP_WUNLOCK(inp);
20247                 return (EINVAL);
20248         }
20249 #ifdef INET6
20250         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20251 #endif
20252 #ifdef INET
20253         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20254 #endif
20255
20256         switch (sopt->sopt_level) {
20257 #ifdef INET6
20258         case IPPROTO_IPV6:
20259                 MPASS(inp->inp_vflag & INP_IPV6PROTO);
20260                 switch (sopt->sopt_name) {
20261                 case IPV6_USE_MIN_MTU:
20262                         tcp6_use_min_mtu(tp);
20263                         break;
20264                 case IPV6_TCLASS:
20265                         /*
20266                          * The DSCP codepoint has changed, update the fsb.
20267                          */
20268                         ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20269                             (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20270                         break;
20271                 }
20272                 INP_WUNLOCK(inp);
20273                 return (0);
20274 #endif
20275 #ifdef INET
20276         case IPPROTO_IP:
20277                 switch (sopt->sopt_name) {
20278                 case IP_TOS:
20279                         /*
20280                          * The DSCP codepoint has changed, update the fsb.
20281                          */
20282                         ip->ip_tos = rack->rc_inp->inp_ip_tos;
20283                         break;
20284                 case IP_TTL:
20285                         /*
20286                          * The TTL has changed, update the fsb.
20287                          */
20288                         ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20289                         break;
20290                 }
20291                 INP_WUNLOCK(inp);
20292                 return (0);
20293 #endif
20294         }
20295
20296         switch (sopt->sopt_name) {
20297         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
20298         /*  Pacing related ones */
20299         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
20300         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
20301         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
20302         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
20303         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
20304         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
20305         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
20306         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
20307         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
20308         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
20309         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
20310         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
20311         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
20312         case TCP_HDWR_RATE_CAP:                 /*  URL:hdwrcap boolean */
20313         case TCP_PACING_RATE_CAP:               /*  URL:cap  -- used by side-channel */
20314         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
20315        /* End pacing related */
20316         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
20317         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20318         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
20319         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
20320         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
20321         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
20322         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
20323         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
20324         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
20325         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
20326         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
20327         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
20328         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
20329         case TCP_NO_PRR:                        /*  URL:noprr */
20330         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
20331         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
20332         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
20333         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
20334         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
20335         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
20336         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
20337         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
20338         case TCP_RACK_PROFILE:                  /*  URL:profile */
20339         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
20340         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
20341         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
20342         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
20343         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
20344         case TCP_RACK_DSACK_OPT:                /*  URL:dsack */
20345         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
20346         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
20347         case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
20348         case TCP_RACK_ENABLE_HYSTART:           /*  URL:hystart */
20349                 break;
20350         default:
20351                 /* Filter off all unknown options to the base stack */
20352                 return (tcp_default_ctloutput(inp, sopt));
20353                 break;
20354         }
20355         INP_WUNLOCK(inp);
20356         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20357                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20358                 /*
20359                  * We truncate it down to 32 bits for the socket-option trace this
20360                  * means rates > 34Gbps won't show right, but thats probably ok.
20361                  */
20362                 optval = (uint32_t)loptval;
20363         } else {
20364                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20365                 /* Save it in 64 bit form too */
20366                 loptval = optval;
20367         }
20368         if (error)
20369                 return (error);
20370         INP_WLOCK(inp);
20371         if (inp->inp_flags & INP_DROPPED) {
20372                 INP_WUNLOCK(inp);
20373                 return (ECONNRESET);
20374         }
20375         if (tp->t_fb != &__tcp_rack) {
20376                 INP_WUNLOCK(inp);
20377                 return (ENOPROTOOPT);
20378         }
20379         if (rack->defer_options && (rack->gp_ready == 0) &&
20380             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20381             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20382             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20383             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20384                 /* Options are beind deferred */
20385                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20386                         INP_WUNLOCK(inp);
20387                         return (0);
20388                 } else {
20389                         /* No memory to defer, fail */
20390                         INP_WUNLOCK(inp);
20391                         return (ENOMEM);
20392                 }
20393         }
20394         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20395         INP_WUNLOCK(inp);
20396         return (error);
20397 }
20398
20399 static void
20400 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20401 {
20402
20403         INP_WLOCK_ASSERT(tptoinpcb(tp));
20404         bzero(ti, sizeof(*ti));
20405
20406         ti->tcpi_state = tp->t_state;
20407         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20408                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20409         if (tp->t_flags & TF_SACK_PERMIT)
20410                 ti->tcpi_options |= TCPI_OPT_SACK;
20411         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20412                 ti->tcpi_options |= TCPI_OPT_WSCALE;
20413                 ti->tcpi_snd_wscale = tp->snd_scale;
20414                 ti->tcpi_rcv_wscale = tp->rcv_scale;
20415         }
20416         if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
20417                 ti->tcpi_options |= TCPI_OPT_ECN;
20418         if (tp->t_flags & TF_FASTOPEN)
20419                 ti->tcpi_options |= TCPI_OPT_TFO;
20420         /* still kept in ticks is t_rcvtime */
20421         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20422         /* Since we hold everything in precise useconds this is easy */
20423         ti->tcpi_rtt = tp->t_srtt;
20424         ti->tcpi_rttvar = tp->t_rttvar;
20425         ti->tcpi_rto = tp->t_rxtcur;
20426         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20427         ti->tcpi_snd_cwnd = tp->snd_cwnd;
20428         /*
20429          * FreeBSD-specific extension fields for tcp_info.
20430          */
20431         ti->tcpi_rcv_space = tp->rcv_wnd;
20432         ti->tcpi_rcv_nxt = tp->rcv_nxt;
20433         ti->tcpi_snd_wnd = tp->snd_wnd;
20434         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
20435         ti->tcpi_snd_nxt = tp->snd_nxt;
20436         ti->tcpi_snd_mss = tp->t_maxseg;
20437         ti->tcpi_rcv_mss = tp->t_maxseg;
20438         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20439         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20440         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20441 #ifdef NETFLIX_STATS
20442         ti->tcpi_total_tlp = tp->t_sndtlppack;
20443         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20444         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20445 #endif
20446 #ifdef TCP_OFFLOAD
20447         if (tp->t_flags & TF_TOE) {
20448                 ti->tcpi_options |= TCPI_OPT_TOE;
20449                 tcp_offload_tcp_info(tp, ti);
20450         }
20451 #endif
20452 }
20453
20454 static int
20455 rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
20456 {
20457         struct tcpcb *tp;
20458         struct tcp_rack *rack;
20459         int32_t error, optval;
20460         uint64_t val, loptval;
20461         struct  tcp_info ti;
20462         /*
20463          * Because all our options are either boolean or an int, we can just
20464          * pull everything into optval and then unlock and copy. If we ever
20465          * add a option that is not a int, then this will have quite an
20466          * impact to this routine.
20467          */
20468         error = 0;
20469         tp = intotcpcb(inp);
20470         rack = (struct tcp_rack *)tp->t_fb_ptr;
20471         if (rack == NULL) {
20472                 INP_WUNLOCK(inp);
20473                 return (EINVAL);
20474         }
20475         switch (sopt->sopt_name) {
20476         case TCP_INFO:
20477                 /* First get the info filled */
20478                 rack_fill_info(tp, &ti);
20479                 /* Fix up the rtt related fields if needed */
20480                 INP_WUNLOCK(inp);
20481                 error = sooptcopyout(sopt, &ti, sizeof ti);
20482                 return (error);
20483         /*
20484          * Beta is the congestion control value for NewReno that influences how
20485          * much of a backoff happens when loss is detected. It is normally set
20486          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20487          * when you exit recovery.
20488          */
20489         case TCP_RACK_PACING_BETA:
20490                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20491                         error = EINVAL;
20492                 else if (rack->rc_pacing_cc_set == 0)
20493                         optval = rack->r_ctl.rc_saved_beta.beta;
20494                 else {
20495                         /*
20496                          * Reach out into the CC data and report back what
20497                          * I have previously set. Yeah it looks hackish but
20498                          * we don't want to report the saved values.
20499                          */
20500                         if (tp->t_ccv.cc_data)
20501                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
20502                         else
20503                                 error = EINVAL;
20504                 }
20505                 break;
20506                 /*
20507                  * Beta_ecn is the congestion control value for NewReno that influences how
20508                  * much of a backoff happens when a ECN mark is detected. It is normally set
20509                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20510                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
20511                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
20512                  */
20513
20514         case TCP_RACK_PACING_BETA_ECN:
20515                 if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20516                         error = EINVAL;
20517                 else if (rack->rc_pacing_cc_set == 0)
20518                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20519                 else {
20520                         /*
20521                          * Reach out into the CC data and report back what
20522                          * I have previously set. Yeah it looks hackish but
20523                          * we don't want to report the saved values.
20524                          */
20525                         if (tp->t_ccv.cc_data)
20526                                 optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
20527                         else
20528                                 error = EINVAL;
20529                 }
20530                 break;
20531         case TCP_RACK_DSACK_OPT:
20532                 optval = 0;
20533                 if (rack->rc_rack_tmr_std_based) {
20534                         optval |= 1;
20535                 }
20536                 if (rack->rc_rack_use_dsack) {
20537                         optval |= 2;
20538                 }
20539                 break;
20540         case TCP_RACK_ENABLE_HYSTART:
20541         {
20542                 if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
20543                         optval = RACK_HYSTART_ON;
20544                         if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
20545                                 optval = RACK_HYSTART_ON_W_SC;
20546                         if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
20547                                 optval = RACK_HYSTART_ON_W_SC_C;
20548                 } else {
20549                         optval = RACK_HYSTART_OFF;
20550                 }
20551         }
20552         break;
20553         case TCP_FAST_RSM_HACK:
20554                 optval = rack->fast_rsm_hack;
20555                 break;
20556         case TCP_DEFER_OPTIONS:
20557                 optval = rack->defer_options;
20558                 break;
20559         case TCP_RACK_MEASURE_CNT:
20560                 optval = rack->r_ctl.req_measurements;
20561                 break;
20562         case TCP_REC_ABC_VAL:
20563                 optval = rack->r_use_labc_for_rec;
20564                 break;
20565         case TCP_RACK_ABC_VAL:
20566                 optval = rack->rc_labc;
20567                 break;
20568         case TCP_HDWR_UP_ONLY:
20569                 optval= rack->r_up_only;
20570                 break;
20571         case TCP_PACING_RATE_CAP:
20572                 loptval = rack->r_ctl.bw_rate_cap;
20573                 break;
20574         case TCP_RACK_PROFILE:
20575                 /* You cannot retrieve a profile, its write only */
20576                 error = EINVAL;
20577                 break;
20578         case TCP_USE_CMP_ACKS:
20579                 optval = rack->r_use_cmp_ack;
20580                 break;
20581         case TCP_RACK_PACE_TO_FILL:
20582                 optval = rack->rc_pace_to_cwnd;
20583                 if (optval && rack->r_fill_less_agg)
20584                         optval++;
20585                 break;
20586         case TCP_RACK_NO_PUSH_AT_MAX:
20587                 optval = rack->r_ctl.rc_no_push_at_mrtt;
20588                 break;
20589         case TCP_SHARED_CWND_ENABLE:
20590                 optval = rack->rack_enable_scwnd;
20591                 break;
20592         case TCP_RACK_NONRXT_CFG_RATE:
20593                 optval = rack->rack_rec_nonrxt_use_cr;
20594                 break;
20595         case TCP_NO_PRR:
20596                 if (rack->rack_no_prr  == 1)
20597                         optval = 1;
20598                 else if (rack->no_prr_addback == 1)
20599                         optval = 2;
20600                 else
20601                         optval = 0;
20602                 break;
20603         case TCP_RACK_DO_DETECTION:
20604                 optval = rack->do_detection;
20605                 break;
20606         case TCP_RACK_MBUF_QUEUE:
20607                 /* Now do we use the LRO mbuf-queue feature */
20608                 optval = rack->r_mbuf_queue;
20609                 break;
20610         case TCP_TIMELY_DYN_ADJ:
20611                 optval = rack->rc_gp_dyn_mul;
20612                 break;
20613         case TCP_BBR_IWINTSO:
20614                 optval = rack->rc_init_win;
20615                 break;
20616         case TCP_RACK_TLP_REDUCE:
20617                 /* RACK TLP cwnd reduction (bool) */
20618                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20619                 break;
20620         case TCP_BBR_RACK_INIT_RATE:
20621                 val = rack->r_ctl.init_rate;
20622                 /* convert to kbits per sec */
20623                 val *= 8;
20624                 val /= 1000;
20625                 optval = (uint32_t)val;
20626                 break;
20627         case TCP_RACK_FORCE_MSEG:
20628                 optval = rack->rc_force_max_seg;
20629                 break;
20630         case TCP_RACK_PACE_MAX_SEG:
20631                 /* Max segments in a pace */
20632                 optval = rack->rc_user_set_max_segs;
20633                 break;
20634         case TCP_RACK_PACE_ALWAYS:
20635                 /* Use the always pace method */
20636                 optval = rack->rc_always_pace;
20637                 break;
20638         case TCP_RACK_PRR_SENDALOT:
20639                 /* Allow PRR to send more than one seg */
20640                 optval = rack->r_ctl.rc_prr_sendalot;
20641                 break;
20642         case TCP_RACK_MIN_TO:
20643                 /* Minimum time between rack t-o's in ms */
20644                 optval = rack->r_ctl.rc_min_to;
20645                 break;
20646         case TCP_RACK_EARLY_SEG:
20647                 /* If early recovery max segments */
20648                 optval = rack->r_ctl.rc_early_recovery_segs;
20649                 break;
20650         case TCP_RACK_REORD_THRESH:
20651                 /* RACK reorder threshold (shift amount) */
20652                 optval = rack->r_ctl.rc_reorder_shift;
20653                 break;
20654         case TCP_RACK_REORD_FADE:
20655                 /* Does reordering fade after ms time */
20656                 optval = rack->r_ctl.rc_reorder_fade;
20657                 break;
20658         case TCP_BBR_USE_RACK_RR:
20659                 /* Do we use the rack cheat for rxt */
20660                 optval = rack->use_rack_rr;
20661                 break;
20662         case TCP_RACK_RR_CONF:
20663                 optval = rack->r_rr_config;
20664                 break;
20665         case TCP_HDWR_RATE_CAP:
20666                 optval = rack->r_rack_hw_rate_caps;
20667                 break;
20668         case TCP_BBR_HDWR_PACE:
20669                 optval = rack->rack_hdw_pace_ena;
20670                 break;
20671         case TCP_RACK_TLP_THRESH:
20672                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
20673                 optval = rack->r_ctl.rc_tlp_threshold;
20674                 break;
20675         case TCP_RACK_PKT_DELAY:
20676                 /* RACK added ms i.e. rack-rtt + reord + N */
20677                 optval = rack->r_ctl.rc_pkt_delay;
20678                 break;
20679         case TCP_RACK_TLP_USE:
20680                 optval = rack->rack_tlp_threshold_use;
20681                 break;
20682         case TCP_RACK_PACE_RATE_CA:
20683                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20684                 break;
20685         case TCP_RACK_PACE_RATE_SS:
20686                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20687                 break;
20688         case TCP_RACK_PACE_RATE_REC:
20689                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20690                 break;
20691         case TCP_RACK_GP_INCREASE_SS:
20692                 optval = rack->r_ctl.rack_per_of_gp_ca;
20693                 break;
20694         case TCP_RACK_GP_INCREASE_CA:
20695                 optval = rack->r_ctl.rack_per_of_gp_ss;
20696                 break;
20697         case TCP_BBR_RACK_RTT_USE:
20698                 optval = rack->r_ctl.rc_rate_sample_method;
20699                 break;
20700         case TCP_DELACK:
20701                 optval = tp->t_delayed_ack;
20702                 break;
20703         case TCP_DATA_AFTER_CLOSE:
20704                 optval = rack->rc_allow_data_af_clo;
20705                 break;
20706         case TCP_SHARED_CWND_TIME_LIMIT:
20707                 optval = rack->r_limit_scw;
20708                 break;
20709         case TCP_RACK_TIMER_SLOP:
20710                 optval = rack->r_ctl.timer_slop;
20711                 break;
20712         default:
20713                 return (tcp_default_ctloutput(inp, sopt));
20714                 break;
20715         }
20716         INP_WUNLOCK(inp);
20717         if (error == 0) {
20718                 if (TCP_PACING_RATE_CAP)
20719                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
20720                 else
20721                         error = sooptcopyout(sopt, &optval, sizeof optval);
20722         }
20723         return (error);
20724 }
20725
20726 static int
20727 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt)
20728 {
20729         if (sopt->sopt_dir == SOPT_SET) {
20730                 return (rack_set_sockopt(inp, sopt));
20731         } else if (sopt->sopt_dir == SOPT_GET) {
20732                 return (rack_get_sockopt(inp, sopt));
20733         } else {
20734                 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
20735         }
20736 }
20737
20738 static const char *rack_stack_names[] = {
20739         __XSTRING(STACKNAME),
20740 #ifdef STACKALIAS
20741         __XSTRING(STACKALIAS),
20742 #endif
20743 };
20744
20745 static int
20746 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20747 {
20748         memset(mem, 0, size);
20749         return (0);
20750 }
20751
20752 static void
20753 rack_dtor(void *mem, int32_t size, void *arg)
20754 {
20755
20756 }
20757
20758 static bool rack_mod_inited = false;
20759
20760 static int
20761 tcp_addrack(module_t mod, int32_t type, void *data)
20762 {
20763         int32_t err = 0;
20764         int num_stacks;
20765
20766         switch (type) {
20767         case MOD_LOAD:
20768                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20769                     sizeof(struct rack_sendmap),
20770                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20771
20772                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20773                     sizeof(struct tcp_rack),
20774                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20775
20776                 sysctl_ctx_init(&rack_sysctl_ctx);
20777                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20778                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20779                     OID_AUTO,
20780 #ifdef STACKALIAS
20781                     __XSTRING(STACKALIAS),
20782 #else
20783                     __XSTRING(STACKNAME),
20784 #endif
20785                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20786                     "");
20787                 if (rack_sysctl_root == NULL) {
20788                         printf("Failed to add sysctl node\n");
20789                         err = EFAULT;
20790                         goto free_uma;
20791                 }
20792                 rack_init_sysctls();
20793                 num_stacks = nitems(rack_stack_names);
20794                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20795                     rack_stack_names, &num_stacks);
20796                 if (err) {
20797                         printf("Failed to register %s stack name for "
20798                             "%s module\n", rack_stack_names[num_stacks],
20799                             __XSTRING(MODNAME));
20800                         sysctl_ctx_free(&rack_sysctl_ctx);
20801 free_uma:
20802                         uma_zdestroy(rack_zone);
20803                         uma_zdestroy(rack_pcb_zone);
20804                         rack_counter_destroy();
20805                         printf("Failed to register rack module -- err:%d\n", err);
20806                         return (err);
20807                 }
20808                 tcp_lro_reg_mbufq();
20809                 rack_mod_inited = true;
20810                 break;
20811         case MOD_QUIESCE:
20812                 err = deregister_tcp_functions(&__tcp_rack, true, false);
20813                 break;
20814         case MOD_UNLOAD:
20815                 err = deregister_tcp_functions(&__tcp_rack, false, true);
20816                 if (err == EBUSY)
20817                         break;
20818                 if (rack_mod_inited) {
20819                         uma_zdestroy(rack_zone);
20820                         uma_zdestroy(rack_pcb_zone);
20821                         sysctl_ctx_free(&rack_sysctl_ctx);
20822                         rack_counter_destroy();
20823                         rack_mod_inited = false;
20824                 }
20825                 tcp_lro_dereg_mbufq();
20826                 err = 0;
20827                 break;
20828         default:
20829                 return (EOPNOTSUPP);
20830         }
20831         return (err);
20832 }
20833
20834 static moduledata_t tcp_rack = {
20835         .name = __XSTRING(MODNAME),
20836         .evhand = tcp_addrack,
20837         .priv = 0
20838 };
20839
20840 MODULE_VERSION(MODNAME, 1);
20841 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20842 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
20843
20844 #endif /* #if !defined(INET) && !defined(INET6) */