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_tcpdebug.h"
  34 #include "opt_ratelimit.h"
  35 #include "opt_kern_tls.h"
  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 TCPDEBUG
 109 #include <netinet/tcp_debug.h>
 110 #endif                          /* TCPDEBUG */
 111 #ifdef TCP_OFFLOAD
 112 #include <netinet/tcp_offload.h>
 113 #endif
 114 #ifdef INET6
 115 #include <netinet6/tcp6_var.h>
 116 #endif
 117 #include <netinet/tcp_ecn.h>
 118
 119 #include <netipsec/ipsec_support.h>
 120
 121 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 122 #include <netipsec/ipsec.h>
 123 #include <netipsec/ipsec6.h>
 124 #endif                          /* IPSEC */
 125
 126 #include <netinet/udp.h>
 127 #include <netinet/udp_var.h>
 128 #include <machine/in_cksum.h>
 129
 130 #ifdef MAC
 131 #include <security/mac/mac_framework.h>
 132 #endif
 133 #include "sack_filter.h"
 134 #include "tcp_rack.h"
 135 #include "rack_bbr_common.h"
 136
 137 uma_zone_t rack_zone;
 138 uma_zone_t rack_pcb_zone;
 139
 140 #ifndef TICKS2SBT
 141 #define TICKS2SBT(__t)  (tick_sbt * ((sbintime_t)(__t)))
 142 #endif
 143
 144 VNET_DECLARE(uint32_t, newreno_beta);
 145 VNET_DECLARE(uint32_t, newreno_beta_ecn);
 146 #define V_newreno_beta VNET(newreno_beta)
 147 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
 148
 149
 150 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
 151 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
 152
 153 struct sysctl_ctx_list rack_sysctl_ctx;
 154 struct sysctl_oid *rack_sysctl_root;
 155
 156 #define CUM_ACKED 1
 157 #define SACKED 2
 158
 159 /*
 160  * The RACK module incorporates a number of
 161  * TCP ideas that have been put out into the IETF
 162  * over the last few years:
 163  * - Matt Mathis's Rate Halving which slowly drops
 164  *    the congestion window so that the ack clock can
 165  *    be maintained during a recovery.
 166  * - Yuchung Cheng's RACK TCP (for which its named) that
 167  *    will stop us using the number of dup acks and instead
 168  *    use time as the gage of when we retransmit.
 169  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 170  *    of Dukkipati et.al.
 171  * RACK depends on SACK, so if an endpoint arrives that
 172  * cannot do SACK the state machine below will shuttle the
 173  * connection back to using the "default" TCP stack that is
 174  * in FreeBSD.
 175  *
 176  * To implement RACK the original TCP stack was first decomposed
 177  * into a functional state machine with individual states
 178  * for each of the possible TCP connection states. The do_segment
 179  * functions role in life is to mandate the connection supports SACK
 180  * initially and then assure that the RACK state matches the conenction
 181  * state before calling the states do_segment function. Each
 182  * state is simplified due to the fact that the original do_segment
 183  * has been decomposed and we *know* what state we are in (no
 184  * switches on the state) and all tests for SACK are gone. This
 185  * greatly simplifies what each state does.
 186  *
 187  * TCP output is also over-written with a new version since it
 188  * must maintain the new rack scoreboard.
 189  *
 190  */
 191 static int32_t rack_tlp_thresh = 1;
 192 static int32_t rack_tlp_limit = 2;      /* No more than 2 TLPs w-out new data */
 193 static int32_t rack_tlp_use_greater = 1;
 194 static int32_t rack_reorder_thresh = 2;
 195 static int32_t rack_reorder_fade = 60000000;    /* 0 - never fade, def 60,000,000
 196                                                  * - 60 seconds */
 197 static uint8_t rack_req_measurements = 1;
 198 /* Attack threshold detections */
 199 static uint32_t rack_highest_sack_thresh_seen = 0;
 200 static uint32_t rack_highest_move_thresh_seen = 0;
 201 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
 202 static int32_t rack_hw_pace_extra_slots = 2;    /* 2 extra MSS time betweens */
 203 static int32_t rack_hw_rate_caps = 1; /* 1; */
 204 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
 205 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
 206 static int32_t rack_hw_up_only = 1;
 207 static int32_t rack_stats_gets_ms_rtt = 1;
 208 static int32_t rack_prr_addbackmax = 2;
 209 static int32_t rack_do_hystart = 0;
 210 static int32_t rack_apply_rtt_with_reduced_conf = 0;
 211
 212 static int32_t rack_pkt_delay = 1000;
 213 static int32_t rack_send_a_lot_in_prr = 1;
 214 static int32_t rack_min_to = 1000;      /* Number of microsecond  min timeout */
 215 static int32_t rack_verbose_logging = 0;
 216 static int32_t rack_ignore_data_after_close = 1;
 217 static int32_t rack_enable_shared_cwnd = 1;
 218 static int32_t rack_use_cmp_acks = 1;
 219 static int32_t rack_use_fsb = 1;
 220 static int32_t rack_use_rfo = 1;
 221 static int32_t rack_use_rsm_rfo = 1;
 222 static int32_t rack_max_abc_post_recovery = 2;
 223 static int32_t rack_client_low_buf = 0;
 224 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 */
 225 #ifdef TCP_ACCOUNTING
 226 static int32_t rack_tcp_accounting = 0;
 227 #endif
 228 static int32_t rack_limits_scwnd = 1;
 229 static int32_t rack_enable_mqueue_for_nonpaced = 0;
 230 static int32_t rack_disable_prr = 0;
 231 static int32_t use_rack_rr = 1;
 232 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
 233 static int32_t rack_persist_min = 250000;       /* 250usec */
 234 static int32_t rack_persist_max = 2000000;      /* 2 Second in usec's */
 235 static int32_t rack_sack_not_required = 1;      /* set to one to allow non-sack to use rack */
 236 static int32_t rack_default_init_window = 0;    /* Use system default */
 237 static int32_t rack_limit_time_with_srtt = 0;
 238 static int32_t rack_autosndbuf_inc = 20;        /* In percentage form */
 239 static int32_t rack_enobuf_hw_boost_mult = 2;   /* How many times the hw rate we boost slot using time_between */
 240 static int32_t rack_enobuf_hw_max = 12000;      /* 12 ms in usecs */
 241 static int32_t rack_enobuf_hw_min = 10000;      /* 10 ms in usecs */
 242 static int32_t rack_hw_rwnd_factor = 2;         /* How many max_segs the rwnd must be before we hold off sending */
 243
 244 /*
 245  * Currently regular tcp has a rto_min of 30ms
 246  * the backoff goes 12 times so that ends up
 247  * being a total of 122.850 seconds before a
 248  * connection is killed.
 249  */
 250 static uint32_t rack_def_data_window = 20;
 251 static uint32_t rack_goal_bdp = 2;
 252 static uint32_t rack_min_srtts = 1;
 253 static uint32_t rack_min_measure_usec = 0;
 254 static int32_t rack_tlp_min = 10000;    /* 10ms */
 255 static int32_t rack_rto_min = 30000;    /* 30,000 usec same as main freebsd */
 256 static int32_t rack_rto_max = 4000000;  /* 4 seconds in usec's */
 257 static const int32_t rack_free_cache = 2;
 258 static int32_t rack_hptsi_segments = 40;
 259 static int32_t rack_rate_sample_method = USE_RTT_LOW;
 260 static int32_t rack_pace_every_seg = 0;
 261 static int32_t rack_delayed_ack_time = 40000;   /* 40ms in usecs */
 262 static int32_t rack_slot_reduction = 4;
 263 static int32_t rack_wma_divisor = 8;            /* For WMA calculation */
 264 static int32_t rack_cwnd_block_ends_measure = 0;
 265 static int32_t rack_rwnd_block_ends_measure = 0;
 266 static int32_t rack_def_profile = 0;
 267
 268 static int32_t rack_lower_cwnd_at_tlp = 0;
 269 static int32_t rack_limited_retran = 0;
 270 static int32_t rack_always_send_oldest = 0;
 271 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
 272
 273 static uint16_t rack_per_of_gp_ss = 250;        /* 250 % slow-start */
 274 static uint16_t rack_per_of_gp_ca = 200;        /* 200 % congestion-avoidance */
 275 static uint16_t rack_per_of_gp_rec = 200;       /* 200 % of bw */
 276
 277 /* Probertt */
 278 static uint16_t rack_per_of_gp_probertt = 60;   /* 60% of bw */
 279 static uint16_t rack_per_of_gp_lowthresh = 40;  /* 40% is bottom */
 280 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
 281 static uint16_t rack_atexit_prtt_hbp = 130;     /* Clamp to 130% on exit prtt if highly buffered path */
 282 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
 283
 284 static uint32_t rack_max_drain_wait = 2;        /* How man gp srtt's before we give up draining */
 285 static uint32_t rack_must_drain = 1;            /* How many GP srtt's we *must* wait */
 286 static uint32_t rack_probertt_use_min_rtt_entry = 1;    /* Use the min to calculate the goal else gp_srtt */
 287 static uint32_t rack_probertt_use_min_rtt_exit = 0;
 288 static uint32_t rack_probe_rtt_sets_cwnd = 0;
 289 static uint32_t rack_probe_rtt_safety_val = 2000000;    /* No more than 2 sec in probe-rtt */
 290 static uint32_t rack_time_between_probertt = 9600000;   /* 9.6 sec in usecs */
 291 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;       /* How many srtt periods does probe-rtt last top fraction */
 292 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;       /* How many srtt periods does probe-rtt last bottom fraction */
 293 static uint32_t rack_min_probertt_hold = 40000;         /* Equal to delayed ack time */
 294 static uint32_t rack_probertt_filter_life = 10000000;
 295 static uint32_t rack_probertt_lower_within = 10;
 296 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds)  to count as a lowering */
 297 static int32_t rack_pace_one_seg = 0;           /* Shall we pace for less than 1.4Meg 1MSS at a time */
 298 static int32_t rack_probertt_clear_is = 1;
 299 static int32_t rack_max_drain_hbp = 1;          /* Extra drain times gpsrtt for highly buffered paths */
 300 static int32_t rack_hbp_thresh = 3;             /* what is the divisor max_rtt/min_rtt to decided a hbp */
 301
 302 /* Part of pacing */
 303 static int32_t rack_max_per_above = 30;         /* When we go to increment stop if above 100+this% */
 304
 305 /* Timely information */
 306 /* Combine these two gives the range of 'no change' to bw */
 307 /* ie the up/down provide the upper and lower bound */
 308 static int32_t rack_gp_per_bw_mul_up = 2;       /* 2% */
 309 static int32_t rack_gp_per_bw_mul_down = 4;     /* 4% */
 310 static int32_t rack_gp_rtt_maxmul = 3;          /* 3 x maxmin */
 311 static int32_t rack_gp_rtt_minmul = 1;          /* minrtt + (minrtt/mindiv) is lower rtt */
 312 static int32_t rack_gp_rtt_mindiv = 4;          /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
 313 static int32_t rack_gp_decrease_per = 20;       /* 20% decrease in multiplier */
 314 static int32_t rack_gp_increase_per = 2;        /* 2% increase in multiplier */
 315 static int32_t rack_per_lower_bound = 50;       /* Don't allow to drop below this multiplier */
 316 static int32_t rack_per_upper_bound_ss = 0;     /* Don't allow SS to grow above this */
 317 static int32_t rack_per_upper_bound_ca = 0;     /* Don't allow CA to grow above this */
 318 static int32_t rack_do_dyn_mul = 0;             /* Are the rack gp multipliers dynamic */
 319 static int32_t rack_gp_no_rec_chg = 1;          /* Prohibit recovery from reducing it's multiplier */
 320 static int32_t rack_timely_dec_clear = 6;       /* Do we clear decrement count at a value (6)? */
 321 static int32_t rack_timely_max_push_rise = 3;   /* One round of pushing */
 322 static int32_t rack_timely_max_push_drop = 3;   /* Three round of pushing */
 323 static int32_t rack_timely_min_segs = 4;        /* 4 segment minimum */
 324 static int32_t rack_use_max_for_nobackoff = 0;
 325 static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */
 326 static int32_t rack_timely_no_stopping = 0;
 327 static int32_t rack_down_raise_thresh = 100;
 328 static int32_t rack_req_segs = 1;
 329 static uint64_t rack_bw_rate_cap = 0;
 330 static uint32_t rack_trace_point_config = 0;
 331 static uint32_t rack_trace_point_bb_mode = 4;
 332 static int32_t rack_trace_point_count = 0;
 333
 334
 335 /* Weird delayed ack mode */
 336 static int32_t rack_use_imac_dack = 0;
 337 /* Rack specific counters */
 338 counter_u64_t rack_saw_enobuf;
 339 counter_u64_t rack_saw_enobuf_hw;
 340 counter_u64_t rack_saw_enetunreach;
 341 counter_u64_t rack_persists_sends;
 342 counter_u64_t rack_persists_acks;
 343 counter_u64_t rack_persists_loss;
 344 counter_u64_t rack_persists_lost_ends;
 345 #ifdef INVARIANTS
 346 counter_u64_t rack_adjust_map_bw;
 347 #endif
 348 /* Tail loss probe counters */
 349 counter_u64_t rack_tlp_tot;
 350 counter_u64_t rack_tlp_newdata;
 351 counter_u64_t rack_tlp_retran;
 352 counter_u64_t rack_tlp_retran_bytes;
 353 counter_u64_t rack_to_tot;
 354 counter_u64_t rack_hot_alloc;
 355 counter_u64_t rack_to_alloc;
 356 counter_u64_t rack_to_alloc_hard;
 357 counter_u64_t rack_to_alloc_emerg;
 358 counter_u64_t rack_to_alloc_limited;
 359 counter_u64_t rack_alloc_limited_conns;
 360 counter_u64_t rack_split_limited;
 361
 362 counter_u64_t rack_multi_single_eq;
 363 counter_u64_t rack_proc_non_comp_ack;
 364
 365 counter_u64_t rack_fto_send;
 366 counter_u64_t rack_fto_rsm_send;
 367 counter_u64_t rack_nfto_resend;
 368 counter_u64_t rack_non_fto_send;
 369 counter_u64_t rack_extended_rfo;
 370
 371 counter_u64_t rack_sack_proc_all;
 372 counter_u64_t rack_sack_proc_short;
 373 counter_u64_t rack_sack_proc_restart;
 374 counter_u64_t rack_sack_attacks_detected;
 375 counter_u64_t rack_sack_attacks_reversed;
 376 counter_u64_t rack_sack_used_next_merge;
 377 counter_u64_t rack_sack_splits;
 378 counter_u64_t rack_sack_used_prev_merge;
 379 counter_u64_t rack_sack_skipped_acked;
 380 counter_u64_t rack_ack_total;
 381 counter_u64_t rack_express_sack;
 382 counter_u64_t rack_sack_total;
 383 counter_u64_t rack_move_none;
 384 counter_u64_t rack_move_some;
 385
 386 counter_u64_t rack_input_idle_reduces;
 387 counter_u64_t rack_collapsed_win;
 388 counter_u64_t rack_collapsed_win_seen;
 389 counter_u64_t rack_collapsed_win_rxt;
 390 counter_u64_t rack_collapsed_win_rxt_bytes;
 391 counter_u64_t rack_try_scwnd;
 392 counter_u64_t rack_hw_pace_init_fail;
 393 counter_u64_t rack_hw_pace_lost;
 394
 395 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
 396 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
 397
 398
 399 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
 400
 401 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {  \
 402         (tv) = (value) + slop;   \
 403         if ((u_long)(tv) < (u_long)(tvmin)) \
 404                 (tv) = (tvmin); \
 405         if ((u_long)(tv) > (u_long)(tvmax)) \
 406                 (tv) = (tvmax); \
 407 } while (0)
 408
 409 static void
 410 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
 411
 412 static int
 413 rack_process_ack(struct mbuf *m, struct tcphdr *th,
 414     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
 415     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
 416 static int
 417 rack_process_data(struct mbuf *m, struct tcphdr *th,
 418     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 419     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
 420 static void
 421 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
 422    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
 423 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
 424 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
 425     uint8_t limit_type);
 426 static struct rack_sendmap *
 427 rack_check_recovery_mode(struct tcpcb *tp,
 428     uint32_t tsused);
 429 static void
 430 rack_cong_signal(struct tcpcb *tp,
 431                  uint32_t type, uint32_t ack, int );
 432 static void rack_counter_destroy(void);
 433 static int
 434 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt);
 435 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
 436 static void
 437 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
 438 static void
 439 rack_do_segment(struct mbuf *m, struct tcphdr *th,
 440     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 441     uint8_t iptos);
 442 static void rack_dtor(void *mem, int32_t size, void *arg);
 443 static void
 444 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
 445     uint32_t flex1, uint32_t flex2,
 446     uint32_t flex3, uint32_t flex4,
 447     uint32_t flex5, uint32_t flex6,
 448     uint16_t flex7, uint8_t mod);
 449
 450 static void
 451 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
 452    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
 453    struct rack_sendmap *rsm, uint8_t quality);
 454 static struct rack_sendmap *
 455 rack_find_high_nonack(struct tcp_rack *rack,
 456     struct rack_sendmap *rsm);
 457 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
 458 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
 459 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
 460 static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt);
 461 static void
 462 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
 463                             tcp_seq th_ack, int line, uint8_t quality);
 464 static uint32_t
 465 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
 466 static int32_t rack_handoff_ok(struct tcpcb *tp);
 467 static int32_t rack_init(struct tcpcb *tp);
 468 static void rack_init_sysctls(void);
 469 static void
 470 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
 471     struct tcphdr *th, int entered_rec, int dup_ack_struck);
 472 static void
 473 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
 474     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
 475     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
 476
 477 static void
 478 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
 479     struct rack_sendmap *rsm);
 480 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
 481 static int32_t rack_output(struct tcpcb *tp);
 482
 483 static uint32_t
 484 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
 485     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
 486     uint32_t cts, int *moved_two);
 487 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
 488 static void rack_remxt_tmr(struct tcpcb *tp);
 489 static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt);
 490 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
 491 static int32_t rack_stopall(struct tcpcb *tp);
 492 static void
 493 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
 494     uint32_t delta);
 495 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
 496 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
 497 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
 498 static uint32_t
 499 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
 500     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
 501 static void
 502 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
 503     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
 504 static int
 505 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
 506     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
 507 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
 508 static int
 509 rack_do_close_wait(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_closing(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 thflags, int32_t nxt_pkt, uint8_t iptos);
 516 static int
 517 rack_do_established(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_fastnewdata(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 nxt_pkt, uint8_t iptos);
 524 static int
 525 rack_do_fin_wait_1(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_fin_wait_2(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_lastack(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 static int
 537 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
 538     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 539     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 540 static int
 541 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
 542     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 543     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 544 struct rack_sendmap *
 545 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
 546     uint32_t tsused);
 547 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
 548     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
 549 static void
 550      tcp_rack_partialack(struct tcpcb *tp);
 551 static int
 552 rack_set_profile(struct tcp_rack *rack, int prof);
 553 static void
 554 rack_apply_deferred_options(struct tcp_rack *rack);
 555
 556 int32_t rack_clear_counter=0;
 557
 558 static inline void
 559 rack_trace_point(struct tcp_rack *rack, int num)
 560 {
 561         if (((rack_trace_point_config == num)  ||
 562              (rack_trace_point_config = 0xffffffff)) &&
 563             (rack_trace_point_bb_mode != 0) &&
 564             (rack_trace_point_count > 0) &&
 565             (rack->rc_tp->t_logstate == 0)) {
 566                 int res;
 567                 res = atomic_fetchadd_int(&rack_trace_point_count, -1);
 568                 if (res > 0) {
 569                         rack->rc_tp->t_logstate = rack_trace_point_bb_mode;
 570                 } else {
 571                         /* Loss a race assure its zero now */
 572                         rack_trace_point_count = 0;
 573                 }
 574         }
 575 }
 576
 577 static void
 578 rack_set_cc_pacing(struct tcp_rack *rack)
 579 {
 580         struct sockopt sopt;
 581         struct cc_newreno_opts opt;
 582         struct newreno old, *ptr;
 583         struct tcpcb *tp;
 584         int error;
 585
 586         if (rack->rc_pacing_cc_set)
 587                 return;
 588
 589         tp = rack->rc_tp;
 590         if (tp->cc_algo == NULL) {
 591                 /* Tcb is leaving */
 592                 return;
 593         }
 594         rack->rc_pacing_cc_set = 1;
 595         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 596                 /* Not new-reno we can't play games with beta! */
 597                 goto out;
 598         }
 599         ptr = ((struct newreno *)tp->ccv->cc_data);
 600         if (CC_ALGO(tp)->ctl_output == NULL)  {
 601                 /* Huh, why does new_reno no longer have a set function? */
 602                 goto out;
 603         }
 604         if (ptr == NULL) {
 605                 /* Just the default values */
 606                 old.beta = V_newreno_beta_ecn;
 607                 old.beta_ecn = V_newreno_beta_ecn;
 608                 old.newreno_flags = 0;
 609         } else {
 610                 old.beta = ptr->beta;
 611                 old.beta_ecn = ptr->beta_ecn;
 612                 old.newreno_flags = ptr->newreno_flags;
 613         }
 614         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 615         sopt.sopt_dir = SOPT_SET;
 616         opt.name = CC_NEWRENO_BETA;
 617         opt.val = rack->r_ctl.rc_saved_beta.beta;
 618         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 619         if (error)  {
 620                 goto out;
 621         }
 622         /*
 623          * Hack alert we need to set in our newreno_flags
 624          * so that Abe behavior is also applied.
 625          */
 626         ((struct newreno *)tp->ccv->cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
 627         opt.name = CC_NEWRENO_BETA_ECN;
 628         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 629         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 630         if (error) {
 631                 goto out;
 632         }
 633         /* Save off the original values for restoral */
 634         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 635 out:
 636         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 637                 union tcp_log_stackspecific log;
 638                 struct timeval tv;
 639
 640                 ptr = ((struct newreno *)tp->ccv->cc_data);
 641                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 642                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 643                 if (ptr) {
 644                         log.u_bbr.flex1 = ptr->beta;
 645                         log.u_bbr.flex2 = ptr->beta_ecn;
 646                         log.u_bbr.flex3 = ptr->newreno_flags;
 647                 }
 648                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 649                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 650                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 651                 log.u_bbr.flex7 = rack->gp_ready;
 652                 log.u_bbr.flex7 <<= 1;
 653                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 654                 log.u_bbr.flex7 <<= 1;
 655                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 656                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 657                 log.u_bbr.flex8 = 3;
 658                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 659                                0, &log, false, NULL, NULL, 0, &tv);
 660         }
 661 }
 662
 663 static void
 664 rack_undo_cc_pacing(struct tcp_rack *rack)
 665 {
 666         struct newreno old, *ptr;
 667         struct tcpcb *tp;
 668
 669         if (rack->rc_pacing_cc_set == 0)
 670                 return;
 671         tp = rack->rc_tp;
 672         rack->rc_pacing_cc_set = 0;
 673         if (tp->cc_algo == NULL)
 674                 /* Tcb is leaving */
 675                 return;
 676         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 677                 /* Not new-reno nothing to do! */
 678                 return;
 679         }
 680         ptr = ((struct newreno *)tp->ccv->cc_data);
 681         if (ptr == NULL) {
 682                 /*
 683                  * This happens at rack_fini() if the
 684                  * cc module gets freed on us. In that
 685                  * case we loose our "new" settings but
 686                  * thats ok, since the tcb is going away anyway.
 687                  */
 688                 return;
 689         }
 690         /* Grab out our set values */
 691         memcpy(&old, ptr, sizeof(struct newreno));
 692         /* Copy back in the original values */
 693         memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
 694         /* Now save back the values we had set in (for when pacing is restored) */
 695         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 696         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 697                 union tcp_log_stackspecific log;
 698                 struct timeval tv;
 699
 700                 ptr = ((struct newreno *)tp->ccv->cc_data);
 701                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 702                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 703                 log.u_bbr.flex1 = ptr->beta;
 704                 log.u_bbr.flex2 = ptr->beta_ecn;
 705                 log.u_bbr.flex3 = ptr->newreno_flags;
 706                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 707                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 708                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 709                 log.u_bbr.flex7 = rack->gp_ready;
 710                 log.u_bbr.flex7 <<= 1;
 711                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 712                 log.u_bbr.flex7 <<= 1;
 713                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 714                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 715                 log.u_bbr.flex8 = 4;
 716                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
 717                                0, &log, false, NULL, NULL, 0, &tv);
 718         }
 719 }
 720
 721 #ifdef NETFLIX_PEAKRATE
 722 static inline void
 723 rack_update_peakrate_thr(struct tcpcb *tp)
 724 {
 725         /* Keep in mind that t_maxpeakrate is in B/s. */
 726         uint64_t peak;
 727         peak = uqmax((tp->t_maxseg * 2),
 728                      (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
 729         tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
 730 }
 731 #endif
 732
 733 static int
 734 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 735 {
 736         uint32_t stat;
 737         int32_t error;
 738
 739         error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
 740         if (error || req->newptr == NULL)
 741                 return error;
 742
 743         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
 744         if (error)
 745                 return (error);
 746         if (stat == 1) {
 747 #ifdef INVARIANTS
 748                 printf("Clearing RACK counters\n");
 749 #endif
 750                 counter_u64_zero(rack_tlp_tot);
 751                 counter_u64_zero(rack_tlp_newdata);
 752                 counter_u64_zero(rack_tlp_retran);
 753                 counter_u64_zero(rack_tlp_retran_bytes);
 754                 counter_u64_zero(rack_to_tot);
 755                 counter_u64_zero(rack_saw_enobuf);
 756                 counter_u64_zero(rack_saw_enobuf_hw);
 757                 counter_u64_zero(rack_saw_enetunreach);
 758                 counter_u64_zero(rack_persists_sends);
 759                 counter_u64_zero(rack_persists_acks);
 760                 counter_u64_zero(rack_persists_loss);
 761                 counter_u64_zero(rack_persists_lost_ends);
 762 #ifdef INVARIANTS
 763                 counter_u64_zero(rack_adjust_map_bw);
 764 #endif
 765                 counter_u64_zero(rack_to_alloc_hard);
 766                 counter_u64_zero(rack_to_alloc_emerg);
 767                 counter_u64_zero(rack_sack_proc_all);
 768                 counter_u64_zero(rack_fto_send);
 769                 counter_u64_zero(rack_fto_rsm_send);
 770                 counter_u64_zero(rack_extended_rfo);
 771                 counter_u64_zero(rack_hw_pace_init_fail);
 772                 counter_u64_zero(rack_hw_pace_lost);
 773                 counter_u64_zero(rack_non_fto_send);
 774                 counter_u64_zero(rack_nfto_resend);
 775                 counter_u64_zero(rack_sack_proc_short);
 776                 counter_u64_zero(rack_sack_proc_restart);
 777                 counter_u64_zero(rack_to_alloc);
 778                 counter_u64_zero(rack_to_alloc_limited);
 779                 counter_u64_zero(rack_alloc_limited_conns);
 780                 counter_u64_zero(rack_split_limited);
 781                 counter_u64_zero(rack_multi_single_eq);
 782                 counter_u64_zero(rack_proc_non_comp_ack);
 783                 counter_u64_zero(rack_sack_attacks_detected);
 784                 counter_u64_zero(rack_sack_attacks_reversed);
 785                 counter_u64_zero(rack_sack_used_next_merge);
 786                 counter_u64_zero(rack_sack_used_prev_merge);
 787                 counter_u64_zero(rack_sack_splits);
 788                 counter_u64_zero(rack_sack_skipped_acked);
 789                 counter_u64_zero(rack_ack_total);
 790                 counter_u64_zero(rack_express_sack);
 791                 counter_u64_zero(rack_sack_total);
 792                 counter_u64_zero(rack_move_none);
 793                 counter_u64_zero(rack_move_some);
 794                 counter_u64_zero(rack_try_scwnd);
 795                 counter_u64_zero(rack_collapsed_win);
 796                 counter_u64_zero(rack_collapsed_win_rxt);
 797                 counter_u64_zero(rack_collapsed_win_seen);
 798                 counter_u64_zero(rack_collapsed_win_rxt_bytes);
 799         }
 800         rack_clear_counter = 0;
 801         return (0);
 802 }
 803
 804 static void
 805 rack_init_sysctls(void)
 806 {
 807         struct sysctl_oid *rack_counters;
 808         struct sysctl_oid *rack_attack;
 809         struct sysctl_oid *rack_pacing;
 810         struct sysctl_oid *rack_timely;
 811         struct sysctl_oid *rack_timers;
 812         struct sysctl_oid *rack_tlp;
 813         struct sysctl_oid *rack_misc;
 814         struct sysctl_oid *rack_features;
 815         struct sysctl_oid *rack_measure;
 816         struct sysctl_oid *rack_probertt;
 817         struct sysctl_oid *rack_hw_pacing;
 818         struct sysctl_oid *rack_tracepoint;
 819
 820         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 821             SYSCTL_CHILDREN(rack_sysctl_root),
 822             OID_AUTO,
 823             "sack_attack",
 824             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 825             "Rack Sack Attack Counters and Controls");
 826         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 827             SYSCTL_CHILDREN(rack_sysctl_root),
 828             OID_AUTO,
 829             "stats",
 830             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 831             "Rack Counters");
 832         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 833             SYSCTL_CHILDREN(rack_sysctl_root),
 834             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 835             &rack_rate_sample_method , USE_RTT_LOW,
 836             "What method should we use for rate sampling 0=high, 1=low ");
 837         /* Probe rtt related controls */
 838         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 839             SYSCTL_CHILDREN(rack_sysctl_root),
 840             OID_AUTO,
 841             "probertt",
 842             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 843             "ProbeRTT related Controls");
 844         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 845             SYSCTL_CHILDREN(rack_probertt),
 846             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 847             &rack_atexit_prtt_hbp, 130,
 848             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 849         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 850             SYSCTL_CHILDREN(rack_probertt),
 851             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 852             &rack_atexit_prtt, 130,
 853             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 854         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 855             SYSCTL_CHILDREN(rack_probertt),
 856             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 857             &rack_per_of_gp_probertt, 60,
 858             "What percentage of goodput do we pace at in probertt");
 859         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 860             SYSCTL_CHILDREN(rack_probertt),
 861             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 862             &rack_per_of_gp_probertt_reduce, 10,
 863             "What percentage of goodput do we reduce every gp_srtt");
 864         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 865             SYSCTL_CHILDREN(rack_probertt),
 866             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 867             &rack_per_of_gp_lowthresh, 40,
 868             "What percentage of goodput do we allow the multiplier to fall to");
 869         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 870             SYSCTL_CHILDREN(rack_probertt),
 871             OID_AUTO, "time_between", CTLFLAG_RW,
 872             & rack_time_between_probertt, 96000000,
 873             "How many useconds between the lowest rtt falling must past before we enter probertt");
 874         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 875             SYSCTL_CHILDREN(rack_probertt),
 876             OID_AUTO, "safety", CTLFLAG_RW,
 877             &rack_probe_rtt_safety_val, 2000000,
 878             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 879         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 880             SYSCTL_CHILDREN(rack_probertt),
 881             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 882             &rack_probe_rtt_sets_cwnd, 0,
 883             "Do we set the cwnd too (if always_lower is on)");
 884         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 885             SYSCTL_CHILDREN(rack_probertt),
 886             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 887             &rack_max_drain_wait, 2,
 888             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 889         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 890             SYSCTL_CHILDREN(rack_probertt),
 891             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 892             &rack_must_drain, 1,
 893             "We must drain this many gp_srtt's waiting for flight to reach goal");
 894         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 895             SYSCTL_CHILDREN(rack_probertt),
 896             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 897             &rack_probertt_use_min_rtt_entry, 1,
 898             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 899         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 900             SYSCTL_CHILDREN(rack_probertt),
 901             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 902             &rack_probertt_use_min_rtt_exit, 0,
 903             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 904         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 905             SYSCTL_CHILDREN(rack_probertt),
 906             OID_AUTO, "length_div", CTLFLAG_RW,
 907             &rack_probertt_gpsrtt_cnt_div, 0,
 908             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 909         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 910             SYSCTL_CHILDREN(rack_probertt),
 911             OID_AUTO, "length_mul", CTLFLAG_RW,
 912             &rack_probertt_gpsrtt_cnt_mul, 0,
 913             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 914         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 915             SYSCTL_CHILDREN(rack_probertt),
 916             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 917             &rack_min_probertt_hold, 200000,
 918             "What is the minimum time we hold probertt at target");
 919         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 920             SYSCTL_CHILDREN(rack_probertt),
 921             OID_AUTO, "filter_life", CTLFLAG_RW,
 922             &rack_probertt_filter_life, 10000000,
 923             "What is the time for the filters life in useconds");
 924         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 925             SYSCTL_CHILDREN(rack_probertt),
 926             OID_AUTO, "lower_within", CTLFLAG_RW,
 927             &rack_probertt_lower_within, 10,
 928             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 929         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 930             SYSCTL_CHILDREN(rack_probertt),
 931             OID_AUTO, "must_move", CTLFLAG_RW,
 932             &rack_min_rtt_movement, 250,
 933             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 934         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 935             SYSCTL_CHILDREN(rack_probertt),
 936             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 937             &rack_probertt_clear_is, 1,
 938             "Do we clear I/S counts on exiting probe-rtt");
 939         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 940             SYSCTL_CHILDREN(rack_probertt),
 941             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 942             &rack_max_drain_hbp, 1,
 943             "How many extra drain gpsrtt's do we get in highly buffered paths");
 944         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 945             SYSCTL_CHILDREN(rack_probertt),
 946             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 947             &rack_hbp_thresh, 3,
 948             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 949
 950         rack_tracepoint = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 951             SYSCTL_CHILDREN(rack_sysctl_root),
 952             OID_AUTO,
 953             "tp",
 954             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 955             "Rack tracepoint facility");
 956         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 957             SYSCTL_CHILDREN(rack_tracepoint),
 958             OID_AUTO, "number", CTLFLAG_RW,
 959             &rack_trace_point_config, 0,
 960             "What is the trace point number to activate (0=none, 0xffffffff = all)?");
 961         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 962             SYSCTL_CHILDREN(rack_tracepoint),
 963             OID_AUTO, "bbmode", CTLFLAG_RW,
 964             &rack_trace_point_bb_mode, 4,
 965             "What is BB logging mode that is activated?");
 966         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 967             SYSCTL_CHILDREN(rack_tracepoint),
 968             OID_AUTO, "count", CTLFLAG_RW,
 969             &rack_trace_point_count, 0,
 970             "How many connections will have BB logging turned on that hit the tracepoint?");
 971         /* Pacing related sysctls */
 972         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 973             SYSCTL_CHILDREN(rack_sysctl_root),
 974             OID_AUTO,
 975             "pacing",
 976             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 977             "Pacing related Controls");
 978         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 979             SYSCTL_CHILDREN(rack_pacing),
 980             OID_AUTO, "max_pace_over", CTLFLAG_RW,
 981             &rack_max_per_above, 30,
 982             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
 983         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 984             SYSCTL_CHILDREN(rack_pacing),
 985             OID_AUTO, "pace_to_one", CTLFLAG_RW,
 986             &rack_pace_one_seg, 0,
 987             "Do we allow low b/w pacing of 1MSS instead of two");
 988         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 989             SYSCTL_CHILDREN(rack_pacing),
 990             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
 991             &rack_limit_time_with_srtt, 0,
 992             "Do we limit pacing time based on srtt");
 993         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 994             SYSCTL_CHILDREN(rack_pacing),
 995             OID_AUTO, "init_win", CTLFLAG_RW,
 996             &rack_default_init_window, 0,
 997             "Do we have a rack initial window 0 = system default");
 998         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 999             SYSCTL_CHILDREN(rack_pacing),
1000             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1001             &rack_per_of_gp_ss, 250,
1002             "If non zero, what percentage of goodput to pace at in slow start");
1003         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1004             SYSCTL_CHILDREN(rack_pacing),
1005             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1006             &rack_per_of_gp_ca, 150,
1007             "If non zero, what percentage of goodput to pace at in congestion avoidance");
1008         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1009             SYSCTL_CHILDREN(rack_pacing),
1010             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1011             &rack_per_of_gp_rec, 200,
1012             "If non zero, what percentage of goodput to pace at in recovery");
1013         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1014             SYSCTL_CHILDREN(rack_pacing),
1015             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1016             &rack_hptsi_segments, 40,
1017             "What size is the max for TSO segments in pacing and burst mitigation");
1018         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1019             SYSCTL_CHILDREN(rack_pacing),
1020             OID_AUTO, "burst_reduces", CTLFLAG_RW,
1021             &rack_slot_reduction, 4,
1022             "When doing only burst mitigation what is the reduce divisor");
1023         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1024             SYSCTL_CHILDREN(rack_sysctl_root),
1025             OID_AUTO, "use_pacing", CTLFLAG_RW,
1026             &rack_pace_every_seg, 0,
1027             "If set we use pacing, if clear we use only the original burst mitigation");
1028         SYSCTL_ADD_U64(&rack_sysctl_ctx,
1029             SYSCTL_CHILDREN(rack_pacing),
1030             OID_AUTO, "rate_cap", CTLFLAG_RW,
1031             &rack_bw_rate_cap, 0,
1032             "If set we apply this value to the absolute rate cap used by pacing");
1033         SYSCTL_ADD_U8(&rack_sysctl_ctx,
1034             SYSCTL_CHILDREN(rack_sysctl_root),
1035             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1036             &rack_req_measurements, 1,
1037             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1038         /* Hardware pacing */
1039         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1040             SYSCTL_CHILDREN(rack_sysctl_root),
1041             OID_AUTO,
1042             "hdwr_pacing",
1043             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1044             "Pacing related Controls");
1045         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1046             SYSCTL_CHILDREN(rack_hw_pacing),
1047             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1048             &rack_hw_rwnd_factor, 2,
1049             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1050         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1051             SYSCTL_CHILDREN(rack_hw_pacing),
1052             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1053             &rack_enobuf_hw_boost_mult, 2,
1054             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1055         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1056             SYSCTL_CHILDREN(rack_hw_pacing),
1057             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1058             &rack_enobuf_hw_max, 2,
1059             "What is the max boost the pacing time if we see a ENOBUFS?");
1060         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1061             SYSCTL_CHILDREN(rack_hw_pacing),
1062             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1063             &rack_enobuf_hw_min, 2,
1064             "What is the min boost the pacing time if we see a ENOBUFS?");
1065         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1066             SYSCTL_CHILDREN(rack_hw_pacing),
1067             OID_AUTO, "enable", CTLFLAG_RW,
1068             &rack_enable_hw_pacing, 0,
1069             "Should RACK attempt to use hw pacing?");
1070         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1071             SYSCTL_CHILDREN(rack_hw_pacing),
1072             OID_AUTO, "rate_cap", CTLFLAG_RW,
1073             &rack_hw_rate_caps, 1,
1074             "Does the highest hardware pacing rate cap the rate we will send at??");
1075         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1076             SYSCTL_CHILDREN(rack_hw_pacing),
1077             OID_AUTO, "rate_min", CTLFLAG_RW,
1078             &rack_hw_rate_min, 0,
1079             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1080         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1081             SYSCTL_CHILDREN(rack_hw_pacing),
1082             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1083             &rack_hw_rate_to_low, 0,
1084             "If we fall below this rate, dis-engage hw pacing?");
1085         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1086             SYSCTL_CHILDREN(rack_hw_pacing),
1087             OID_AUTO, "up_only", CTLFLAG_RW,
1088             &rack_hw_up_only, 1,
1089             "Do we allow hw pacing to lower the rate selected?");
1090         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1091             SYSCTL_CHILDREN(rack_hw_pacing),
1092             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1093             &rack_hw_pace_extra_slots, 2,
1094             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1095         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1096             SYSCTL_CHILDREN(rack_sysctl_root),
1097             OID_AUTO,
1098             "timely",
1099             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1100             "Rack Timely RTT Controls");
1101         /* Timely based GP dynmics */
1102         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1103             SYSCTL_CHILDREN(rack_timely),
1104             OID_AUTO, "upper", CTLFLAG_RW,
1105             &rack_gp_per_bw_mul_up, 2,
1106             "Rack timely upper range for equal b/w (in percentage)");
1107         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1108             SYSCTL_CHILDREN(rack_timely),
1109             OID_AUTO, "lower", CTLFLAG_RW,
1110             &rack_gp_per_bw_mul_down, 4,
1111             "Rack timely lower range for equal b/w (in percentage)");
1112         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1113             SYSCTL_CHILDREN(rack_timely),
1114             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1115             &rack_gp_rtt_maxmul, 3,
1116             "Rack timely multiplier of lowest rtt for rtt_max");
1117         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1118             SYSCTL_CHILDREN(rack_timely),
1119             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1120             &rack_gp_rtt_mindiv, 4,
1121             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1122         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1123             SYSCTL_CHILDREN(rack_timely),
1124             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1125             &rack_gp_rtt_minmul, 1,
1126             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1127         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1128             SYSCTL_CHILDREN(rack_timely),
1129             OID_AUTO, "decrease", CTLFLAG_RW,
1130             &rack_gp_decrease_per, 20,
1131             "Rack timely decrease percentage of our GP multiplication factor");
1132         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1133             SYSCTL_CHILDREN(rack_timely),
1134             OID_AUTO, "increase", CTLFLAG_RW,
1135             &rack_gp_increase_per, 2,
1136             "Rack timely increase perentage of our GP multiplication factor");
1137         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1138             SYSCTL_CHILDREN(rack_timely),
1139             OID_AUTO, "lowerbound", CTLFLAG_RW,
1140             &rack_per_lower_bound, 50,
1141             "Rack timely lowest percentage we allow GP multiplier to fall to");
1142         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1143             SYSCTL_CHILDREN(rack_timely),
1144             OID_AUTO, "upperboundss", CTLFLAG_RW,
1145             &rack_per_upper_bound_ss, 0,
1146             "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1147         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1148             SYSCTL_CHILDREN(rack_timely),
1149             OID_AUTO, "upperboundca", CTLFLAG_RW,
1150             &rack_per_upper_bound_ca, 0,
1151             "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1152         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1153             SYSCTL_CHILDREN(rack_timely),
1154             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1155             &rack_do_dyn_mul, 0,
1156             "Rack timely do we enable dynmaic timely goodput by default");
1157         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1158             SYSCTL_CHILDREN(rack_timely),
1159             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1160             &rack_gp_no_rec_chg, 1,
1161             "Rack timely do we prohibit the recovery multiplier from being lowered");
1162         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1163             SYSCTL_CHILDREN(rack_timely),
1164             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1165             &rack_timely_dec_clear, 6,
1166             "Rack timely what threshold do we count to before another boost during b/w decent");
1167         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1168             SYSCTL_CHILDREN(rack_timely),
1169             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1170             &rack_timely_max_push_rise, 3,
1171             "Rack timely how many times do we push up with b/w increase");
1172         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1173             SYSCTL_CHILDREN(rack_timely),
1174             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1175             &rack_timely_max_push_drop, 3,
1176             "Rack timely how many times do we push back on b/w decent");
1177         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1178             SYSCTL_CHILDREN(rack_timely),
1179             OID_AUTO, "min_segs", CTLFLAG_RW,
1180             &rack_timely_min_segs, 4,
1181             "Rack timely when setting the cwnd what is the min num segments");
1182         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1183             SYSCTL_CHILDREN(rack_timely),
1184             OID_AUTO, "noback_max", CTLFLAG_RW,
1185             &rack_use_max_for_nobackoff, 0,
1186             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1187         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1188             SYSCTL_CHILDREN(rack_timely),
1189             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1190             &rack_timely_int_timely_only, 0,
1191             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1192         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1193             SYSCTL_CHILDREN(rack_timely),
1194             OID_AUTO, "nonstop", CTLFLAG_RW,
1195             &rack_timely_no_stopping, 0,
1196             "Rack timely don't stop increase");
1197         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1198             SYSCTL_CHILDREN(rack_timely),
1199             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1200             &rack_down_raise_thresh, 100,
1201             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1202         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1203             SYSCTL_CHILDREN(rack_timely),
1204             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1205             &rack_req_segs, 1,
1206             "Bottom dragging if not these many segments outstanding and room");
1207
1208         /* TLP and Rack related parameters */
1209         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1210             SYSCTL_CHILDREN(rack_sysctl_root),
1211             OID_AUTO,
1212             "tlp",
1213             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1214             "TLP and Rack related Controls");
1215         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1216             SYSCTL_CHILDREN(rack_tlp),
1217             OID_AUTO, "use_rrr", CTLFLAG_RW,
1218             &use_rack_rr, 1,
1219             "Do we use Rack Rapid Recovery");
1220         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1221             SYSCTL_CHILDREN(rack_tlp),
1222             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1223             &rack_max_abc_post_recovery, 2,
1224             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1225         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1226             SYSCTL_CHILDREN(rack_tlp),
1227             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1228             &rack_non_rxt_use_cr, 0,
1229             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1230         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1231             SYSCTL_CHILDREN(rack_tlp),
1232             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1233             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1234             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1235         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1236             SYSCTL_CHILDREN(rack_tlp),
1237             OID_AUTO, "limit", CTLFLAG_RW,
1238             &rack_tlp_limit, 2,
1239             "How many TLP's can be sent without sending new data");
1240         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1241             SYSCTL_CHILDREN(rack_tlp),
1242             OID_AUTO, "use_greater", CTLFLAG_RW,
1243             &rack_tlp_use_greater, 1,
1244             "Should we use the rack_rtt time if its greater than srtt");
1245         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1246             SYSCTL_CHILDREN(rack_tlp),
1247             OID_AUTO, "tlpminto", CTLFLAG_RW,
1248             &rack_tlp_min, 10000,
1249             "TLP minimum timeout per the specification (in microseconds)");
1250         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1251             SYSCTL_CHILDREN(rack_tlp),
1252             OID_AUTO, "send_oldest", CTLFLAG_RW,
1253             &rack_always_send_oldest, 0,
1254             "Should we always send the oldest TLP and RACK-TLP");
1255         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1256             SYSCTL_CHILDREN(rack_tlp),
1257             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1258             &rack_limited_retran, 0,
1259             "How many times can a rack timeout drive out sends");
1260         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1261             SYSCTL_CHILDREN(rack_tlp),
1262             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1263             &rack_lower_cwnd_at_tlp, 0,
1264             "When a TLP completes a retran should we enter recovery");
1265         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1266             SYSCTL_CHILDREN(rack_tlp),
1267             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1268             &rack_reorder_thresh, 2,
1269             "What factor for rack will be added when seeing reordering (shift right)");
1270         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1271             SYSCTL_CHILDREN(rack_tlp),
1272             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1273             &rack_tlp_thresh, 1,
1274             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1275         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1276             SYSCTL_CHILDREN(rack_tlp),
1277             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1278             &rack_reorder_fade, 60000000,
1279             "Does reorder detection fade, if so how many microseconds (0 means never)");
1280         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1281             SYSCTL_CHILDREN(rack_tlp),
1282             OID_AUTO, "pktdelay", CTLFLAG_RW,
1283             &rack_pkt_delay, 1000,
1284             "Extra RACK time (in microseconds) besides reordering thresh");
1285
1286         /* Timer related controls */
1287         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1288             SYSCTL_CHILDREN(rack_sysctl_root),
1289             OID_AUTO,
1290             "timers",
1291             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1292             "Timer related controls");
1293         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1294             SYSCTL_CHILDREN(rack_timers),
1295             OID_AUTO, "persmin", CTLFLAG_RW,
1296             &rack_persist_min, 250000,
1297             "What is the minimum time in microseconds between persists");
1298         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1299             SYSCTL_CHILDREN(rack_timers),
1300             OID_AUTO, "persmax", CTLFLAG_RW,
1301             &rack_persist_max, 2000000,
1302             "What is the largest delay in microseconds between persists");
1303         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1304             SYSCTL_CHILDREN(rack_timers),
1305             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1306             &rack_delayed_ack_time, 40000,
1307             "Delayed ack time (40ms in microseconds)");
1308         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1309             SYSCTL_CHILDREN(rack_timers),
1310             OID_AUTO, "minrto", CTLFLAG_RW,
1311             &rack_rto_min, 30000,
1312             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1313         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1314             SYSCTL_CHILDREN(rack_timers),
1315             OID_AUTO, "maxrto", CTLFLAG_RW,
1316             &rack_rto_max, 4000000,
1317             "Maximum RTO in microseconds -- should be at least as large as min_rto");
1318         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1319             SYSCTL_CHILDREN(rack_timers),
1320             OID_AUTO, "minto", CTLFLAG_RW,
1321             &rack_min_to, 1000,
1322             "Minimum rack timeout in microseconds");
1323         /* Measure controls */
1324         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1325             SYSCTL_CHILDREN(rack_sysctl_root),
1326             OID_AUTO,
1327             "measure",
1328             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1329             "Measure related controls");
1330         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1331             SYSCTL_CHILDREN(rack_measure),
1332             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1333             &rack_wma_divisor, 8,
1334             "When doing b/w calculation what is the  divisor for the WMA");
1335         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1336             SYSCTL_CHILDREN(rack_measure),
1337             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1338             &rack_cwnd_block_ends_measure, 0,
1339             "Does a cwnd just-return end the measurement window (app limited)");
1340         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1341             SYSCTL_CHILDREN(rack_measure),
1342             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1343             &rack_rwnd_block_ends_measure, 0,
1344             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1345         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1346             SYSCTL_CHILDREN(rack_measure),
1347             OID_AUTO, "min_target", CTLFLAG_RW,
1348             &rack_def_data_window, 20,
1349             "What is the minimum target window (in mss) for a GP measurements");
1350         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1351             SYSCTL_CHILDREN(rack_measure),
1352             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1353             &rack_goal_bdp, 2,
1354             "What is the goal BDP to measure");
1355         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1356             SYSCTL_CHILDREN(rack_measure),
1357             OID_AUTO, "min_srtts", CTLFLAG_RW,
1358             &rack_min_srtts, 1,
1359             "What is the goal BDP to measure");
1360         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1361             SYSCTL_CHILDREN(rack_measure),
1362             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1363             &rack_min_measure_usec, 0,
1364             "What is the Minimum time time for a measurement if 0, this is off");
1365         /* Features */
1366         rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1367             SYSCTL_CHILDREN(rack_sysctl_root),
1368             OID_AUTO,
1369             "features",
1370             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1371             "Feature controls");
1372         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1373             SYSCTL_CHILDREN(rack_features),
1374             OID_AUTO, "cmpack", CTLFLAG_RW,
1375             &rack_use_cmp_acks, 1,
1376             "Should RACK have LRO send compressed acks");
1377         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1378             SYSCTL_CHILDREN(rack_features),
1379             OID_AUTO, "fsb", CTLFLAG_RW,
1380             &rack_use_fsb, 1,
1381             "Should RACK use the fast send block?");
1382         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1383             SYSCTL_CHILDREN(rack_features),
1384             OID_AUTO, "rfo", CTLFLAG_RW,
1385             &rack_use_rfo, 1,
1386             "Should RACK use rack_fast_output()?");
1387         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1388             SYSCTL_CHILDREN(rack_features),
1389             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1390             &rack_use_rsm_rfo, 1,
1391             "Should RACK use rack_fast_rsm_output()?");
1392         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1393             SYSCTL_CHILDREN(rack_features),
1394             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1395             &rack_enable_mqueue_for_nonpaced, 0,
1396             "Should RACK use mbuf queuing for non-paced connections");
1397         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1398             SYSCTL_CHILDREN(rack_features),
1399             OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1400             &rack_do_hystart, 0,
1401             "Should RACK enable HyStart++ on connections?");
1402         /* Misc rack controls */
1403         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1404             SYSCTL_CHILDREN(rack_sysctl_root),
1405             OID_AUTO,
1406             "misc",
1407             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1408             "Misc related controls");
1409 #ifdef TCP_ACCOUNTING
1410         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1411             SYSCTL_CHILDREN(rack_misc),
1412             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1413             &rack_tcp_accounting, 0,
1414             "Should we turn on TCP accounting for all rack sessions?");
1415 #endif
1416         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1417             SYSCTL_CHILDREN(rack_misc),
1418             OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1419             &rack_apply_rtt_with_reduced_conf, 0,
1420             "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1421         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1422             SYSCTL_CHILDREN(rack_misc),
1423             OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1424             &rack_dsack_std_based, 3,
1425             "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1426         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1427             SYSCTL_CHILDREN(rack_misc),
1428             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1429             &rack_prr_addbackmax, 2,
1430             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1431         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1432             SYSCTL_CHILDREN(rack_misc),
1433             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1434             &rack_stats_gets_ms_rtt, 1,
1435             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1436         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1437             SYSCTL_CHILDREN(rack_misc),
1438             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1439             &rack_client_low_buf, 0,
1440             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1441         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1442             SYSCTL_CHILDREN(rack_misc),
1443             OID_AUTO, "defprofile", CTLFLAG_RW,
1444             &rack_def_profile, 0,
1445             "Should RACK use a default profile (0=no, num == profile num)?");
1446         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1447             SYSCTL_CHILDREN(rack_misc),
1448             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1449             &rack_enable_shared_cwnd, 1,
1450             "Should RACK try to use the shared cwnd on connections where allowed");
1451         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1452             SYSCTL_CHILDREN(rack_misc),
1453             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1454             &rack_limits_scwnd, 1,
1455             "Should RACK place low end time limits on the shared cwnd feature");
1456         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1457             SYSCTL_CHILDREN(rack_misc),
1458             OID_AUTO, "iMac_dack", CTLFLAG_RW,
1459             &rack_use_imac_dack, 0,
1460             "Should RACK try to emulate iMac delayed ack");
1461         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1462             SYSCTL_CHILDREN(rack_misc),
1463             OID_AUTO, "no_prr", CTLFLAG_RW,
1464             &rack_disable_prr, 0,
1465             "Should RACK not use prr and only pace (must have pacing on)");
1466         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1467             SYSCTL_CHILDREN(rack_misc),
1468             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1469             &rack_verbose_logging, 0,
1470             "Should RACK black box logging be verbose");
1471         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1472             SYSCTL_CHILDREN(rack_misc),
1473             OID_AUTO, "data_after_close", CTLFLAG_RW,
1474             &rack_ignore_data_after_close, 1,
1475             "Do we hold off sending a RST until all pending data is ack'd");
1476         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1477             SYSCTL_CHILDREN(rack_misc),
1478             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1479             &rack_sack_not_required, 1,
1480             "Do we allow rack to run on connections not supporting SACK");
1481         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1482             SYSCTL_CHILDREN(rack_misc),
1483             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1484             &rack_send_a_lot_in_prr, 1,
1485             "Send a lot in prr");
1486         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1487             SYSCTL_CHILDREN(rack_misc),
1488             OID_AUTO, "autoscale", CTLFLAG_RW,
1489             &rack_autosndbuf_inc, 20,
1490             "What percentage should rack scale up its snd buffer by?");
1491         /* Sack Attacker detection stuff */
1492         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1493             SYSCTL_CHILDREN(rack_attack),
1494             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1495             &rack_highest_sack_thresh_seen, 0,
1496             "Highest sack to ack ratio seen");
1497         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1498             SYSCTL_CHILDREN(rack_attack),
1499             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1500             &rack_highest_move_thresh_seen, 0,
1501             "Highest move to non-move ratio seen");
1502         rack_ack_total = counter_u64_alloc(M_WAITOK);
1503         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1504             SYSCTL_CHILDREN(rack_attack),
1505             OID_AUTO, "acktotal", CTLFLAG_RD,
1506             &rack_ack_total,
1507             "Total number of Ack's");
1508         rack_express_sack = counter_u64_alloc(M_WAITOK);
1509         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1510             SYSCTL_CHILDREN(rack_attack),
1511             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1512             &rack_express_sack,
1513             "Total expresss number of Sack's");
1514         rack_sack_total = counter_u64_alloc(M_WAITOK);
1515         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1516             SYSCTL_CHILDREN(rack_attack),
1517             OID_AUTO, "sacktotal", CTLFLAG_RD,
1518             &rack_sack_total,
1519             "Total number of SACKs");
1520         rack_move_none = counter_u64_alloc(M_WAITOK);
1521         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1522             SYSCTL_CHILDREN(rack_attack),
1523             OID_AUTO, "move_none", CTLFLAG_RD,
1524             &rack_move_none,
1525             "Total number of SACK index reuse of positions under threshold");
1526         rack_move_some = counter_u64_alloc(M_WAITOK);
1527         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1528             SYSCTL_CHILDREN(rack_attack),
1529             OID_AUTO, "move_some", CTLFLAG_RD,
1530             &rack_move_some,
1531             "Total number of SACK index reuse of positions over threshold");
1532         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1533         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1534             SYSCTL_CHILDREN(rack_attack),
1535             OID_AUTO, "attacks", CTLFLAG_RD,
1536             &rack_sack_attacks_detected,
1537             "Total number of SACK attackers that had sack disabled");
1538         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1539         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1540             SYSCTL_CHILDREN(rack_attack),
1541             OID_AUTO, "reversed", CTLFLAG_RD,
1542             &rack_sack_attacks_reversed,
1543             "Total number of SACK attackers that were later determined false positive");
1544         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1545         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1546             SYSCTL_CHILDREN(rack_attack),
1547             OID_AUTO, "nextmerge", CTLFLAG_RD,
1548             &rack_sack_used_next_merge,
1549             "Total number of times we used the next merge");
1550         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1551         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1552             SYSCTL_CHILDREN(rack_attack),
1553             OID_AUTO, "prevmerge", CTLFLAG_RD,
1554             &rack_sack_used_prev_merge,
1555             "Total number of times we used the prev merge");
1556         /* Counters */
1557         rack_fto_send = counter_u64_alloc(M_WAITOK);
1558         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1559             SYSCTL_CHILDREN(rack_counters),
1560             OID_AUTO, "fto_send", CTLFLAG_RD,
1561             &rack_fto_send, "Total number of rack_fast_output sends");
1562         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1563         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1564             SYSCTL_CHILDREN(rack_counters),
1565             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1566             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1567         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1568         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1569             SYSCTL_CHILDREN(rack_counters),
1570             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1571             &rack_nfto_resend, "Total number of rack_output retransmissions");
1572         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1573         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1574             SYSCTL_CHILDREN(rack_counters),
1575             OID_AUTO, "nfto_send", CTLFLAG_RD,
1576             &rack_non_fto_send, "Total number of rack_output first sends");
1577         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1578         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1579             SYSCTL_CHILDREN(rack_counters),
1580             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1581             &rack_extended_rfo, "Total number of times we extended rfo");
1582
1583         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1584         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1585             SYSCTL_CHILDREN(rack_counters),
1586             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1587             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1588         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1589
1590         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1591             SYSCTL_CHILDREN(rack_counters),
1592             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1593             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1594         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1595         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1596             SYSCTL_CHILDREN(rack_counters),
1597             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1598             &rack_tlp_tot,
1599             "Total number of tail loss probe expirations");
1600         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1601         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1602             SYSCTL_CHILDREN(rack_counters),
1603             OID_AUTO, "tlp_new", CTLFLAG_RD,
1604             &rack_tlp_newdata,
1605             "Total number of tail loss probe sending new data");
1606         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1607         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1608             SYSCTL_CHILDREN(rack_counters),
1609             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1610             &rack_tlp_retran,
1611             "Total number of tail loss probe sending retransmitted data");
1612         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1613         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1614             SYSCTL_CHILDREN(rack_counters),
1615             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1616             &rack_tlp_retran_bytes,
1617             "Total bytes of tail loss probe sending retransmitted data");
1618         rack_to_tot = counter_u64_alloc(M_WAITOK);
1619         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1620             SYSCTL_CHILDREN(rack_counters),
1621             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1622             &rack_to_tot,
1623             "Total number of times the rack to expired");
1624         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1625         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1626             SYSCTL_CHILDREN(rack_counters),
1627             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1628             &rack_saw_enobuf,
1629             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1630         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1631         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1632             SYSCTL_CHILDREN(rack_counters),
1633             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1634             &rack_saw_enobuf_hw,
1635             "Total number of times a send returned enobuf for hdwr paced connections");
1636         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1637         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1638             SYSCTL_CHILDREN(rack_counters),
1639             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1640             &rack_saw_enetunreach,
1641             "Total number of times a send received a enetunreachable");
1642         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1643         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1644             SYSCTL_CHILDREN(rack_counters),
1645             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1646             &rack_hot_alloc,
1647             "Total allocations from the top of our list");
1648         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1649         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1650             SYSCTL_CHILDREN(rack_counters),
1651             OID_AUTO, "allocs", CTLFLAG_RD,
1652             &rack_to_alloc,
1653             "Total allocations of tracking structures");
1654         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1655         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1656             SYSCTL_CHILDREN(rack_counters),
1657             OID_AUTO, "allochard", CTLFLAG_RD,
1658             &rack_to_alloc_hard,
1659             "Total allocations done with sleeping the hard way");
1660         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1661         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1662             SYSCTL_CHILDREN(rack_counters),
1663             OID_AUTO, "allocemerg", CTLFLAG_RD,
1664             &rack_to_alloc_emerg,
1665             "Total allocations done from emergency cache");
1666         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1667         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1668             SYSCTL_CHILDREN(rack_counters),
1669             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1670             &rack_to_alloc_limited,
1671             "Total allocations dropped due to limit");
1672         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1673         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674             SYSCTL_CHILDREN(rack_counters),
1675             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1676             &rack_alloc_limited_conns,
1677             "Connections with allocations dropped due to limit");
1678         rack_split_limited = counter_u64_alloc(M_WAITOK);
1679         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680             SYSCTL_CHILDREN(rack_counters),
1681             OID_AUTO, "split_limited", CTLFLAG_RD,
1682             &rack_split_limited,
1683             "Split allocations dropped due to limit");
1684         rack_persists_sends = counter_u64_alloc(M_WAITOK);
1685         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1686             SYSCTL_CHILDREN(rack_counters),
1687             OID_AUTO, "persist_sends", CTLFLAG_RD,
1688             &rack_persists_sends,
1689             "Number of times we sent a persist probe");
1690         rack_persists_acks = counter_u64_alloc(M_WAITOK);
1691         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1692             SYSCTL_CHILDREN(rack_counters),
1693             OID_AUTO, "persist_acks", CTLFLAG_RD,
1694             &rack_persists_acks,
1695             "Number of times a persist probe was acked");
1696         rack_persists_loss = counter_u64_alloc(M_WAITOK);
1697         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1698             SYSCTL_CHILDREN(rack_counters),
1699             OID_AUTO, "persist_loss", CTLFLAG_RD,
1700             &rack_persists_loss,
1701             "Number of times we detected a lost persist probe (no ack)");
1702         rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1703         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1704             SYSCTL_CHILDREN(rack_counters),
1705             OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1706             &rack_persists_lost_ends,
1707             "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1708 #ifdef INVARIANTS
1709         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1710         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1711             SYSCTL_CHILDREN(rack_counters),
1712             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1713             &rack_adjust_map_bw,
1714             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1715 #endif
1716         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1717         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1718             SYSCTL_CHILDREN(rack_counters),
1719             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1720             &rack_multi_single_eq,
1721             "Number of compressed acks total represented");
1722         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1723         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1724             SYSCTL_CHILDREN(rack_counters),
1725             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1726             &rack_proc_non_comp_ack,
1727             "Number of non compresseds acks that we processed");
1728
1729
1730         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1731         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1732             SYSCTL_CHILDREN(rack_counters),
1733             OID_AUTO, "sack_long", CTLFLAG_RD,
1734             &rack_sack_proc_all,
1735             "Total times we had to walk whole list for sack processing");
1736         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1737         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1738             SYSCTL_CHILDREN(rack_counters),
1739             OID_AUTO, "sack_restart", CTLFLAG_RD,
1740             &rack_sack_proc_restart,
1741             "Total times we had to walk whole list due to a restart");
1742         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1743         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744             SYSCTL_CHILDREN(rack_counters),
1745             OID_AUTO, "sack_short", CTLFLAG_RD,
1746             &rack_sack_proc_short,
1747             "Total times we took shortcut for sack processing");
1748         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1749         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750             SYSCTL_CHILDREN(rack_attack),
1751             OID_AUTO, "skipacked", CTLFLAG_RD,
1752             &rack_sack_skipped_acked,
1753             "Total number of times we skipped previously sacked");
1754         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1755         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756             SYSCTL_CHILDREN(rack_attack),
1757             OID_AUTO, "ofsplit", CTLFLAG_RD,
1758             &rack_sack_splits,
1759             "Total number of times we did the old fashion tree split");
1760         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1761         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1762             SYSCTL_CHILDREN(rack_counters),
1763             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1764             &rack_input_idle_reduces,
1765             "Total number of idle reductions on input");
1766         rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1767         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1768             SYSCTL_CHILDREN(rack_counters),
1769             OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1770             &rack_collapsed_win_seen,
1771             "Total number of collapsed window events seen (where our window shrinks)");
1772
1773         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1774         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1775             SYSCTL_CHILDREN(rack_counters),
1776             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1777             &rack_collapsed_win,
1778             "Total number of collapsed window events where we mark packets");
1779         rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1780         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1781             SYSCTL_CHILDREN(rack_counters),
1782             OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1783             &rack_collapsed_win_rxt,
1784             "Total number of packets that were retransmitted");
1785         rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1786         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1787             SYSCTL_CHILDREN(rack_counters),
1788             OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1789             &rack_collapsed_win_rxt_bytes,
1790             "Total number of bytes that were retransmitted");
1791         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1792         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1793             SYSCTL_CHILDREN(rack_counters),
1794             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1795             &rack_try_scwnd,
1796             "Total number of scwnd attempts");
1797         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1798         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1799             OID_AUTO, "outsize", CTLFLAG_RD,
1800             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1801         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1802         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1803             OID_AUTO, "opts", CTLFLAG_RD,
1804             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1805         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1806             SYSCTL_CHILDREN(rack_sysctl_root),
1807             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1808             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1809 }
1810
1811 static __inline int
1812 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1813 {
1814         if (SEQ_GEQ(b->r_start, a->r_start) &&
1815             SEQ_LT(b->r_start, a->r_end)) {
1816                 /*
1817                  * The entry b is within the
1818                  * block a. i.e.:
1819                  * a --   |-------------|
1820                  * b --   |----|
1821                  * <or>
1822                  * b --       |------|
1823                  * <or>
1824                  * b --       |-----------|
1825                  */
1826                 return (0);
1827         } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1828                 /*
1829                  * b falls as either the next
1830                  * sequence block after a so a
1831                  * is said to be smaller than b.
1832                  * i.e:
1833                  * a --   |------|
1834                  * b --          |--------|
1835                  * or
1836                  * b --              |-----|
1837                  */
1838                 return (1);
1839         }
1840         /*
1841          * Whats left is where a is
1842          * larger than b. i.e:
1843          * a --         |-------|
1844          * b --  |---|
1845          * or even possibly
1846          * b --   |--------------|
1847          */
1848         return (-1);
1849 }
1850
1851 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1852 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1853
1854 static uint32_t
1855 rc_init_window(struct tcp_rack *rack)
1856 {
1857         uint32_t win;
1858
1859         if (rack->rc_init_win == 0) {
1860                 /*
1861                  * Nothing set by the user, use the system stack
1862                  * default.
1863                  */
1864                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1865         }
1866         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1867         return (win);
1868 }
1869
1870 static uint64_t
1871 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1872 {
1873         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1874                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1875         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1876                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1877         else
1878                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1879 }
1880
1881 static uint64_t
1882 rack_get_bw(struct tcp_rack *rack)
1883 {
1884         if (rack->use_fixed_rate) {
1885                 /* Return the fixed pacing rate */
1886                 return (rack_get_fixed_pacing_bw(rack));
1887         }
1888         if (rack->r_ctl.gp_bw == 0) {
1889                 /*
1890                  * We have yet no b/w measurement,
1891                  * if we have a user set initial bw
1892                  * return it. If we don't have that and
1893                  * we have an srtt, use the tcp IW (10) to
1894                  * calculate a fictional b/w over the SRTT
1895                  * which is more or less a guess. Note
1896                  * we don't use our IW from rack on purpose
1897                  * so if we have like IW=30, we are not
1898                  * calculating a "huge" b/w.
1899                  */
1900                 uint64_t bw, srtt;
1901                 if (rack->r_ctl.init_rate)
1902                         return (rack->r_ctl.init_rate);
1903
1904                 /* Has the user set a max peak rate? */
1905 #ifdef NETFLIX_PEAKRATE
1906                 if (rack->rc_tp->t_maxpeakrate)
1907                         return (rack->rc_tp->t_maxpeakrate);
1908 #endif
1909                 /* Ok lets come up with the IW guess, if we have a srtt */
1910                 if (rack->rc_tp->t_srtt == 0) {
1911                         /*
1912                          * Go with old pacing method
1913                          * i.e. burst mitigation only.
1914                          */
1915                         return (0);
1916                 }
1917                 /* Ok lets get the initial TCP win (not racks) */
1918                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1919                 srtt = (uint64_t)rack->rc_tp->t_srtt;
1920                 bw *= (uint64_t)USECS_IN_SECOND;
1921                 bw /= srtt;
1922                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1923                         bw = rack->r_ctl.bw_rate_cap;
1924                 return (bw);
1925         } else {
1926                 uint64_t bw;
1927
1928                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
1929                         /* Averaging is done, we can return the value */
1930                         bw = rack->r_ctl.gp_bw;
1931                 } else {
1932                         /* Still doing initial average must calculate */
1933                         bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
1934                 }
1935 #ifdef NETFLIX_PEAKRATE
1936                 if ((rack->rc_tp->t_maxpeakrate) &&
1937                     (bw > rack->rc_tp->t_maxpeakrate)) {
1938                         /* The user has set a peak rate to pace at
1939                          * don't allow us to pace faster than that.
1940                          */
1941                         return (rack->rc_tp->t_maxpeakrate);
1942                 }
1943 #endif
1944                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1945                         bw = rack->r_ctl.bw_rate_cap;
1946                 return (bw);
1947         }
1948 }
1949
1950 static uint16_t
1951 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1952 {
1953         if (rack->use_fixed_rate) {
1954                 return (100);
1955         } else if (rack->in_probe_rtt && (rsm == NULL))
1956                 return (rack->r_ctl.rack_per_of_gp_probertt);
1957         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
1958                   rack->r_ctl.rack_per_of_gp_rec)) {
1959                 if (rsm) {
1960                         /* a retransmission always use the recovery rate */
1961                         return (rack->r_ctl.rack_per_of_gp_rec);
1962                 } else if (rack->rack_rec_nonrxt_use_cr) {
1963                         /* Directed to use the configured rate */
1964                         goto configured_rate;
1965                 } else if (rack->rack_no_prr &&
1966                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1967                         /* No PRR, lets just use the b/w estimate only */
1968                         return (100);
1969                 } else {
1970                         /*
1971                          * Here we may have a non-retransmit but we
1972                          * have no overrides, so just use the recovery
1973                          * rate (prr is in effect).
1974                          */
1975                         return (rack->r_ctl.rack_per_of_gp_rec);
1976                 }
1977         }
1978 configured_rate:
1979         /* For the configured rate we look at our cwnd vs the ssthresh */
1980         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1981                 return (rack->r_ctl.rack_per_of_gp_ss);
1982         else
1983                 return (rack->r_ctl.rack_per_of_gp_ca);
1984 }
1985
1986 static void
1987 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
1988 {
1989         /*
1990          * Types of logs (mod value)
1991          * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
1992          * 2 = a dsack round begins, persist is reset to 16.
1993          * 3 = a dsack round ends
1994          * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
1995          * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
1996          * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
1997          */
1998         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1999                 union tcp_log_stackspecific log;
2000                 struct timeval tv;
2001
2002                 memset(&log, 0, sizeof(log));
2003                 log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2004                 log.u_bbr.flex1 <<= 1;
2005                 log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2006                 log.u_bbr.flex1 <<= 1;
2007                 log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2008                 log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2009                 log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2010                 log.u_bbr.flex4 = flex4;
2011                 log.u_bbr.flex5 = flex5;
2012                 log.u_bbr.flex6 = flex6;
2013                 log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2014                 log.u_bbr.flex8 = mod;
2015                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2016                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2017                     &rack->rc_inp->inp_socket->so_rcv,
2018                     &rack->rc_inp->inp_socket->so_snd,
2019                     RACK_DSACK_HANDLING, 0,
2020                     0, &log, false, &tv);
2021         }
2022 }
2023
2024 static void
2025 rack_log_hdwr_pacing(struct tcp_rack *rack,
2026                      uint64_t rate, uint64_t hw_rate, int line,
2027                      int error, uint16_t mod)
2028 {
2029         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2030                 union tcp_log_stackspecific log;
2031                 struct timeval tv;
2032                 const struct ifnet *ifp;
2033
2034                 memset(&log, 0, sizeof(log));
2035                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2036                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2037                 if (rack->r_ctl.crte) {
2038                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2039                 } else if (rack->rc_inp->inp_route.ro_nh &&
2040                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2041                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2042                 } else
2043                         ifp = NULL;
2044                 if (ifp) {
2045                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2046                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2047                 }
2048                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2049                 log.u_bbr.bw_inuse = rate;
2050                 log.u_bbr.flex5 = line;
2051                 log.u_bbr.flex6 = error;
2052                 log.u_bbr.flex7 = mod;
2053                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2054                 log.u_bbr.flex8 = rack->use_fixed_rate;
2055                 log.u_bbr.flex8 <<= 1;
2056                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2057                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2058                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2059                 if (rack->r_ctl.crte)
2060                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2061                 else
2062                         log.u_bbr.cur_del_rate = 0;
2063                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2064                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2065                     &rack->rc_inp->inp_socket->so_rcv,
2066                     &rack->rc_inp->inp_socket->so_snd,
2067                     BBR_LOG_HDWR_PACE, 0,
2068                     0, &log, false, &tv);
2069         }
2070 }
2071
2072 static uint64_t
2073 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2074 {
2075         /*
2076          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2077          */
2078         uint64_t bw_est, high_rate;
2079         uint64_t gain;
2080
2081         gain = (uint64_t)rack_get_output_gain(rack, rsm);
2082         bw_est = bw * gain;
2083         bw_est /= (uint64_t)100;
2084         /* Never fall below the minimum (def 64kbps) */
2085         if (bw_est < RACK_MIN_BW)
2086                 bw_est = RACK_MIN_BW;
2087         if (rack->r_rack_hw_rate_caps) {
2088                 /* Rate caps are in place */
2089                 if (rack->r_ctl.crte != NULL) {
2090                         /* We have a hdwr rate already */
2091                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2092                         if (bw_est >= high_rate) {
2093                                 /* We are capping bw at the highest rate table entry */
2094                                 rack_log_hdwr_pacing(rack,
2095                                                      bw_est, high_rate, __LINE__,
2096                                                      0, 3);
2097                                 bw_est = high_rate;
2098                                 if (capped)
2099                                         *capped = 1;
2100                         }
2101                 } else if ((rack->rack_hdrw_pacing == 0) &&
2102                            (rack->rack_hdw_pace_ena) &&
2103                            (rack->rack_attempt_hdwr_pace == 0) &&
2104                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2105                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2106                         /*
2107                          * Special case, we have not yet attempted hardware
2108                          * pacing, and yet we may, when we do, find out if we are
2109                          * above the highest rate. We need to know the maxbw for the interface
2110                          * in question (if it supports ratelimiting). We get back
2111                          * a 0, if the interface is not found in the RL lists.
2112                          */
2113                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2114                         if (high_rate) {
2115                                 /* Yep, we have a rate is it above this rate? */
2116                                 if (bw_est > high_rate) {
2117                                         bw_est = high_rate;
2118                                         if (capped)
2119                                                 *capped = 1;
2120                                 }
2121                         }
2122                 }
2123         }
2124         return (bw_est);
2125 }
2126
2127 static void
2128 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2129 {
2130         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2131                 union tcp_log_stackspecific log;
2132                 struct timeval tv;
2133
2134                 if ((mod != 1) && (rack_verbose_logging == 0)) {
2135                         /*
2136                          * We get 3 values currently for mod
2137                          * 1 - We are retransmitting and this tells the reason.
2138                          * 2 - We are clearing a dup-ack count.
2139                          * 3 - We are incrementing a dup-ack count.
2140                          *
2141                          * The clear/increment are only logged
2142                          * if you have BBverbose on.
2143                          */
2144                         return;
2145                 }
2146                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2147                 log.u_bbr.flex1 = tsused;
2148                 log.u_bbr.flex2 = thresh;
2149                 log.u_bbr.flex3 = rsm->r_flags;
2150                 log.u_bbr.flex4 = rsm->r_dupack;
2151                 log.u_bbr.flex5 = rsm->r_start;
2152                 log.u_bbr.flex6 = rsm->r_end;
2153                 log.u_bbr.flex8 = mod;
2154                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2155                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2156                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2157                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2158                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2159                 log.u_bbr.pacing_gain = rack->r_must_retran;
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_SETTINGS_CHG, 0,
2164                     0, &log, false, &tv);
2165         }
2166 }
2167
2168 static void
2169 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
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.flex1 = rack->rc_tp->t_srtt;
2177                 log.u_bbr.flex2 = to;
2178                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2179                 log.u_bbr.flex4 = slot;
2180                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2181                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2182                 log.u_bbr.flex7 = rack->rc_in_persist;
2183                 log.u_bbr.flex8 = which;
2184                 if (rack->rack_no_prr)
2185                         log.u_bbr.pkts_out = 0;
2186                 else
2187                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2188                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2189                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2190                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2191                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2192                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2193                 log.u_bbr.pacing_gain = rack->r_must_retran;
2194                 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2195                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2196                 log.u_bbr.lost = rack_rto_min;
2197                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2198                     &rack->rc_inp->inp_socket->so_rcv,
2199                     &rack->rc_inp->inp_socket->so_snd,
2200                     BBR_LOG_TIMERSTAR, 0,
2201                     0, &log, false, &tv);
2202         }
2203 }
2204
2205 static void
2206 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2207 {
2208         if (rack->rc_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.inhpts = tcp_in_hpts(rack->rc_inp);
2214                 log.u_bbr.flex8 = to_num;
2215                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2216                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2217                 if (rsm == NULL)
2218                         log.u_bbr.flex3 = 0;
2219                 else
2220                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2221                 if (rack->rack_no_prr)
2222                         log.u_bbr.flex5 = 0;
2223                 else
2224                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2225                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2226                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2227                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2228                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2229                 log.u_bbr.pacing_gain = rack->r_must_retran;
2230                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2231                     &rack->rc_inp->inp_socket->so_rcv,
2232                     &rack->rc_inp->inp_socket->so_snd,
2233                     BBR_LOG_RTO, 0,
2234                     0, &log, false, &tv);
2235         }
2236 }
2237
2238 static void
2239 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2240                  struct rack_sendmap *prev,
2241                  struct rack_sendmap *rsm,
2242                  struct rack_sendmap *next,
2243                  int flag, uint32_t th_ack, int line)
2244 {
2245         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2246                 union tcp_log_stackspecific log;
2247                 struct timeval tv;
2248
2249                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2250                 log.u_bbr.flex8 = flag;
2251                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2252                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2253                 log.u_bbr.delRate = (uint64_t)rsm;
2254                 log.u_bbr.rttProp = (uint64_t)next;
2255                 log.u_bbr.flex7 = 0;
2256                 if (prev) {
2257                         log.u_bbr.flex1 = prev->r_start;
2258                         log.u_bbr.flex2 = prev->r_end;
2259                         log.u_bbr.flex7 |= 0x4;
2260                 }
2261                 if (rsm) {
2262                         log.u_bbr.flex3 = rsm->r_start;
2263                         log.u_bbr.flex4 = rsm->r_end;
2264                         log.u_bbr.flex7 |= 0x2;
2265                 }
2266                 if (next) {
2267                         log.u_bbr.flex5 = next->r_start;
2268                         log.u_bbr.flex6 = next->r_end;
2269                         log.u_bbr.flex7 |= 0x1;
2270                 }
2271                 log.u_bbr.applimited = line;
2272                 log.u_bbr.pkts_out = th_ack;
2273                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2274                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2275                 if (rack->rack_no_prr)
2276                         log.u_bbr.lost = 0;
2277                 else
2278                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2279                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2280                     &rack->rc_inp->inp_socket->so_rcv,
2281                     &rack->rc_inp->inp_socket->so_snd,
2282                     TCP_LOG_MAPCHG, 0,
2283                     0, &log, false, &tv);
2284         }
2285 }
2286
2287 static void
2288 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2289                  struct rack_sendmap *rsm, int conf)
2290 {
2291         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2292                 union tcp_log_stackspecific log;
2293                 struct timeval tv;
2294                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2295                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2296                 log.u_bbr.flex1 = t;
2297                 log.u_bbr.flex2 = len;
2298                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2299                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2300                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2301                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2302                 log.u_bbr.flex7 = conf;
2303                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2304                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2305                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2306                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2307                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2308                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2309                 if (rsm) {
2310                         log.u_bbr.pkt_epoch = rsm->r_start;
2311                         log.u_bbr.lost = rsm->r_end;
2312                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2313                         /* We loose any upper of the 24 bits */
2314                         log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2315                 } else {
2316                         /* Its a SYN */
2317                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2318                         log.u_bbr.lost = 0;
2319                         log.u_bbr.cwnd_gain = 0;
2320                         log.u_bbr.pacing_gain = 0;
2321                 }
2322                 /* Write out general bits of interest rrs here */
2323                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2324                 log.u_bbr.use_lt_bw <<= 1;
2325                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2326                 log.u_bbr.use_lt_bw <<= 1;
2327                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2328                 log.u_bbr.use_lt_bw <<= 1;
2329                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2330                 log.u_bbr.use_lt_bw <<= 1;
2331                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2332                 log.u_bbr.use_lt_bw <<= 1;
2333                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2334                 log.u_bbr.use_lt_bw <<= 1;
2335                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2336                 log.u_bbr.use_lt_bw <<= 1;
2337                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2338                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2339                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2340                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2341                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2342                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2343                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2344                 log.u_bbr.bw_inuse <<= 32;
2345                 if (rsm)
2346                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2347                 TCP_LOG_EVENTP(tp, NULL,
2348                     &rack->rc_inp->inp_socket->so_rcv,
2349                     &rack->rc_inp->inp_socket->so_snd,
2350                     BBR_LOG_BBRRTT, 0,
2351                     0, &log, false, &tv);
2352
2353
2354         }
2355 }
2356
2357 static void
2358 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2359 {
2360         /*
2361          * Log the rtt sample we are
2362          * applying to the srtt algorithm in
2363          * useconds.
2364          */
2365         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2366                 union tcp_log_stackspecific log;
2367                 struct timeval tv;
2368
2369                 /* Convert our ms to a microsecond */
2370                 memset(&log, 0, sizeof(log));
2371                 log.u_bbr.flex1 = rtt;
2372                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2373                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2374                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2375                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2376                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2377                 log.u_bbr.flex7 = 1;
2378                 log.u_bbr.flex8 = rack->sack_attack_disable;
2379                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2380                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2381                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2382                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2383                 log.u_bbr.pacing_gain = rack->r_must_retran;
2384                 /*
2385                  * We capture in delRate the upper 32 bits as
2386                  * the confidence level we had declared, and the
2387                  * lower 32 bits as the actual RTT using the arrival
2388                  * timestamp.
2389                  */
2390                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2391                 log.u_bbr.delRate <<= 32;
2392                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2393                 /* Lets capture all the things that make up t_rtxcur */
2394                 log.u_bbr.applimited = rack_rto_min;
2395                 log.u_bbr.epoch = rack_rto_max;
2396                 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2397                 log.u_bbr.lost = rack_rto_min;
2398                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2399                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2400                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2401                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2402                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2403                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2404                     &rack->rc_inp->inp_socket->so_rcv,
2405                     &rack->rc_inp->inp_socket->so_snd,
2406                     TCP_LOG_RTT, 0,
2407                     0, &log, false, &tv);
2408         }
2409 }
2410
2411 static void
2412 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2413 {
2414         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2415                 union tcp_log_stackspecific log;
2416                 struct timeval tv;
2417
2418                 /* Convert our ms to a microsecond */
2419                 memset(&log, 0, sizeof(log));
2420                 log.u_bbr.flex1 = rtt;
2421                 log.u_bbr.flex2 = send_time;
2422                 log.u_bbr.flex3 = ack_time;
2423                 log.u_bbr.flex4 = where;
2424                 log.u_bbr.flex7 = 2;
2425                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2426                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2427                     &rack->rc_inp->inp_socket->so_rcv,
2428                     &rack->rc_inp->inp_socket->so_snd,
2429                     TCP_LOG_RTT, 0,
2430                     0, &log, false, &tv);
2431         }
2432 }
2433
2434
2435
2436 static inline void
2437 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2438 {
2439         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2440                 union tcp_log_stackspecific log;
2441                 struct timeval tv;
2442
2443                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2444                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2445                 log.u_bbr.flex1 = line;
2446                 log.u_bbr.flex2 = tick;
2447                 log.u_bbr.flex3 = tp->t_maxunacktime;
2448                 log.u_bbr.flex4 = tp->t_acktime;
2449                 log.u_bbr.flex8 = event;
2450                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2451                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2452                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2453                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2454                 log.u_bbr.pacing_gain = rack->r_must_retran;
2455                 TCP_LOG_EVENTP(tp, NULL,
2456                     &rack->rc_inp->inp_socket->so_rcv,
2457                     &rack->rc_inp->inp_socket->so_snd,
2458                     BBR_LOG_PROGRESS, 0,
2459                     0, &log, false, &tv);
2460         }
2461 }
2462
2463 static void
2464 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2465 {
2466         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2467                 union tcp_log_stackspecific log;
2468
2469                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2470                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2471                 log.u_bbr.flex1 = slot;
2472                 if (rack->rack_no_prr)
2473                         log.u_bbr.flex2 = 0;
2474                 else
2475                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2476                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2477                 log.u_bbr.flex8 = rack->rc_in_persist;
2478                 log.u_bbr.timeStamp = cts;
2479                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2480                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2481                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2482                 log.u_bbr.pacing_gain = rack->r_must_retran;
2483                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2484                     &rack->rc_inp->inp_socket->so_rcv,
2485                     &rack->rc_inp->inp_socket->so_snd,
2486                     BBR_LOG_BBRSND, 0,
2487                     0, &log, false, tv);
2488         }
2489 }
2490
2491 static void
2492 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2493 {
2494         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2495                 union tcp_log_stackspecific log;
2496                 struct timeval tv;
2497
2498                 memset(&log, 0, sizeof(log));
2499                 log.u_bbr.flex1 = did_out;
2500                 log.u_bbr.flex2 = nxt_pkt;
2501                 log.u_bbr.flex3 = way_out;
2502                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2503                 if (rack->rack_no_prr)
2504                         log.u_bbr.flex5 = 0;
2505                 else
2506                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2507                 log.u_bbr.flex6 = nsegs;
2508                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2509                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2510                 log.u_bbr.flex7 <<= 1;
2511                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2512                 log.u_bbr.flex7 <<= 1;
2513                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2514                 log.u_bbr.flex8 = rack->rc_in_persist;
2515                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2516                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2517                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2518                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2519                 log.u_bbr.use_lt_bw <<= 1;
2520                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2521                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2522                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2523                 log.u_bbr.pacing_gain = rack->r_must_retran;
2524                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2525                     &rack->rc_inp->inp_socket->so_rcv,
2526                     &rack->rc_inp->inp_socket->so_snd,
2527                     BBR_LOG_DOSEG_DONE, 0,
2528                     0, &log, false, &tv);
2529         }
2530 }
2531
2532 static void
2533 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2534 {
2535         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2536                 union tcp_log_stackspecific log;
2537                 struct timeval tv;
2538
2539                 memset(&log, 0, sizeof(log));
2540                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2541                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2542                 log.u_bbr.flex4 = arg1;
2543                 log.u_bbr.flex5 = arg2;
2544                 log.u_bbr.flex6 = arg3;
2545                 log.u_bbr.flex8 = frm;
2546                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2547                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2548                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2549                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2550                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2551                 log.u_bbr.pacing_gain = rack->r_must_retran;
2552                 TCP_LOG_EVENTP(tp, NULL,
2553                     &tp->t_inpcb->inp_socket->so_rcv,
2554                     &tp->t_inpcb->inp_socket->so_snd,
2555                     TCP_HDWR_PACE_SIZE, 0,
2556                     0, &log, false, &tv);
2557         }
2558 }
2559
2560 static void
2561 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2562                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2563 {
2564         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2565                 union tcp_log_stackspecific log;
2566                 struct timeval tv;
2567
2568                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2569                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2570                 log.u_bbr.flex1 = slot;
2571                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2572                 log.u_bbr.flex4 = reason;
2573                 if (rack->rack_no_prr)
2574                         log.u_bbr.flex5 = 0;
2575                 else
2576                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2577                 log.u_bbr.flex7 = hpts_calling;
2578                 log.u_bbr.flex8 = rack->rc_in_persist;
2579                 log.u_bbr.lt_epoch = cwnd_to_use;
2580                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2581                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2582                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2583                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2584                 log.u_bbr.pacing_gain = rack->r_must_retran;
2585                 log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2586                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2587                     &rack->rc_inp->inp_socket->so_rcv,
2588                     &rack->rc_inp->inp_socket->so_snd,
2589                     BBR_LOG_JUSTRET, 0,
2590                     tlen, &log, false, &tv);
2591         }
2592 }
2593
2594 static void
2595 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2596                    struct timeval *tv, uint32_t flags_on_entry)
2597 {
2598         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2599                 union tcp_log_stackspecific log;
2600
2601                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2602                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2603                 log.u_bbr.flex1 = line;
2604                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2605                 log.u_bbr.flex3 = flags_on_entry;
2606                 log.u_bbr.flex4 = us_cts;
2607                 if (rack->rack_no_prr)
2608                         log.u_bbr.flex5 = 0;
2609                 else
2610                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2611                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2612                 log.u_bbr.flex7 = hpts_removed;
2613                 log.u_bbr.flex8 = 1;
2614                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2615                 log.u_bbr.timeStamp = us_cts;
2616                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2617                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2618                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2619                 log.u_bbr.pacing_gain = rack->r_must_retran;
2620                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2621                     &rack->rc_inp->inp_socket->so_rcv,
2622                     &rack->rc_inp->inp_socket->so_snd,
2623                     BBR_LOG_TIMERCANC, 0,
2624                     0, &log, false, tv);
2625         }
2626 }
2627
2628 static void
2629 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2630                           uint32_t flex1, uint32_t flex2,
2631                           uint32_t flex3, uint32_t flex4,
2632                           uint32_t flex5, uint32_t flex6,
2633                           uint16_t flex7, uint8_t mod)
2634 {
2635         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2636                 union tcp_log_stackspecific log;
2637                 struct timeval tv;
2638
2639                 if (mod == 1) {
2640                         /* No you can't use 1, its for the real to cancel */
2641                         return;
2642                 }
2643                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2644                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2645                 log.u_bbr.flex1 = flex1;
2646                 log.u_bbr.flex2 = flex2;
2647                 log.u_bbr.flex3 = flex3;
2648                 log.u_bbr.flex4 = flex4;
2649                 log.u_bbr.flex5 = flex5;
2650                 log.u_bbr.flex6 = flex6;
2651                 log.u_bbr.flex7 = flex7;
2652                 log.u_bbr.flex8 = mod;
2653                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2654                     &rack->rc_inp->inp_socket->so_rcv,
2655                     &rack->rc_inp->inp_socket->so_snd,
2656                     BBR_LOG_TIMERCANC, 0,
2657                     0, &log, false, &tv);
2658         }
2659 }
2660
2661 static void
2662 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2663 {
2664         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2665                 union tcp_log_stackspecific log;
2666                 struct timeval tv;
2667
2668                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2669                 log.u_bbr.flex1 = timers;
2670                 log.u_bbr.flex2 = ret;
2671                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2672                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2673                 log.u_bbr.flex5 = cts;
2674                 if (rack->rack_no_prr)
2675                         log.u_bbr.flex6 = 0;
2676                 else
2677                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2678                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2679                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2680                 log.u_bbr.pacing_gain = rack->r_must_retran;
2681                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2682                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2683                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2684                     &rack->rc_inp->inp_socket->so_rcv,
2685                     &rack->rc_inp->inp_socket->so_snd,
2686                     BBR_LOG_TO_PROCESS, 0,
2687                     0, &log, false, &tv);
2688         }
2689 }
2690
2691 static void
2692 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
2693 {
2694         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2695                 union tcp_log_stackspecific log;
2696                 struct timeval tv;
2697
2698                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2699                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2700                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2701                 if (rack->rack_no_prr)
2702                         log.u_bbr.flex3 = 0;
2703                 else
2704                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2705                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2706                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2707                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2708                 log.u_bbr.flex7 = line;
2709                 log.u_bbr.flex8 = frm;
2710                 log.u_bbr.pkts_out = orig_cwnd;
2711                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2712                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2713                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2714                 log.u_bbr.use_lt_bw <<= 1;
2715                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2716                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2717                     &rack->rc_inp->inp_socket->so_rcv,
2718                     &rack->rc_inp->inp_socket->so_snd,
2719                     BBR_LOG_BBRUPD, 0,
2720                     0, &log, false, &tv);
2721         }
2722 }
2723
2724 #ifdef NETFLIX_EXP_DETECTION
2725 static void
2726 rack_log_sad(struct tcp_rack *rack, int event)
2727 {
2728         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2729                 union tcp_log_stackspecific log;
2730                 struct timeval tv;
2731
2732                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2733                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2734                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2735                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2736                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2737                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2738                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2739                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2740                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2741                 log.u_bbr.lt_epoch |= rack->do_detection;
2742                 log.u_bbr.applimited = tcp_map_minimum;
2743                 log.u_bbr.flex7 = rack->sack_attack_disable;
2744                 log.u_bbr.flex8 = event;
2745                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2746                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2747                 log.u_bbr.delivered = tcp_sad_decay_val;
2748                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2749                     &rack->rc_inp->inp_socket->so_rcv,
2750                     &rack->rc_inp->inp_socket->so_snd,
2751                     TCP_SAD_DETECTION, 0,
2752                     0, &log, false, &tv);
2753         }
2754 }
2755 #endif
2756
2757 static void
2758 rack_counter_destroy(void)
2759 {
2760         counter_u64_free(rack_fto_send);
2761         counter_u64_free(rack_fto_rsm_send);
2762         counter_u64_free(rack_nfto_resend);
2763         counter_u64_free(rack_hw_pace_init_fail);
2764         counter_u64_free(rack_hw_pace_lost);
2765         counter_u64_free(rack_non_fto_send);
2766         counter_u64_free(rack_extended_rfo);
2767         counter_u64_free(rack_ack_total);
2768         counter_u64_free(rack_express_sack);
2769         counter_u64_free(rack_sack_total);
2770         counter_u64_free(rack_move_none);
2771         counter_u64_free(rack_move_some);
2772         counter_u64_free(rack_sack_attacks_detected);
2773         counter_u64_free(rack_sack_attacks_reversed);
2774         counter_u64_free(rack_sack_used_next_merge);
2775         counter_u64_free(rack_sack_used_prev_merge);
2776         counter_u64_free(rack_tlp_tot);
2777         counter_u64_free(rack_tlp_newdata);
2778         counter_u64_free(rack_tlp_retran);
2779         counter_u64_free(rack_tlp_retran_bytes);
2780         counter_u64_free(rack_to_tot);
2781         counter_u64_free(rack_saw_enobuf);
2782         counter_u64_free(rack_saw_enobuf_hw);
2783         counter_u64_free(rack_saw_enetunreach);
2784         counter_u64_free(rack_hot_alloc);
2785         counter_u64_free(rack_to_alloc);
2786         counter_u64_free(rack_to_alloc_hard);
2787         counter_u64_free(rack_to_alloc_emerg);
2788         counter_u64_free(rack_to_alloc_limited);
2789         counter_u64_free(rack_alloc_limited_conns);
2790         counter_u64_free(rack_split_limited);
2791         counter_u64_free(rack_multi_single_eq);
2792         counter_u64_free(rack_proc_non_comp_ack);
2793         counter_u64_free(rack_sack_proc_all);
2794         counter_u64_free(rack_sack_proc_restart);
2795         counter_u64_free(rack_sack_proc_short);
2796         counter_u64_free(rack_sack_skipped_acked);
2797         counter_u64_free(rack_sack_splits);
2798         counter_u64_free(rack_input_idle_reduces);
2799         counter_u64_free(rack_collapsed_win);
2800         counter_u64_free(rack_collapsed_win_rxt);
2801         counter_u64_free(rack_collapsed_win_rxt_bytes);
2802         counter_u64_free(rack_collapsed_win_seen);
2803         counter_u64_free(rack_try_scwnd);
2804         counter_u64_free(rack_persists_sends);
2805         counter_u64_free(rack_persists_acks);
2806         counter_u64_free(rack_persists_loss);
2807         counter_u64_free(rack_persists_lost_ends);
2808 #ifdef INVARIANTS
2809         counter_u64_free(rack_adjust_map_bw);
2810 #endif
2811         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2812         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2813 }
2814
2815 static struct rack_sendmap *
2816 rack_alloc(struct tcp_rack *rack)
2817 {
2818         struct rack_sendmap *rsm;
2819
2820         /*
2821          * First get the top of the list it in
2822          * theory is the "hottest" rsm we have,
2823          * possibly just freed by ack processing.
2824          */
2825         if (rack->rc_free_cnt > rack_free_cache) {
2826                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2827                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2828                 counter_u64_add(rack_hot_alloc, 1);
2829                 rack->rc_free_cnt--;
2830                 return (rsm);
2831         }
2832         /*
2833          * Once we get under our free cache we probably
2834          * no longer have a "hot" one available. Lets
2835          * get one from UMA.
2836          */
2837         rsm = uma_zalloc(rack_zone, M_NOWAIT);
2838         if (rsm) {
2839                 rack->r_ctl.rc_num_maps_alloced++;
2840                 counter_u64_add(rack_to_alloc, 1);
2841                 return (rsm);
2842         }
2843         /*
2844          * Dig in to our aux rsm's (the last two) since
2845          * UMA failed to get us one.
2846          */
2847         if (rack->rc_free_cnt) {
2848                 counter_u64_add(rack_to_alloc_emerg, 1);
2849                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2850                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2851                 rack->rc_free_cnt--;
2852                 return (rsm);
2853         }
2854         return (NULL);
2855 }
2856
2857 static struct rack_sendmap *
2858 rack_alloc_full_limit(struct tcp_rack *rack)
2859 {
2860         if ((V_tcp_map_entries_limit > 0) &&
2861             (rack->do_detection == 0) &&
2862             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2863                 counter_u64_add(rack_to_alloc_limited, 1);
2864                 if (!rack->alloc_limit_reported) {
2865                         rack->alloc_limit_reported = 1;
2866                         counter_u64_add(rack_alloc_limited_conns, 1);
2867                 }
2868                 return (NULL);
2869         }
2870         return (rack_alloc(rack));
2871 }
2872
2873 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2874 static struct rack_sendmap *
2875 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2876 {
2877         struct rack_sendmap *rsm;
2878
2879         if (limit_type) {
2880                 /* currently there is only one limit type */
2881                 if (V_tcp_map_split_limit > 0 &&
2882                     (rack->do_detection == 0) &&
2883                     rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2884                         counter_u64_add(rack_split_limited, 1);
2885                         if (!rack->alloc_limit_reported) {
2886                                 rack->alloc_limit_reported = 1;
2887                                 counter_u64_add(rack_alloc_limited_conns, 1);
2888                         }
2889                         return (NULL);
2890                 }
2891         }
2892
2893         /* allocate and mark in the limit type, if set */
2894         rsm = rack_alloc(rack);
2895         if (rsm != NULL && limit_type) {
2896                 rsm->r_limit_type = limit_type;
2897                 rack->r_ctl.rc_num_split_allocs++;
2898         }
2899         return (rsm);
2900 }
2901
2902 static void
2903 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2904 {
2905         if (rsm->r_flags & RACK_APP_LIMITED) {
2906                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2907                         rack->r_ctl.rc_app_limited_cnt--;
2908                 }
2909         }
2910         if (rsm->r_limit_type) {
2911                 /* currently there is only one limit type */
2912                 rack->r_ctl.rc_num_split_allocs--;
2913         }
2914         if (rsm == rack->r_ctl.rc_first_appl) {
2915                 if (rack->r_ctl.rc_app_limited_cnt == 0)
2916                         rack->r_ctl.rc_first_appl = NULL;
2917                 else {
2918                         /* Follow the next one out */
2919                         struct rack_sendmap fe;
2920
2921                         fe.r_start = rsm->r_nseq_appl;
2922                         rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2923                 }
2924         }
2925         if (rsm == rack->r_ctl.rc_resend)
2926                 rack->r_ctl.rc_resend = NULL;
2927         if (rsm == rack->r_ctl.rc_end_appl)
2928                 rack->r_ctl.rc_end_appl = NULL;
2929         if (rack->r_ctl.rc_tlpsend == rsm)
2930                 rack->r_ctl.rc_tlpsend = NULL;
2931         if (rack->r_ctl.rc_sacklast == rsm)
2932                 rack->r_ctl.rc_sacklast = NULL;
2933         memset(rsm, 0, sizeof(struct rack_sendmap));
2934         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
2935         rack->rc_free_cnt++;
2936 }
2937
2938 static void
2939 rack_free_trim(struct tcp_rack *rack)
2940 {
2941         struct rack_sendmap *rsm;
2942
2943         /*
2944          * Free up all the tail entries until
2945          * we get our list down to the limit.
2946          */
2947         while (rack->rc_free_cnt > rack_free_cache) {
2948                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
2949                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2950                 rack->rc_free_cnt--;
2951                 uma_zfree(rack_zone, rsm);
2952         }
2953 }
2954
2955
2956 static uint32_t
2957 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2958 {
2959         uint64_t srtt, bw, len, tim;
2960         uint32_t segsiz, def_len, minl;
2961
2962         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2963         def_len = rack_def_data_window * segsiz;
2964         if (rack->rc_gp_filled == 0) {
2965                 /*
2966                  * We have no measurement (IW is in flight?) so
2967                  * we can only guess using our data_window sysctl
2968                  * value (usually 20MSS).
2969                  */
2970                 return (def_len);
2971         }
2972         /*
2973          * Now we have a number of factors to consider.
2974          *
2975          * 1) We have a desired BDP which is usually
2976          *    at least 2.
2977          * 2) We have a minimum number of rtt's usually 1 SRTT
2978          *    but we allow it too to be more.
2979          * 3) We want to make sure a measurement last N useconds (if
2980          *    we have set rack_min_measure_usec.
2981          *
2982          * We handle the first concern here by trying to create a data
2983          * window of max(rack_def_data_window, DesiredBDP). The
2984          * second concern we handle in not letting the measurement
2985          * window end normally until at least the required SRTT's
2986          * have gone by which is done further below in
2987          * rack_enough_for_measurement(). Finally the third concern
2988          * we also handle here by calculating how long that time
2989          * would take at the current BW and then return the
2990          * max of our first calculation and that length. Note
2991          * that if rack_min_measure_usec is 0, we don't deal
2992          * with concern 3. Also for both Concern 1 and 3 an
2993          * application limited period could end the measurement
2994          * earlier.
2995          *
2996          * So lets calculate the BDP with the "known" b/w using
2997          * the SRTT has our rtt and then multiply it by the
2998          * goal.
2999          */
3000         bw = rack_get_bw(rack);
3001         srtt = (uint64_t)tp->t_srtt;
3002         len = bw * srtt;
3003         len /= (uint64_t)HPTS_USEC_IN_SEC;
3004         len *= max(1, rack_goal_bdp);
3005         /* Now we need to round up to the nearest MSS */
3006         len = roundup(len, segsiz);
3007         if (rack_min_measure_usec) {
3008                 /* Now calculate our min length for this b/w */
3009                 tim = rack_min_measure_usec;
3010                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3011                 if (minl == 0)
3012                         minl = 1;
3013                 minl = roundup(minl, segsiz);
3014                 if (len < minl)
3015                         len = minl;
3016         }
3017         /*
3018          * Now if we have a very small window we want
3019          * to attempt to get the window that is
3020          * as small as possible. This happens on
3021          * low b/w connections and we don't want to
3022          * span huge numbers of rtt's between measurements.
3023          *
3024          * We basically include 2 over our "MIN window" so
3025          * that the measurement can be shortened (possibly) by
3026          * an ack'ed packet.
3027          */
3028         if (len < def_len)
3029                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3030         else
3031                 return (max((uint32_t)len, def_len));
3032
3033 }
3034
3035 static int
3036 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3037 {
3038         uint32_t tim, srtts, segsiz;
3039
3040         /*
3041          * Has enough time passed for the GP measurement to be valid?
3042          */
3043         if ((tp->snd_max == tp->snd_una) ||
3044             (th_ack == tp->snd_max)){
3045                 /* All is acked */
3046                 *quality = RACK_QUALITY_ALLACKED;
3047                 return (1);
3048         }
3049         if (SEQ_LT(th_ack, tp->gput_seq)) {
3050                 /* Not enough bytes yet */
3051                 return (0);
3052         }
3053         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3054         if (SEQ_LT(th_ack, tp->gput_ack) &&
3055             ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3056                 /* Not enough bytes yet */
3057                 return (0);
3058         }
3059         if (rack->r_ctl.rc_first_appl &&
3060             (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3061                 /*
3062                  * We are up to the app limited send point
3063                  * we have to measure irrespective of the time..
3064                  */
3065                 *quality = RACK_QUALITY_APPLIMITED;
3066                 return (1);
3067         }
3068         /* Now what about time? */
3069         srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3070         tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3071         if (tim >= srtts) {
3072                 *quality = RACK_QUALITY_HIGH;
3073                 return (1);
3074         }
3075         /* Nope not even a full SRTT has passed */
3076         return (0);
3077 }
3078
3079 static void
3080 rack_log_timely(struct tcp_rack *rack,
3081                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3082                 uint64_t up_bnd, int line, uint8_t method)
3083 {
3084         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3085                 union tcp_log_stackspecific log;
3086                 struct timeval tv;
3087
3088                 memset(&log, 0, sizeof(log));
3089                 log.u_bbr.flex1 = logged;
3090                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3091                 log.u_bbr.flex2 <<= 4;
3092                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3093                 log.u_bbr.flex2 <<= 4;
3094                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3095                 log.u_bbr.flex2 <<= 4;
3096                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3097                 log.u_bbr.flex3 = rack->rc_gp_incr;
3098                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3099                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3100                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3101                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3102                 log.u_bbr.flex8 = method;
3103                 log.u_bbr.cur_del_rate = cur_bw;
3104                 log.u_bbr.delRate = low_bnd;
3105                 log.u_bbr.bw_inuse = up_bnd;
3106                 log.u_bbr.rttProp = rack_get_bw(rack);
3107                 log.u_bbr.pkt_epoch = line;
3108                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3109                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3110                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3111                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3112                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3113                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3114                 log.u_bbr.cwnd_gain <<= 1;
3115                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3116                 log.u_bbr.cwnd_gain <<= 1;
3117                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3118                 log.u_bbr.cwnd_gain <<= 1;
3119                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3120                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3121                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3122                     &rack->rc_inp->inp_socket->so_rcv,
3123                     &rack->rc_inp->inp_socket->so_snd,
3124                     TCP_TIMELY_WORK, 0,
3125                     0, &log, false, &tv);
3126         }
3127 }
3128
3129 static int
3130 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3131 {
3132         /*
3133          * Before we increase we need to know if
3134          * the estimate just made was less than
3135          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3136          *
3137          * If we already are pacing at a fast enough
3138          * rate to push us faster there is no sense of
3139          * increasing.
3140          *
3141          * We first caculate our actual pacing rate (ss or ca multiplier
3142          * times our cur_bw).
3143          *
3144          * Then we take the last measured rate and multipy by our
3145          * maximum pacing overage to give us a max allowable rate.
3146          *
3147          * If our act_rate is smaller than our max_allowable rate
3148          * then we should increase. Else we should hold steady.
3149          *
3150          */
3151         uint64_t act_rate, max_allow_rate;
3152
3153         if (rack_timely_no_stopping)
3154                 return (1);
3155
3156         if ((cur_bw == 0) || (last_bw_est == 0)) {
3157                 /*
3158                  * Initial startup case or
3159                  * everything is acked case.
3160                  */
3161                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3162                                 __LINE__, 9);
3163                 return (1);
3164         }
3165         if (mult <= 100) {
3166                 /*
3167                  * We can always pace at or slightly above our rate.
3168                  */
3169                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3170                                 __LINE__, 9);
3171                 return (1);
3172         }
3173         act_rate = cur_bw * (uint64_t)mult;
3174         act_rate /= 100;
3175         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3176         max_allow_rate /= 100;
3177         if (act_rate < max_allow_rate) {
3178                 /*
3179                  * Here the rate we are actually pacing at
3180                  * is smaller than 10% above our last measurement.
3181                  * This means we are pacing below what we would
3182                  * like to try to achieve (plus some wiggle room).
3183                  */
3184                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3185                                 __LINE__, 9);
3186                 return (1);
3187         } else {
3188                 /*
3189                  * Here we are already pacing at least rack_max_per_above(10%)
3190                  * what we are getting back. This indicates most likely
3191                  * that we are being limited (cwnd/rwnd/app) and can't
3192                  * get any more b/w. There is no sense of trying to
3193                  * raise up the pacing rate its not speeding us up
3194                  * and we already are pacing faster than we are getting.
3195                  */
3196                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3197                                 __LINE__, 8);
3198                 return (0);
3199         }
3200 }
3201
3202 static void
3203 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3204 {
3205         /*
3206          * When we drag bottom, we want to assure
3207          * that no multiplier is below 1.0, if so
3208          * we want to restore it to at least that.
3209          */
3210         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3211                 /* This is unlikely we usually do not touch recovery */
3212                 rack->r_ctl.rack_per_of_gp_rec = 100;
3213         }
3214         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3215                 rack->r_ctl.rack_per_of_gp_ca = 100;
3216         }
3217         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3218                 rack->r_ctl.rack_per_of_gp_ss = 100;
3219         }
3220 }
3221
3222 static void
3223 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3224 {
3225         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3226                 rack->r_ctl.rack_per_of_gp_ca = 100;
3227         }
3228         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3229                 rack->r_ctl.rack_per_of_gp_ss = 100;
3230         }
3231 }
3232
3233 static void
3234 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3235 {
3236         int32_t  calc, logged, plus;
3237
3238         logged = 0;
3239
3240         if (override) {
3241                 /*
3242                  * override is passed when we are
3243                  * loosing b/w and making one last
3244                  * gasp at trying to not loose out
3245                  * to a new-reno flow.
3246                  */
3247                 goto extra_boost;
3248         }
3249         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3250         if (rack->rc_gp_incr &&
3251             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3252                 /*
3253                  * Reset and get 5 strokes more before the boost. Note
3254                  * that the count is 0 based so we have to add one.
3255                  */
3256 extra_boost:
3257                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3258                 rack->rc_gp_timely_inc_cnt = 0;
3259         } else
3260                 plus = (uint32_t)rack_gp_increase_per;
3261         /* Must be at least 1% increase for true timely increases */
3262         if ((plus < 1) &&
3263             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3264                 plus = 1;
3265         if (rack->rc_gp_saw_rec &&
3266             (rack->rc_gp_no_rec_chg == 0) &&
3267             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3268                                   rack->r_ctl.rack_per_of_gp_rec)) {
3269                 /* We have been in recovery ding it too */
3270                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3271                 if (calc > 0xffff)
3272                         calc = 0xffff;
3273                 logged |= 1;
3274                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3275                 if (rack_per_upper_bound_ss &&
3276                     (rack->rc_dragged_bottom == 0) &&
3277                     (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3278                         rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3279         }
3280         if (rack->rc_gp_saw_ca &&
3281             (rack->rc_gp_saw_ss == 0) &&
3282             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3283                                   rack->r_ctl.rack_per_of_gp_ca)) {
3284                 /* In CA */
3285                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3286                 if (calc > 0xffff)
3287                         calc = 0xffff;
3288                 logged |= 2;
3289                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3290                 if (rack_per_upper_bound_ca &&
3291                     (rack->rc_dragged_bottom == 0) &&
3292                     (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3293                         rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3294         }
3295         if (rack->rc_gp_saw_ss &&
3296             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3297                                   rack->r_ctl.rack_per_of_gp_ss)) {
3298                 /* In SS */
3299                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3300                 if (calc > 0xffff)
3301                         calc = 0xffff;
3302                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3303                 if (rack_per_upper_bound_ss &&
3304                     (rack->rc_dragged_bottom == 0) &&
3305                     (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3306                         rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3307                 logged |= 4;
3308         }
3309         if (logged &&
3310             (rack->rc_gp_incr == 0)){
3311                 /* Go into increment mode */
3312                 rack->rc_gp_incr = 1;
3313                 rack->rc_gp_timely_inc_cnt = 0;
3314         }
3315         if (rack->rc_gp_incr &&
3316             logged &&
3317             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3318                 rack->rc_gp_timely_inc_cnt++;
3319         }
3320         rack_log_timely(rack,  logged, plus, 0, 0,
3321                         __LINE__, 1);
3322 }
3323
3324 static uint32_t
3325 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3326 {
3327         /*
3328          * norm_grad = rtt_diff / minrtt;
3329          * new_per = curper * (1 - B * norm_grad)
3330          *
3331          * B = rack_gp_decrease_per (default 10%)
3332          * rtt_dif = input var current rtt-diff
3333          * curper = input var current percentage
3334          * minrtt = from rack filter
3335          *
3336          */
3337         uint64_t perf;
3338
3339         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3340                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3341                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3342                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3343                      (uint64_t)1000000)) /
3344                 (uint64_t)1000000);
3345         if (perf > curper) {
3346                 /* TSNH */
3347                 perf = curper - 1;
3348         }
3349         return ((uint32_t)perf);
3350 }
3351
3352 static uint32_t
3353 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3354 {
3355         /*
3356          *                                   highrttthresh
3357          * result = curper * (1 - (B * ( 1 -  ------          ))
3358          *                                     gp_srtt
3359          *
3360          * B = rack_gp_decrease_per (default 10%)
3361          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3362          */
3363         uint64_t perf;
3364         uint32_t highrttthresh;
3365
3366         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3367
3368         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3369                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3370                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3371                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3372         return (perf);
3373 }
3374
3375 static void
3376 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3377 {
3378         uint64_t logvar, logvar2, logvar3;
3379         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3380
3381         if (rack->rc_gp_incr) {
3382                 /* Turn off increment counting */
3383                 rack->rc_gp_incr = 0;
3384                 rack->rc_gp_timely_inc_cnt = 0;
3385         }
3386         ss_red = ca_red = rec_red = 0;
3387         logged = 0;
3388         /* Calculate the reduction value */
3389         if (rtt_diff < 0) {
3390                 rtt_diff *= -1;
3391         }
3392         /* Must be at least 1% reduction */
3393         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3394                 /* We have been in recovery ding it too */
3395                 if (timely_says == 2) {
3396                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3397                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3398                         if (alt < new_per)
3399                                 val = alt;
3400                         else
3401                                 val = new_per;
3402                 } else
3403                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3404                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3405                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3406                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3407                 } else {
3408                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3409                         rec_red = 0;
3410                 }
3411                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3412                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3413                 logged |= 1;
3414         }
3415         if (rack->rc_gp_saw_ss) {
3416                 /* Sent in SS */
3417                 if (timely_says == 2) {
3418                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3419                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3420                         if (alt < new_per)
3421                                 val = alt;
3422                         else
3423                                 val = new_per;
3424                 } else
3425                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3426                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3427                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3428                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3429                 } else {
3430                         ss_red = new_per;
3431                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3432                         logvar = new_per;
3433                         logvar <<= 32;
3434                         logvar |= alt;
3435                         logvar2 = (uint32_t)rtt;
3436                         logvar2 <<= 32;
3437                         logvar2 |= (uint32_t)rtt_diff;
3438                         logvar3 = rack_gp_rtt_maxmul;
3439                         logvar3 <<= 32;
3440                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3441                         rack_log_timely(rack, timely_says,
3442                                         logvar2, logvar3,
3443                                         logvar, __LINE__, 10);
3444                 }
3445                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3446                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3447                 logged |= 4;
3448         } else if (rack->rc_gp_saw_ca) {
3449                 /* Sent in CA */
3450                 if (timely_says == 2) {
3451                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3452                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3453                         if (alt < new_per)
3454                                 val = alt;
3455                         else
3456                                 val = new_per;
3457                 } else
3458                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3459                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3460                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3461                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3462                 } else {
3463                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3464                         ca_red = 0;
3465                         logvar = new_per;
3466                         logvar <<= 32;
3467                         logvar |= alt;
3468                         logvar2 = (uint32_t)rtt;
3469                         logvar2 <<= 32;
3470                         logvar2 |= (uint32_t)rtt_diff;
3471                         logvar3 = rack_gp_rtt_maxmul;
3472                         logvar3 <<= 32;
3473                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3474                         rack_log_timely(rack, timely_says,
3475                                         logvar2, logvar3,
3476                                         logvar, __LINE__, 10);
3477                 }
3478                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3479                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3480                 logged |= 2;
3481         }
3482         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3483                 rack->rc_gp_timely_dec_cnt++;
3484                 if (rack_timely_dec_clear &&
3485                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3486                         rack->rc_gp_timely_dec_cnt = 0;
3487         }
3488         logvar = ss_red;
3489         logvar <<= 32;
3490         logvar |= ca_red;
3491         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3492                         __LINE__, 2);
3493 }
3494
3495 static void
3496 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3497                      uint32_t rtt, uint32_t line, uint8_t reas)
3498 {
3499         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3500                 union tcp_log_stackspecific log;
3501                 struct timeval tv;
3502
3503                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3504                 log.u_bbr.flex1 = line;
3505                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3506                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3507                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3508                 log.u_bbr.flex5 = rtt;
3509                 log.u_bbr.flex6 = rack->rc_highly_buffered;
3510                 log.u_bbr.flex6 <<= 1;
3511                 log.u_bbr.flex6 |= rack->forced_ack;
3512                 log.u_bbr.flex6 <<= 1;
3513                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3514                 log.u_bbr.flex6 <<= 1;
3515                 log.u_bbr.flex6 |= rack->in_probe_rtt;
3516                 log.u_bbr.flex6 <<= 1;
3517                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3518                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3519                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3520                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3521                 log.u_bbr.flex8 = reas;
3522                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3523                 log.u_bbr.delRate = rack_get_bw(rack);
3524                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3525                 log.u_bbr.cur_del_rate <<= 32;
3526                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3527                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3528                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3529                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3530                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3531                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3532                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3533                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3534                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3535                 log.u_bbr.rttProp = us_cts;
3536                 log.u_bbr.rttProp <<= 32;
3537                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3538                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3539                     &rack->rc_inp->inp_socket->so_rcv,
3540                     &rack->rc_inp->inp_socket->so_snd,
3541                     BBR_LOG_RTT_SHRINKS, 0,
3542                     0, &log, false, &rack->r_ctl.act_rcv_time);
3543         }
3544 }
3545
3546 static void
3547 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3548 {
3549         uint64_t bwdp;
3550
3551         bwdp = rack_get_bw(rack);
3552         bwdp *= (uint64_t)rtt;
3553         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3554         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3555         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3556                 /*
3557                  * A window protocol must be able to have 4 packets
3558                  * outstanding as the floor in order to function
3559                  * (especially considering delayed ack :D).
3560                  */
3561                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3562         }
3563 }
3564
3565 static void
3566 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3567 {
3568         /**
3569          * ProbeRTT is a bit different in rack_pacing than in
3570          * BBR. It is like BBR in that it uses the lowering of
3571          * the RTT as a signal that we saw something new and
3572          * counts from there for how long between. But it is
3573          * different in that its quite simple. It does not
3574          * play with the cwnd and wait until we get down
3575          * to N segments outstanding and hold that for
3576          * 200ms. Instead it just sets the pacing reduction
3577          * rate to a set percentage (70 by default) and hold
3578          * that for a number of recent GP Srtt's.
3579          */
3580         uint32_t segsiz;
3581
3582         if (rack->rc_gp_dyn_mul == 0)
3583                 return;
3584
3585         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3586                 /* We are idle */
3587                 return;
3588         }
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         }
3602         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3603         rack->r_ctl.rc_time_probertt_entered = us_cts;
3604         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3605                      rack->r_ctl.rc_pace_min_segs);
3606         rack->in_probe_rtt = 1;
3607         rack->measure_saw_probe_rtt = 1;
3608         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3609         rack->r_ctl.rc_time_probertt_starts = 0;
3610         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3611         if (rack_probertt_use_min_rtt_entry)
3612                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3613         else
3614                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3615         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3616                              __LINE__, RACK_RTTS_ENTERPROBE);
3617 }
3618
3619 static void
3620 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3621 {
3622         struct rack_sendmap *rsm;
3623         uint32_t segsiz;
3624
3625         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3626                      rack->r_ctl.rc_pace_min_segs);
3627         rack->in_probe_rtt = 0;
3628         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3629             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3630                 /*
3631                  * Stop the goodput now, the idea here is
3632                  * that future measurements with in_probe_rtt
3633                  * won't register if they are not greater so
3634                  * we want to get what info (if any) is available
3635                  * now.
3636                  */
3637                 rack_do_goodput_measurement(rack->rc_tp, rack,
3638                                             rack->rc_tp->snd_una, __LINE__,
3639                                             RACK_QUALITY_PROBERTT);
3640         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3641                 /*
3642                  * We don't have enough data to make a measurement.
3643                  * So lets just stop and start here after exiting
3644                  * probe-rtt. We probably are not interested in
3645                  * the results anyway.
3646                  */
3647                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3648         }
3649         /*
3650          * Measurements through the current snd_max are going
3651          * to be limited by the slower pacing rate.
3652          *
3653          * We need to mark these as app-limited so we
3654          * don't collapse the b/w.
3655          */
3656         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3657         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3658                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3659                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3660                 else {
3661                         /*
3662                          * Go out to the end app limited and mark
3663                          * this new one as next and move the end_appl up
3664                          * to this guy.
3665                          */
3666                         if (rack->r_ctl.rc_end_appl)
3667                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3668                         rack->r_ctl.rc_end_appl = rsm;
3669                 }
3670                 rsm->r_flags |= RACK_APP_LIMITED;
3671                 rack->r_ctl.rc_app_limited_cnt++;
3672         }
3673         /*
3674          * Now, we need to examine our pacing rate multipliers.
3675          * If its under 100%, we need to kick it back up to
3676          * 100%. We also don't let it be over our "max" above
3677          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3678          * Note setting clamp_atexit_prtt to 0 has the effect
3679          * of setting CA/SS to 100% always at exit (which is
3680          * the default behavior).
3681          */
3682         if (rack_probertt_clear_is) {
3683                 rack->rc_gp_incr = 0;
3684                 rack->rc_gp_bwred = 0;
3685                 rack->rc_gp_timely_inc_cnt = 0;
3686                 rack->rc_gp_timely_dec_cnt = 0;
3687         }
3688         /* Do we do any clamping at exit? */
3689         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3690                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3691                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3692         }
3693         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3694                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3695                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3696         }
3697         /*
3698          * Lets set rtt_diff to 0, so that we will get a "boost"
3699          * after exiting.
3700          */
3701         rack->r_ctl.rc_rtt_diff = 0;
3702
3703         /* Clear all flags so we start fresh */
3704         rack->rc_tp->t_bytes_acked = 0;
3705         rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3706         /*
3707          * If configured to, set the cwnd and ssthresh to
3708          * our targets.
3709          */
3710         if (rack_probe_rtt_sets_cwnd) {
3711                 uint64_t ebdp;
3712                 uint32_t setto;
3713
3714                 /* Set ssthresh so we get into CA once we hit our target */
3715                 if (rack_probertt_use_min_rtt_exit == 1) {
3716                         /* Set to min rtt */
3717                         rack_set_prtt_target(rack, segsiz,
3718                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3719                 } else if (rack_probertt_use_min_rtt_exit == 2) {
3720                         /* Set to current gp rtt */
3721                         rack_set_prtt_target(rack, segsiz,
3722                                              rack->r_ctl.rc_gp_srtt);
3723                 } else if (rack_probertt_use_min_rtt_exit == 3) {
3724                         /* Set to entry gp rtt */
3725                         rack_set_prtt_target(rack, segsiz,
3726                                              rack->r_ctl.rc_entry_gp_rtt);
3727                 } else {
3728                         uint64_t sum;
3729                         uint32_t setval;
3730
3731                         sum = rack->r_ctl.rc_entry_gp_rtt;
3732                         sum *= 10;
3733                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3734                         if (sum >= 20) {
3735                                 /*
3736                                  * A highly buffered path needs
3737                                  * cwnd space for timely to work.
3738                                  * Lets set things up as if
3739                                  * we are heading back here again.
3740                                  */
3741                                 setval = rack->r_ctl.rc_entry_gp_rtt;
3742                         } else if (sum >= 15) {
3743                                 /*
3744                                  * Lets take the smaller of the
3745                                  * two since we are just somewhat
3746                                  * buffered.
3747                                  */
3748                                 setval = rack->r_ctl.rc_gp_srtt;
3749                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3750                                         setval = rack->r_ctl.rc_entry_gp_rtt;
3751                         } else {
3752                                 /*
3753                                  * Here we are not highly buffered
3754                                  * and should pick the min we can to
3755                                  * keep from causing loss.
3756                                  */
3757                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3758                         }
3759                         rack_set_prtt_target(rack, segsiz,
3760                                              setval);
3761                 }
3762                 if (rack_probe_rtt_sets_cwnd > 1) {
3763                         /* There is a percentage here to boost */
3764                         ebdp = rack->r_ctl.rc_target_probertt_flight;
3765                         ebdp *= rack_probe_rtt_sets_cwnd;
3766                         ebdp /= 100;
3767                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3768                 } else
3769                         setto = rack->r_ctl.rc_target_probertt_flight;
3770                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3771                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3772                         /* Enforce a min */
3773                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3774                 }
3775                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3776                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3777         }
3778         rack_log_rtt_shrinks(rack,  us_cts,
3779                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3780                              __LINE__, RACK_RTTS_EXITPROBE);
3781         /* Clear times last so log has all the info */
3782         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3783         rack->r_ctl.rc_time_probertt_entered = us_cts;
3784         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3785         rack->r_ctl.rc_time_of_last_probertt = us_cts;
3786 }
3787
3788 static void
3789 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3790 {
3791         /* Check in on probe-rtt */
3792         if (rack->rc_gp_filled == 0) {
3793                 /* We do not do p-rtt unless we have gp measurements */
3794                 return;
3795         }
3796         if (rack->in_probe_rtt) {
3797                 uint64_t no_overflow;
3798                 uint32_t endtime, must_stay;
3799
3800                 if (rack->r_ctl.rc_went_idle_time &&
3801                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3802                         /*
3803                          * We went idle during prtt, just exit now.
3804                          */
3805                         rack_exit_probertt(rack, us_cts);
3806                 } else if (rack_probe_rtt_safety_val &&
3807                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3808                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3809                         /*
3810                          * Probe RTT safety value triggered!
3811                          */
3812                         rack_log_rtt_shrinks(rack,  us_cts,
3813                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3814                                              __LINE__, RACK_RTTS_SAFETY);
3815                         rack_exit_probertt(rack, us_cts);
3816                 }
3817                 /* Calculate the max we will wait */
3818                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3819                 if (rack->rc_highly_buffered)
3820                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3821                 /* Calculate the min we must wait */
3822                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3823                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3824                     TSTMP_LT(us_cts, endtime)) {
3825                         uint32_t calc;
3826                         /* Do we lower more? */
3827 no_exit:
3828                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3829                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3830                         else
3831                                 calc = 0;
3832                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3833                         if (calc) {
3834                                 /* Maybe */
3835                                 calc *= rack_per_of_gp_probertt_reduce;
3836                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3837                                 /* Limit it too */
3838                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3839                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3840                         }
3841                         /* We must reach target or the time set */
3842                         return;
3843                 }
3844                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3845                         if ((TSTMP_LT(us_cts, must_stay) &&
3846                              rack->rc_highly_buffered) ||
3847                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3848                               rack->r_ctl.rc_target_probertt_flight)) {
3849                                 /* We are not past the must_stay time */
3850                                 goto no_exit;
3851                         }
3852                         rack_log_rtt_shrinks(rack,  us_cts,
3853                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3854                                              __LINE__, RACK_RTTS_REACHTARGET);
3855                         rack->r_ctl.rc_time_probertt_starts = us_cts;
3856                         if (rack->r_ctl.rc_time_probertt_starts == 0)
3857                                 rack->r_ctl.rc_time_probertt_starts = 1;
3858                         /* Restore back to our rate we want to pace at in prtt */
3859                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3860                 }
3861                 /*
3862                  * Setup our end time, some number of gp_srtts plus 200ms.
3863                  */
3864                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3865                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3866                 if (rack_probertt_gpsrtt_cnt_div)
3867                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3868                 else
3869                         endtime = 0;
3870                 endtime += rack_min_probertt_hold;
3871                 endtime += rack->r_ctl.rc_time_probertt_starts;
3872                 if (TSTMP_GEQ(us_cts,  endtime)) {
3873                         /* yes, exit probertt */
3874                         rack_exit_probertt(rack, us_cts);
3875                 }
3876
3877         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3878                 /* Go into probertt, its been too long since we went lower */
3879                 rack_enter_probertt(rack, us_cts);
3880         }
3881 }
3882
3883 static void
3884 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3885                        uint32_t rtt, int32_t rtt_diff)
3886 {
3887         uint64_t cur_bw, up_bnd, low_bnd, subfr;
3888         uint32_t losses;
3889
3890         if ((rack->rc_gp_dyn_mul == 0) ||
3891             (rack->use_fixed_rate) ||
3892             (rack->in_probe_rtt) ||
3893             (rack->rc_always_pace == 0)) {
3894                 /* No dynamic GP multiplier in play */
3895                 return;
3896         }
3897         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3898         cur_bw = rack_get_bw(rack);
3899         /* Calculate our up and down range */
3900         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3901         up_bnd /= 100;
3902         up_bnd += rack->r_ctl.last_gp_comp_bw;
3903
3904         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3905         subfr /= 100;
3906         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3907         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3908                 /*
3909                  * This is the case where our RTT is above
3910                  * the max target and we have been configured
3911                  * to just do timely no bonus up stuff in that case.
3912                  *
3913                  * There are two configurations, set to 1, and we
3914                  * just do timely if we are over our max. If its
3915                  * set above 1 then we slam the multipliers down
3916                  * to 100 and then decrement per timely.
3917                  */
3918                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3919                                 __LINE__, 3);
3920                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3921                         rack_validate_multipliers_at_or_below_100(rack);
3922                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3923         } else if ((last_bw_est < low_bnd) && !losses) {
3924                 /*
3925                  * We are decreasing this is a bit complicated this
3926                  * means we are loosing ground. This could be
3927                  * because another flow entered and we are competing
3928                  * for b/w with it. This will push the RTT up which
3929                  * makes timely unusable unless we want to get shoved
3930                  * into a corner and just be backed off (the age
3931                  * old problem with delay based CC).
3932                  *
3933                  * On the other hand if it was a route change we
3934                  * would like to stay somewhat contained and not
3935                  * blow out the buffers.
3936                  */
3937                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3938                                 __LINE__, 3);
3939                 rack->r_ctl.last_gp_comp_bw = cur_bw;
3940                 if (rack->rc_gp_bwred == 0) {
3941                         /* Go into reduction counting */
3942                         rack->rc_gp_bwred = 1;
3943                         rack->rc_gp_timely_dec_cnt = 0;
3944                 }
3945                 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3946                     (timely_says == 0)) {
3947                         /*
3948                          * Push another time with a faster pacing
3949                          * to try to gain back (we include override to
3950                          * get a full raise factor).
3951                          */
3952                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3953                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3954                             (timely_says == 0) ||
3955                             (rack_down_raise_thresh == 0)) {
3956                                 /*
3957                                  * Do an override up in b/w if we were
3958                                  * below the threshold or if the threshold
3959                                  * is zero we always do the raise.
3960                                  */
3961                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3962                         } else {
3963                                 /* Log it stays the same */
3964                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3965                                                 __LINE__, 11);
3966                         }
3967                         rack->rc_gp_timely_dec_cnt++;
3968                         /* We are not incrementing really no-count */
3969                         rack->rc_gp_incr = 0;
3970                         rack->rc_gp_timely_inc_cnt = 0;
3971                 } else {
3972                         /*
3973                          * Lets just use the RTT
3974                          * information and give up
3975                          * pushing.
3976                          */
3977                         goto use_timely;
3978                 }
3979         } else if ((timely_says != 2) &&
3980                     !losses &&
3981                     (last_bw_est > up_bnd)) {
3982                 /*
3983                  * We are increasing b/w lets keep going, updating
3984                  * our b/w and ignoring any timely input, unless
3985                  * of course we are at our max raise (if there is one).
3986                  */
3987
3988                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3989                                 __LINE__, 3);
3990                 rack->r_ctl.last_gp_comp_bw = cur_bw;
3991                 if (rack->rc_gp_saw_ss &&
3992                     rack_per_upper_bound_ss &&
3993                      (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3994                             /*
3995                              * In cases where we can't go higher
3996                              * we should just use timely.
3997                              */
3998                             goto use_timely;
3999                 }
4000                 if (rack->rc_gp_saw_ca &&
4001                     rack_per_upper_bound_ca &&
4002                     (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4003                             /*
4004                              * In cases where we can't go higher
4005                              * we should just use timely.
4006                              */
4007                             goto use_timely;
4008                 }
4009                 rack->rc_gp_bwred = 0;
4010                 rack->rc_gp_timely_dec_cnt = 0;
4011                 /* You get a set number of pushes if timely is trying to reduce */
4012                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4013                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4014                 } else {
4015                         /* Log it stays the same */
4016                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4017                             __LINE__, 12);
4018                 }
4019                 return;
4020         } else {
4021                 /*
4022                  * We are staying between the lower and upper range bounds
4023                  * so use timely to decide.
4024                  */
4025                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4026                                 __LINE__, 3);
4027 use_timely:
4028                 if (timely_says) {
4029                         rack->rc_gp_incr = 0;
4030                         rack->rc_gp_timely_inc_cnt = 0;
4031                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4032                             !losses &&
4033                             (last_bw_est < low_bnd)) {
4034                                 /* We are loosing ground */
4035                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4036                                 rack->rc_gp_timely_dec_cnt++;
4037                                 /* We are not incrementing really no-count */
4038                                 rack->rc_gp_incr = 0;
4039                                 rack->rc_gp_timely_inc_cnt = 0;
4040                         } else
4041                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4042                 } else {
4043                         rack->rc_gp_bwred = 0;
4044                         rack->rc_gp_timely_dec_cnt = 0;
4045                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4046                 }
4047         }
4048 }
4049
4050 static int32_t
4051 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4052 {
4053         int32_t timely_says;
4054         uint64_t log_mult, log_rtt_a_diff;
4055
4056         log_rtt_a_diff = rtt;
4057         log_rtt_a_diff <<= 32;
4058         log_rtt_a_diff |= (uint32_t)rtt_diff;
4059         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4060                     rack_gp_rtt_maxmul)) {
4061                 /* Reduce the b/w multiplier */
4062                 timely_says = 2;
4063                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4064                 log_mult <<= 32;
4065                 log_mult |= prev_rtt;
4066                 rack_log_timely(rack,  timely_says, log_mult,
4067                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4068                                 log_rtt_a_diff, __LINE__, 4);
4069         } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4070                            ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4071                             max(rack_gp_rtt_mindiv , 1)))) {
4072                 /* Increase the b/w multiplier */
4073                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4074                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4075                          max(rack_gp_rtt_mindiv , 1));
4076                 log_mult <<= 32;
4077                 log_mult |= prev_rtt;
4078                 timely_says = 0;
4079                 rack_log_timely(rack,  timely_says, log_mult ,
4080                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4081                                 log_rtt_a_diff, __LINE__, 5);
4082         } else {
4083                 /*
4084                  * Use a gradient to find it the timely gradient
4085                  * is:
4086                  * grad = rc_rtt_diff / min_rtt;
4087                  *
4088                  * anything below or equal to 0 will be
4089                  * a increase indication. Anything above
4090                  * zero is a decrease. Note we take care
4091                  * of the actual gradient calculation
4092                  * in the reduction (its not needed for
4093                  * increase).
4094                  */
4095                 log_mult = prev_rtt;
4096                 if (rtt_diff <= 0) {
4097                         /*
4098                          * Rttdiff is less than zero, increase the
4099                          * b/w multiplier (its 0 or negative)
4100                          */
4101                         timely_says = 0;
4102                         rack_log_timely(rack,  timely_says, log_mult,
4103                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4104                 } else {
4105                         /* Reduce the b/w multiplier */
4106                         timely_says = 1;
4107                         rack_log_timely(rack,  timely_says, log_mult,
4108                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4109                 }
4110         }
4111         return (timely_says);
4112 }
4113
4114 static void
4115 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4116                             tcp_seq th_ack, int line, uint8_t quality)
4117 {
4118         uint64_t tim, bytes_ps, ltim, stim, utim;
4119         uint32_t segsiz, bytes, reqbytes, us_cts;
4120         int32_t gput, new_rtt_diff, timely_says;
4121         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4122         int did_add = 0;
4123
4124         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4125         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4126         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4127                 tim = us_cts - tp->gput_ts;
4128         else
4129                 tim = 0;
4130         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4131                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4132         else
4133                 stim = 0;
4134         /*
4135          * Use the larger of the send time or ack time. This prevents us
4136          * from being influenced by ack artifacts to come up with too
4137          * high of measurement. Note that since we are spanning over many more
4138          * bytes in most of our measurements hopefully that is less likely to
4139          * occur.
4140          */
4141         if (tim > stim)
4142                 utim = max(tim, 1);
4143         else
4144                 utim = max(stim, 1);
4145         /* Lets get a msec time ltim too for the old stuff */
4146         ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4147         gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4148         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4149         if ((tim == 0) && (stim == 0)) {
4150                 /*
4151                  * Invalid measurement time, maybe
4152                  * all on one ack/one send?
4153                  */
4154                 bytes = 0;
4155                 bytes_ps = 0;
4156                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4157                                            0, 0, 0, 10, __LINE__, NULL, quality);
4158                 goto skip_measurement;
4159         }
4160         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4161                 /* We never made a us_rtt measurement? */
4162                 bytes = 0;
4163                 bytes_ps = 0;
4164                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4165                                            0, 0, 0, 10, __LINE__, NULL, quality);
4166                 goto skip_measurement;
4167         }
4168         /*
4169          * Calculate the maximum possible b/w this connection
4170          * could have. We base our calculation on the lowest
4171          * rtt we have seen during the measurement and the
4172          * largest rwnd the client has given us in that time. This
4173          * forms a BDP that is the maximum that we could ever
4174          * get to the client. Anything larger is not valid.
4175          *
4176          * I originally had code here that rejected measurements
4177          * where the time was less than 1/2 the latest us_rtt.
4178          * But after thinking on that I realized its wrong since
4179          * say you had a 150Mbps or even 1Gbps link, and you
4180          * were a long way away.. example I am in Europe (100ms rtt)
4181          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4182          * bytes my time would be 1.2ms, and yet my rtt would say
4183          * the measurement was invalid the time was < 50ms. The
4184          * same thing is true for 150Mb (8ms of time).
4185          *
4186          * A better way I realized is to look at what the maximum
4187          * the connection could possibly do. This is gated on
4188          * the lowest RTT we have seen and the highest rwnd.
4189          * We should in theory never exceed that, if we are
4190          * then something on the path is storing up packets
4191          * and then feeding them all at once to our endpoint
4192          * messing up our measurement.
4193          */
4194         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4195         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4196         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4197         if (SEQ_LT(th_ack, tp->gput_seq)) {
4198                 /* No measurement can be made */
4199                 bytes = 0;
4200                 bytes_ps = 0;
4201                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4202                                            0, 0, 0, 10, __LINE__, NULL, quality);
4203                 goto skip_measurement;
4204         } else
4205                 bytes = (th_ack - tp->gput_seq);
4206         bytes_ps = (uint64_t)bytes;
4207         /*
4208          * Don't measure a b/w for pacing unless we have gotten at least
4209          * an initial windows worth of data in this measurement interval.
4210          *
4211          * Small numbers of bytes get badly influenced by delayed ack and
4212          * other artifacts. Note we take the initial window or our
4213          * defined minimum GP (defaulting to 10 which hopefully is the
4214          * IW).
4215          */
4216         if (rack->rc_gp_filled == 0) {
4217                 /*
4218                  * The initial estimate is special. We
4219                  * have blasted out an IW worth of packets
4220                  * without a real valid ack ts results. We
4221                  * then setup the app_limited_needs_set flag,
4222                  * this should get the first ack in (probably 2
4223                  * MSS worth) to be recorded as the timestamp.
4224                  * We thus allow a smaller number of bytes i.e.
4225                  * IW - 2MSS.
4226                  */
4227                 reqbytes -= (2 * segsiz);
4228                 /* Also lets fill previous for our first measurement to be neutral */
4229                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4230         }
4231         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4232                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4233                                            rack->r_ctl.rc_app_limited_cnt,
4234                                            0, 0, 10, __LINE__, NULL, quality);
4235                 goto skip_measurement;
4236         }
4237         /*
4238          * We now need to calculate the Timely like status so
4239          * we can update (possibly) the b/w multipliers.
4240          */
4241         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4242         if (rack->rc_gp_filled == 0) {
4243                 /* No previous reading */
4244                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4245         } else {
4246                 if (rack->measure_saw_probe_rtt == 0) {
4247                         /*
4248                          * We don't want a probertt to be counted
4249                          * since it will be negative incorrectly. We
4250                          * expect to be reducing the RTT when we
4251                          * pace at a slower rate.
4252                          */
4253                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4254                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4255                 }
4256         }
4257         timely_says = rack_make_timely_judgement(rack,
4258                 rack->r_ctl.rc_gp_srtt,
4259                 rack->r_ctl.rc_rtt_diff,
4260                 rack->r_ctl.rc_prev_gp_srtt
4261                 );
4262         bytes_ps *= HPTS_USEC_IN_SEC;
4263         bytes_ps /= utim;
4264         if (bytes_ps > rack->r_ctl.last_max_bw) {
4265                 /*
4266                  * Something is on path playing
4267                  * since this b/w is not possible based
4268                  * on our BDP (highest rwnd and lowest rtt
4269                  * we saw in the measurement window).
4270                  *
4271                  * Another option here would be to
4272                  * instead skip the measurement.
4273                  */
4274                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4275                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4276                                            11, __LINE__, NULL, quality);
4277                 bytes_ps = rack->r_ctl.last_max_bw;
4278         }
4279         /* We store gp for b/w in bytes per second */
4280         if (rack->rc_gp_filled == 0) {
4281                 /* Initial measurement */
4282                 if (bytes_ps) {
4283                         rack->r_ctl.gp_bw = bytes_ps;
4284                         rack->rc_gp_filled = 1;
4285                         rack->r_ctl.num_measurements = 1;
4286                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4287                 } else {
4288                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4289                                                    rack->r_ctl.rc_app_limited_cnt,
4290                                                    0, 0, 10, __LINE__, NULL, quality);
4291                 }
4292                 if (tcp_in_hpts(rack->rc_inp) &&
4293                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4294                         /*
4295                          * Ok we can't trust the pacer in this case
4296                          * where we transition from un-paced to paced.
4297                          * Or for that matter when the burst mitigation
4298                          * was making a wild guess and got it wrong.
4299                          * Stop the pacer and clear up all the aggregate
4300                          * delays etc.
4301                          */
4302                         tcp_hpts_remove(rack->rc_inp);
4303                         rack->r_ctl.rc_hpts_flags = 0;
4304                         rack->r_ctl.rc_last_output_to = 0;
4305                 }
4306                 did_add = 2;
4307         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4308                 /* Still a small number run an average */
4309                 rack->r_ctl.gp_bw += bytes_ps;
4310                 addpart = rack->r_ctl.num_measurements;
4311                 rack->r_ctl.num_measurements++;
4312                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4313                         /* We have collected enough to move forward */
4314                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4315                 }
4316                 did_add = 3;
4317         } else {
4318                 /*
4319                  * We want to take 1/wma of the goodput and add in to 7/8th
4320                  * of the old value weighted by the srtt. So if your measurement
4321                  * period is say 2 SRTT's long you would get 1/4 as the
4322                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4323                  *
4324                  * But we must be careful not to take too much i.e. if the
4325                  * srtt is say 20ms and the measurement is taken over
4326                  * 400ms our weight would be 400/20 i.e. 20. On the
4327                  * other hand if we get a measurement over 1ms with a
4328                  * 10ms rtt we only want to take a much smaller portion.
4329                  */
4330                 if (rack->r_ctl.num_measurements < 0xff) {
4331                         rack->r_ctl.num_measurements++;
4332                 }
4333                 srtt = (uint64_t)tp->t_srtt;
4334                 if (srtt == 0) {
4335                         /*
4336                          * Strange why did t_srtt go back to zero?
4337                          */
4338                         if (rack->r_ctl.rc_rack_min_rtt)
4339                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4340                         else
4341                                 srtt = HPTS_USEC_IN_MSEC;
4342                 }
4343                 /*
4344                  * XXXrrs: Note for reviewers, in playing with
4345                  * dynamic pacing I discovered this GP calculation
4346                  * as done originally leads to some undesired results.
4347                  * Basically you can get longer measurements contributing
4348                  * too much to the WMA. Thus I changed it if you are doing
4349                  * dynamic adjustments to only do the aportioned adjustment
4350                  * if we have a very small (time wise) measurement. Longer
4351                  * measurements just get there weight (defaulting to 1/8)
4352                  * add to the WMA. We may want to think about changing
4353                  * this to always do that for both sides i.e. dynamic
4354                  * and non-dynamic... but considering lots of folks
4355                  * were playing with this I did not want to change the
4356                  * calculation per.se. without your thoughts.. Lawerence?
4357                  * Peter??
4358                  */
4359                 if (rack->rc_gp_dyn_mul == 0) {
4360                         subpart = rack->r_ctl.gp_bw * utim;
4361                         subpart /= (srtt * 8);
4362                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4363                                 /*
4364                                  * The b/w update takes no more
4365                                  * away then 1/2 our running total
4366                                  * so factor it in.
4367                                  */
4368                                 addpart = bytes_ps * utim;
4369                                 addpart /= (srtt * 8);
4370                         } else {
4371                                 /*
4372                                  * Don't allow a single measurement
4373                                  * to account for more than 1/2 of the
4374                                  * WMA. This could happen on a retransmission
4375                                  * where utim becomes huge compared to
4376                                  * srtt (multiple retransmissions when using
4377                                  * the sending rate which factors in all the
4378                                  * transmissions from the first one).
4379                                  */
4380                                 subpart = rack->r_ctl.gp_bw / 2;
4381                                 addpart = bytes_ps / 2;
4382                         }
4383                         resid_bw = rack->r_ctl.gp_bw - subpart;
4384                         rack->r_ctl.gp_bw = resid_bw + addpart;
4385                         did_add = 1;
4386                 } else {
4387                         if ((utim / srtt) <= 1) {
4388                                 /*
4389                                  * The b/w update was over a small period
4390                                  * of time. The idea here is to prevent a small
4391                                  * measurement time period from counting
4392                                  * too much. So we scale it based on the
4393                                  * time so it attributes less than 1/rack_wma_divisor
4394                                  * of its measurement.
4395                                  */
4396                                 subpart = rack->r_ctl.gp_bw * utim;
4397                                 subpart /= (srtt * rack_wma_divisor);
4398                                 addpart = bytes_ps * utim;
4399                                 addpart /= (srtt * rack_wma_divisor);
4400                         } else {
4401                                 /*
4402                                  * The scaled measurement was long
4403                                  * enough so lets just add in the
4404                                  * portion of the measurement i.e. 1/rack_wma_divisor
4405                                  */
4406                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4407                                 addpart = bytes_ps / rack_wma_divisor;
4408                         }
4409                         if ((rack->measure_saw_probe_rtt == 0) ||
4410                             (bytes_ps > rack->r_ctl.gp_bw)) {
4411                                 /*
4412                                  * For probe-rtt we only add it in
4413                                  * if its larger, all others we just
4414                                  * add in.
4415                                  */
4416                                 did_add = 1;
4417                                 resid_bw = rack->r_ctl.gp_bw - subpart;
4418                                 rack->r_ctl.gp_bw = resid_bw + addpart;
4419                         }
4420                 }
4421         }
4422         if ((rack->gp_ready == 0) &&
4423             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4424                 /* We have enough measurements now */
4425                 rack->gp_ready = 1;
4426                 rack_set_cc_pacing(rack);
4427                 if (rack->defer_options)
4428                         rack_apply_deferred_options(rack);
4429         }
4430         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4431                                    rack_get_bw(rack), 22, did_add, NULL, quality);
4432         /* We do not update any multipliers if we are in or have seen a probe-rtt */
4433         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4434                 rack_update_multiplier(rack, timely_says, bytes_ps,
4435                                        rack->r_ctl.rc_gp_srtt,
4436                                        rack->r_ctl.rc_rtt_diff);
4437         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4438                                    rack_get_bw(rack), 3, line, NULL, quality);
4439         /* reset the gp srtt and setup the new prev */
4440         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4441         /* Record the lost count for the next measurement */
4442         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4443         /*
4444          * We restart our diffs based on the gpsrtt in the
4445          * measurement window.
4446          */
4447         rack->rc_gp_rtt_set = 0;
4448         rack->rc_gp_saw_rec = 0;
4449         rack->rc_gp_saw_ca = 0;
4450         rack->rc_gp_saw_ss = 0;
4451         rack->rc_dragged_bottom = 0;
4452 skip_measurement:
4453
4454 #ifdef STATS
4455         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4456                                  gput);
4457         /*
4458          * XXXLAS: This is a temporary hack, and should be
4459          * chained off VOI_TCP_GPUT when stats(9) grows an
4460          * API to deal with chained VOIs.
4461          */
4462         if (tp->t_stats_gput_prev > 0)
4463                 stats_voi_update_abs_s32(tp->t_stats,
4464                                          VOI_TCP_GPUT_ND,
4465                                          ((gput - tp->t_stats_gput_prev) * 100) /
4466                                          tp->t_stats_gput_prev);
4467 #endif
4468         tp->t_flags &= ~TF_GPUTINPROG;
4469         tp->t_stats_gput_prev = gput;
4470         /*
4471          * Now are we app limited now and there is space from where we
4472          * were to where we want to go?
4473          *
4474          * We don't do the other case i.e. non-applimited here since
4475          * the next send will trigger us picking up the missing data.
4476          */
4477         if (rack->r_ctl.rc_first_appl &&
4478             TCPS_HAVEESTABLISHED(tp->t_state) &&
4479             rack->r_ctl.rc_app_limited_cnt &&
4480             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4481             ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4482              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4483                 /*
4484                  * Yep there is enough outstanding to make a measurement here.
4485                  */
4486                 struct rack_sendmap *rsm, fe;
4487
4488                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4489                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4490                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4491                 rack->app_limited_needs_set = 0;
4492                 tp->gput_seq = th_ack;
4493                 if (rack->in_probe_rtt)
4494                         rack->measure_saw_probe_rtt = 1;
4495                 else if ((rack->measure_saw_probe_rtt) &&
4496                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4497                         rack->measure_saw_probe_rtt = 0;
4498                 if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4499                         /* There is a full window to gain info from */
4500                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4501                 } else {
4502                         /* We can only measure up to the applimited point */
4503                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4504                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4505                                 /*
4506                                  * We don't have enough to make a measurement.
4507                                  */
4508                                 tp->t_flags &= ~TF_GPUTINPROG;
4509                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4510                                                            0, 0, 0, 6, __LINE__, NULL, quality);
4511                                 return;
4512                         }
4513                 }
4514                 if (tp->t_state >= TCPS_FIN_WAIT_1) {
4515                         /*
4516                          * We will get no more data into the SB
4517                          * this means we need to have the data available
4518                          * before we start a measurement.
4519                          */
4520                         if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4521                                 /* Nope not enough data. */
4522                                 return;
4523                         }
4524                 }
4525                 tp->t_flags |= TF_GPUTINPROG;
4526                 /*
4527                  * Now we need to find the timestamp of the send at tp->gput_seq
4528                  * for the send based measurement.
4529                  */
4530                 fe.r_start = tp->gput_seq;
4531                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4532                 if (rsm) {
4533                         /* Ok send-based limit is set */
4534                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4535                                 /*
4536                                  * Move back to include the earlier part
4537                                  * so our ack time lines up right (this may
4538                                  * make an overlapping measurement but thats
4539                                  * ok).
4540                                  */
4541                                 tp->gput_seq = rsm->r_start;
4542                         }
4543                         if (rsm->r_flags & RACK_ACKED)
4544                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4545                         else
4546                                 rack->app_limited_needs_set = 1;
4547                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4548                 } else {
4549                         /*
4550                          * If we don't find the rsm due to some
4551                          * send-limit set the current time, which
4552                          * basically disables the send-limit.
4553                          */
4554                         struct timeval tv;
4555
4556                         microuptime(&tv);
4557                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4558                 }
4559                 rack_log_pacing_delay_calc(rack,
4560                                            tp->gput_seq,
4561                                            tp->gput_ack,
4562                                            (uint64_t)rsm,
4563                                            tp->gput_ts,
4564                                            rack->r_ctl.rc_app_limited_cnt,
4565                                            9,
4566                                            __LINE__, NULL, quality);
4567         }
4568 }
4569
4570 /*
4571  * CC wrapper hook functions
4572  */
4573 static void
4574 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4575     uint16_t type, int32_t recovery)
4576 {
4577         uint32_t prior_cwnd, acked;
4578         struct tcp_log_buffer *lgb = NULL;
4579         uint8_t labc_to_use, quality;
4580
4581         INP_WLOCK_ASSERT(tp->t_inpcb);
4582         tp->ccv->nsegs = nsegs;
4583         acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4584         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4585                 uint32_t max;
4586
4587                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4588                 if (tp->ccv->bytes_this_ack > max) {
4589                         tp->ccv->bytes_this_ack = max;
4590                 }
4591         }
4592 #ifdef STATS
4593         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4594             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4595 #endif
4596         quality = RACK_QUALITY_NONE;
4597         if ((tp->t_flags & TF_GPUTINPROG) &&
4598             rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4599                 /* Measure the Goodput */
4600                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4601 #ifdef NETFLIX_PEAKRATE
4602                 if ((type == CC_ACK) &&
4603                     (tp->t_maxpeakrate)) {
4604                         /*
4605                          * We update t_peakrate_thr. This gives us roughly
4606                          * one update per round trip time. Note
4607                          * it will only be used if pace_always is off i.e
4608                          * we don't do this for paced flows.
4609                          */
4610                         rack_update_peakrate_thr(tp);
4611                 }
4612 #endif
4613         }
4614         /* Which way our we limited, if not cwnd limited no advance in CA */
4615         if (tp->snd_cwnd <= tp->snd_wnd)
4616                 tp->ccv->flags |= CCF_CWND_LIMITED;
4617         else
4618                 tp->ccv->flags &= ~CCF_CWND_LIMITED;
4619         if (tp->snd_cwnd > tp->snd_ssthresh) {
4620                 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4621                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4622                 /* For the setting of a window past use the actual scwnd we are using */
4623                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4624                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4625                         tp->ccv->flags |= CCF_ABC_SENTAWND;
4626                 }
4627         } else {
4628                 tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4629                 tp->t_bytes_acked = 0;
4630         }
4631         prior_cwnd = tp->snd_cwnd;
4632         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4633             (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4634                 labc_to_use = rack->rc_labc;
4635         else
4636                 labc_to_use = rack_max_abc_post_recovery;
4637         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4638                 union tcp_log_stackspecific log;
4639                 struct timeval tv;
4640
4641                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4642                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4643                 log.u_bbr.flex1 = th_ack;
4644                 log.u_bbr.flex2 = tp->ccv->flags;
4645                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4646                 log.u_bbr.flex4 = tp->ccv->nsegs;
4647                 log.u_bbr.flex5 = labc_to_use;
4648                 log.u_bbr.flex6 = prior_cwnd;
4649                 log.u_bbr.flex7 = V_tcp_do_newsack;
4650                 log.u_bbr.flex8 = 1;
4651                 lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4652                                      0, &log, false, NULL, NULL, 0, &tv);
4653         }
4654         if (CC_ALGO(tp)->ack_received != NULL) {
4655                 /* XXXLAS: Find a way to live without this */
4656                 tp->ccv->curack = th_ack;
4657                 tp->ccv->labc = labc_to_use;
4658                 tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4659                 CC_ALGO(tp)->ack_received(tp->ccv, type);
4660         }
4661         if (lgb) {
4662                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4663         }
4664         if (rack->r_must_retran) {
4665                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4666                         /*
4667                          * We now are beyond the rxt point so lets disable
4668                          * the flag.
4669                          */
4670                         rack->r_ctl.rc_out_at_rto = 0;
4671                         rack->r_must_retran = 0;
4672                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4673                         /*
4674                          * Only decrement the rc_out_at_rto if the cwnd advances
4675                          * at least a whole segment. Otherwise next time the peer
4676                          * acks, we won't be able to send this generaly happens
4677                          * when we are in Congestion Avoidance.
4678                          */
4679                         if (acked <= rack->r_ctl.rc_out_at_rto){
4680                                 rack->r_ctl.rc_out_at_rto -= acked;
4681                         } else {
4682                                 rack->r_ctl.rc_out_at_rto = 0;
4683                         }
4684                 }
4685         }
4686 #ifdef STATS
4687         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4688 #endif
4689         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4690                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4691         }
4692 #ifdef NETFLIX_PEAKRATE
4693         /* we enforce max peak rate if it is set and we are not pacing */
4694         if ((rack->rc_always_pace == 0) &&
4695             tp->t_peakrate_thr &&
4696             (tp->snd_cwnd > tp->t_peakrate_thr)) {
4697                 tp->snd_cwnd = tp->t_peakrate_thr;
4698         }
4699 #endif
4700 }
4701
4702 static void
4703 tcp_rack_partialack(struct tcpcb *tp)
4704 {
4705         struct tcp_rack *rack;
4706
4707         rack = (struct tcp_rack *)tp->t_fb_ptr;
4708         INP_WLOCK_ASSERT(tp->t_inpcb);
4709         /*
4710          * If we are doing PRR and have enough
4711          * room to send <or> we are pacing and prr
4712          * is disabled we will want to see if we
4713          * can send data (by setting r_wanted_output to
4714          * true).
4715          */
4716         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4717             rack->rack_no_prr)
4718                 rack->r_wanted_output = 1;
4719 }
4720
4721 static void
4722 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4723 {
4724         struct tcp_rack *rack;
4725         uint32_t orig_cwnd;
4726
4727         orig_cwnd = tp->snd_cwnd;
4728         INP_WLOCK_ASSERT(tp->t_inpcb);
4729         rack = (struct tcp_rack *)tp->t_fb_ptr;
4730         /* only alert CC if we alerted when we entered */
4731         if (CC_ALGO(tp)->post_recovery != NULL) {
4732                 tp->ccv->curack = th_ack;
4733                 CC_ALGO(tp)->post_recovery(tp->ccv);
4734                 if (tp->snd_cwnd < tp->snd_ssthresh) {
4735                         /*
4736                          * Rack has burst control and pacing
4737                          * so lets not set this any lower than
4738                          * snd_ssthresh per RFC-6582 (option 2).
4739                          */
4740                         tp->snd_cwnd = tp->snd_ssthresh;
4741                 }
4742         }
4743         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4744                 union tcp_log_stackspecific log;
4745                 struct timeval tv;
4746
4747                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4748                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4749                 log.u_bbr.flex1 = th_ack;
4750                 log.u_bbr.flex2 = tp->ccv->flags;
4751                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4752                 log.u_bbr.flex4 = tp->ccv->nsegs;
4753                 log.u_bbr.flex5 = V_tcp_abc_l_var;
4754                 log.u_bbr.flex6 = orig_cwnd;
4755                 log.u_bbr.flex7 = V_tcp_do_newsack;
4756                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4757                 log.u_bbr.flex8 = 2;
4758                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4759                                0, &log, false, NULL, NULL, 0, &tv);
4760         }
4761         if ((rack->rack_no_prr == 0) &&
4762             (rack->no_prr_addback == 0) &&
4763             (rack->r_ctl.rc_prr_sndcnt > 0)) {
4764                 /*
4765                  * Suck the next prr cnt back into cwnd, but
4766                  * only do that if we are not application limited.
4767                  */
4768                 if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4769                         /*
4770                          * We are allowed to add back to the cwnd the amount we did
4771                          * not get out if:
4772                          * a) no_prr_addback is off.
4773                          * b) we are not app limited
4774                          * c) we are doing prr
4775                          * <and>
4776                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4777                          */
4778                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4779                                             rack->r_ctl.rc_prr_sndcnt);
4780                 }
4781                 rack->r_ctl.rc_prr_sndcnt = 0;
4782                 rack_log_to_prr(rack, 1, 0, __LINE__);
4783         }
4784         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
4785         tp->snd_recover = tp->snd_una;
4786         if (rack->r_ctl.dsack_persist) {
4787                 rack->r_ctl.dsack_persist--;
4788                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4789                         rack->r_ctl.num_dsack = 0;
4790                 }
4791                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4792         }
4793         EXIT_RECOVERY(tp->t_flags);
4794 }
4795
4796 static void
4797 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
4798 {
4799         struct tcp_rack *rack;
4800         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4801
4802         INP_WLOCK_ASSERT(tp->t_inpcb);
4803 #ifdef STATS
4804         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4805 #endif
4806         if (IN_RECOVERY(tp->t_flags) == 0) {
4807                 in_rec_at_entry = 0;
4808                 ssthresh_enter = tp->snd_ssthresh;
4809                 cwnd_enter = tp->snd_cwnd;
4810         } else
4811                 in_rec_at_entry = 1;
4812         rack = (struct tcp_rack *)tp->t_fb_ptr;
4813         switch (type) {
4814         case CC_NDUPACK:
4815                 tp->t_flags &= ~TF_WASFRECOVERY;
4816                 tp->t_flags &= ~TF_WASCRECOVERY;
4817                 if (!IN_FASTRECOVERY(tp->t_flags)) {
4818                         rack->r_ctl.rc_prr_delivered = 0;
4819                         rack->r_ctl.rc_prr_out = 0;
4820                         if (rack->rack_no_prr == 0) {
4821                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4822                                 rack_log_to_prr(rack, 2, in_rec_at_entry, line);
4823                         }
4824                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4825                         tp->snd_recover = tp->snd_max;
4826                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4827                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4828                 }
4829                 break;
4830         case CC_ECN:
4831                 if (!IN_CONGRECOVERY(tp->t_flags) ||
4832                     /*
4833                      * Allow ECN reaction on ACK to CWR, if
4834                      * that data segment was also CE marked.
4835                      */
4836                     SEQ_GEQ(ack, tp->snd_recover)) {
4837                         EXIT_CONGRECOVERY(tp->t_flags);
4838                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4839                         tp->snd_recover = tp->snd_max + 1;
4840                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4841                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4842                 }
4843                 break;
4844         case CC_RTO:
4845                 tp->t_dupacks = 0;
4846                 tp->t_bytes_acked = 0;
4847                 EXIT_RECOVERY(tp->t_flags);
4848                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4849                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4850                 orig_cwnd = tp->snd_cwnd;
4851                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4852                 rack_log_to_prr(rack, 16, orig_cwnd, line);
4853                 if (tp->t_flags2 & TF2_ECN_PERMIT)
4854                         tp->t_flags2 |= TF2_ECN_SND_CWR;
4855                 break;
4856         case CC_RTO_ERR:
4857                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4858                 /* RTO was unnecessary, so reset everything. */
4859                 tp->snd_cwnd = tp->snd_cwnd_prev;
4860                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4861                 tp->snd_recover = tp->snd_recover_prev;
4862                 if (tp->t_flags & TF_WASFRECOVERY) {
4863                         ENTER_FASTRECOVERY(tp->t_flags);
4864                         tp->t_flags &= ~TF_WASFRECOVERY;
4865                 }
4866                 if (tp->t_flags & TF_WASCRECOVERY) {
4867                         ENTER_CONGRECOVERY(tp->t_flags);
4868                         tp->t_flags &= ~TF_WASCRECOVERY;
4869                 }
4870                 tp->snd_nxt = tp->snd_max;
4871                 tp->t_badrxtwin = 0;
4872                 break;
4873         }
4874         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4875             (type != CC_RTO)){
4876                 tp->ccv->curack = ack;
4877                 CC_ALGO(tp)->cong_signal(tp->ccv, type);
4878         }
4879         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4880                 rack_log_to_prr(rack, 15, cwnd_enter, line);
4881                 rack->r_ctl.dsack_byte_cnt = 0;
4882                 rack->r_ctl.retran_during_recovery = 0;
4883                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4884                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4885                 rack->r_ent_rec_ns = 1;
4886         }
4887 }
4888
4889 static inline void
4890 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4891 {
4892         uint32_t i_cwnd;
4893
4894         INP_WLOCK_ASSERT(tp->t_inpcb);
4895
4896 #ifdef NETFLIX_STATS
4897         KMOD_TCPSTAT_INC(tcps_idle_restarts);
4898         if (tp->t_state == TCPS_ESTABLISHED)
4899                 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4900 #endif
4901         if (CC_ALGO(tp)->after_idle != NULL)
4902                 CC_ALGO(tp)->after_idle(tp->ccv);
4903
4904         if (tp->snd_cwnd == 1)
4905                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
4906         else
4907                 i_cwnd = rc_init_window(rack);
4908
4909         /*
4910          * Being idle is no different than the initial window. If the cc
4911          * clamps it down below the initial window raise it to the initial
4912          * window.
4913          */
4914         if (tp->snd_cwnd < i_cwnd) {
4915                 tp->snd_cwnd = i_cwnd;
4916         }
4917 }
4918
4919 /*
4920  * Indicate whether this ack should be delayed.  We can delay the ack if
4921  * following conditions are met:
4922  *      - There is no delayed ack timer in progress.
4923  *      - Our last ack wasn't a 0-sized window. We never want to delay
4924  *        the ack that opens up a 0-sized window.
4925  *      - LRO wasn't used for this segment. We make sure by checking that the
4926  *        segment size is not larger than the MSS.
4927  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4928  *        connection.
4929  */
4930 #define DELAY_ACK(tp, tlen)                      \
4931         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4932         ((tp->t_flags & TF_DELACK) == 0) &&      \
4933         (tlen <= tp->t_maxseg) &&                \
4934         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4935
4936 static struct rack_sendmap *
4937 rack_find_lowest_rsm(struct tcp_rack *rack)
4938 {
4939         struct rack_sendmap *rsm;
4940
4941         /*
4942          * Walk the time-order transmitted list looking for an rsm that is
4943          * not acked. This will be the one that was sent the longest time
4944          * ago that is still outstanding.
4945          */
4946         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4947                 if (rsm->r_flags & RACK_ACKED) {
4948                         continue;
4949                 }
4950                 goto finish;
4951         }
4952 finish:
4953         return (rsm);
4954 }
4955
4956 static struct rack_sendmap *
4957 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4958 {
4959         struct rack_sendmap *prsm;
4960
4961         /*
4962          * Walk the sequence order list backward until we hit and arrive at
4963          * the highest seq not acked. In theory when this is called it
4964          * should be the last segment (which it was not).
4965          */
4966         prsm = rsm;
4967         RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4968                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4969                         continue;
4970                 }
4971                 return (prsm);
4972         }
4973         return (NULL);
4974 }
4975
4976 static uint32_t
4977 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4978 {
4979         int32_t lro;
4980         uint32_t thresh;
4981
4982         /*
4983          * lro is the flag we use to determine if we have seen reordering.
4984          * If it gets set we have seen reordering. The reorder logic either
4985          * works in one of two ways:
4986          *
4987          * If reorder-fade is configured, then we track the last time we saw
4988          * re-ordering occur. If we reach the point where enough time as
4989          * passed we no longer consider reordering has occuring.
4990          *
4991          * Or if reorder-face is 0, then once we see reordering we consider
4992          * the connection to alway be subject to reordering and just set lro
4993          * to 1.
4994          *
4995          * In the end if lro is non-zero we add the extra time for
4996          * reordering in.
4997          */
4998         if (srtt == 0)
4999                 srtt = 1;
5000         if (rack->r_ctl.rc_reorder_ts) {
5001                 if (rack->r_ctl.rc_reorder_fade) {
5002                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5003                                 lro = cts - rack->r_ctl.rc_reorder_ts;
5004                                 if (lro == 0) {
5005                                         /*
5006                                          * No time as passed since the last
5007                                          * reorder, mark it as reordering.
5008                                          */
5009                                         lro = 1;
5010                                 }
5011                         } else {
5012                                 /* Negative time? */
5013                                 lro = 0;
5014                         }
5015                         if (lro > rack->r_ctl.rc_reorder_fade) {
5016                                 /* Turn off reordering seen too */
5017                                 rack->r_ctl.rc_reorder_ts = 0;
5018                                 lro = 0;
5019                         }
5020                 } else {
5021                         /* Reodering does not fade */
5022                         lro = 1;
5023                 }
5024         } else {
5025                 lro = 0;
5026         }
5027         if (rack->rc_rack_tmr_std_based == 0) {
5028                 thresh = srtt + rack->r_ctl.rc_pkt_delay;
5029         } else {
5030                 /* Standards based pkt-delay is 1/4 srtt */
5031                 thresh = srtt +  (srtt >> 2);
5032         }
5033         if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5034                 /* It must be set, if not you get 1/4 rtt */
5035                 if (rack->r_ctl.rc_reorder_shift)
5036                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5037                 else
5038                         thresh += (srtt >> 2);
5039         }
5040         if (rack->rc_rack_use_dsack &&
5041             lro &&
5042             (rack->r_ctl.num_dsack > 0)) {
5043                 /*
5044                  * We only increase the reordering window if we
5045                  * have seen reordering <and> we have a DSACK count.
5046                  */
5047                 thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5048                 rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
5049         }
5050         /* SRTT * 2 is the ceiling */
5051         if (thresh > (srtt * 2)) {
5052                 thresh = srtt * 2;
5053         }
5054         /* And we don't want it above the RTO max either */
5055         if (thresh > rack_rto_max) {
5056                 thresh = rack_rto_max;
5057         }
5058         rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
5059         return (thresh);
5060 }
5061
5062 static uint32_t
5063 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5064                      struct rack_sendmap *rsm, uint32_t srtt)
5065 {
5066         struct rack_sendmap *prsm;
5067         uint32_t thresh, len;
5068         int segsiz;
5069
5070         if (srtt == 0)
5071                 srtt = 1;
5072         if (rack->r_ctl.rc_tlp_threshold)
5073                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5074         else
5075                 thresh = (srtt * 2);
5076
5077         /* Get the previous sent packet, if any */
5078         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5079         len = rsm->r_end - rsm->r_start;
5080         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5081                 /* Exactly like the ID */
5082                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= 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         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5092                 /* 2.1 behavior */
5093                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5094                 if (prsm && (len <= segsiz)) {
5095                         /*
5096                          * Two packets outstanding, thresh should be (2*srtt) +
5097                          * possible inter-packet delay (if any).
5098                          */
5099                         uint32_t inter_gap = 0;
5100                         int idx, nidx;
5101
5102                         idx = rsm->r_rtr_cnt - 1;
5103                         nidx = prsm->r_rtr_cnt - 1;
5104                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5105                                 /* Yes it was sent later (or at the same time) */
5106                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5107                         }
5108                         thresh += inter_gap;
5109                 } else if (len <= segsiz) {
5110                         /*
5111                          * Possibly compensate for delayed-ack.
5112                          */
5113                         uint32_t alt_thresh;
5114
5115                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5116                         if (alt_thresh > thresh)
5117                                 thresh = alt_thresh;
5118                 }
5119         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5120                 /* 2.2 behavior */
5121                 if (len <= segsiz) {
5122                         uint32_t alt_thresh;
5123                         /*
5124                          * Compensate for delayed-ack with the d-ack time.
5125                          */
5126                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5127                         if (alt_thresh > thresh)
5128                                 thresh = alt_thresh;
5129                 }
5130         }
5131         /* Not above an RTO */
5132         if (thresh > tp->t_rxtcur) {
5133                 thresh = tp->t_rxtcur;
5134         }
5135         /* Not above a RTO max */
5136         if (thresh > rack_rto_max) {
5137                 thresh = rack_rto_max;
5138         }
5139         /* Apply user supplied min TLP */
5140         if (thresh < rack_tlp_min) {
5141                 thresh = rack_tlp_min;
5142         }
5143         return (thresh);
5144 }
5145
5146 static uint32_t
5147 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5148 {
5149         /*
5150          * We want the rack_rtt which is the
5151          * last rtt we measured. However if that
5152          * does not exist we fallback to the srtt (which
5153          * we probably will never do) and then as a last
5154          * resort we use RACK_INITIAL_RTO if no srtt is
5155          * yet set.
5156          */
5157         if (rack->rc_rack_rtt)
5158                 return (rack->rc_rack_rtt);
5159         else if (tp->t_srtt == 0)
5160                 return (RACK_INITIAL_RTO);
5161         return (tp->t_srtt);
5162 }
5163
5164 static struct rack_sendmap *
5165 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5166 {
5167         /*
5168          * Check to see that we don't need to fall into recovery. We will
5169          * need to do so if our oldest transmit is past the time we should
5170          * have had an ack.
5171          */
5172         struct tcp_rack *rack;
5173         struct rack_sendmap *rsm;
5174         int32_t idx;
5175         uint32_t srtt, thresh;
5176
5177         rack = (struct tcp_rack *)tp->t_fb_ptr;
5178         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5179                 return (NULL);
5180         }
5181         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5182         if (rsm == NULL)
5183                 return (NULL);
5184
5185
5186         if (rsm->r_flags & RACK_ACKED) {
5187                 rsm = rack_find_lowest_rsm(rack);
5188                 if (rsm == NULL)
5189                         return (NULL);
5190         }
5191         idx = rsm->r_rtr_cnt - 1;
5192         srtt = rack_grab_rtt(tp, rack);
5193         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5194         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5195                 return (NULL);
5196         }
5197         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5198                 return (NULL);
5199         }
5200         /* Ok if we reach here we are over-due and this guy can be sent */
5201         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
5202         return (rsm);
5203 }
5204
5205 static uint32_t
5206 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5207 {
5208         int32_t t;
5209         int32_t tt;
5210         uint32_t ret_val;
5211
5212         t = (tp->t_srtt + (tp->t_rttvar << 2));
5213         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5214             rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5215         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5216         ret_val = (uint32_t)tt;
5217         return (ret_val);
5218 }
5219
5220 static uint32_t
5221 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5222 {
5223         /*
5224          * Start the FR timer, we do this based on getting the first one in
5225          * the rc_tmap. Note that if its NULL we must stop the timer. in all
5226          * events we need to stop the running timer (if its running) before
5227          * starting the new one.
5228          */
5229         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5230         uint32_t srtt_cur;
5231         int32_t idx;
5232         int32_t is_tlp_timer = 0;
5233         struct rack_sendmap *rsm;
5234
5235         if (rack->t_timers_stopped) {
5236                 /* All timers have been stopped none are to run */
5237                 return (0);
5238         }
5239         if (rack->rc_in_persist) {
5240                 /* We can't start any timer in persists */
5241                 return (rack_get_persists_timer_val(tp, rack));
5242         }
5243         rack->rc_on_min_to = 0;
5244         if ((tp->t_state < TCPS_ESTABLISHED) ||
5245             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5246                 goto activate_rxt;
5247         }
5248         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5249         if ((rsm == NULL) || sup_rack) {
5250                 /* Nothing on the send map or no rack */
5251 activate_rxt:
5252                 time_since_sent = 0;
5253                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5254                 if (rsm) {
5255                         /*
5256                          * Should we discount the RTX timer any?
5257                          *
5258                          * We want to discount it the smallest amount.
5259                          * If a timer (Rack/TLP or RXT) has gone off more
5260                          * recently thats the discount we want to use (now - timer time).
5261                          * If the retransmit of the oldest packet was more recent then
5262                          * we want to use that (now - oldest-packet-last_transmit_time).
5263                          *
5264                          */
5265                         idx = rsm->r_rtr_cnt - 1;
5266                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5267                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5268                         else
5269                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5270                         if (TSTMP_GT(cts, tstmp_touse))
5271                             time_since_sent = cts - tstmp_touse;
5272                 }
5273                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5274                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5275                         to = tp->t_rxtcur;
5276                         if (to > time_since_sent)
5277                                 to -= time_since_sent;
5278                         else
5279                                 to = rack->r_ctl.rc_min_to;
5280                         if (to == 0)
5281                                 to = 1;
5282                         /* Special case for KEEPINIT */
5283                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5284                             (TP_KEEPINIT(tp) != 0) &&
5285                             rsm) {
5286                                 /*
5287                                  * We have to put a ceiling on the rxt timer
5288                                  * of the keep-init timeout.
5289                                  */
5290                                 uint32_t max_time, red;
5291
5292                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5293                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5294                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5295                                         if (red < max_time)
5296                                                 max_time -= red;
5297                                         else
5298                                                 max_time = 1;
5299                                 }
5300                                 /* Reduce timeout to the keep value if needed */
5301                                 if (max_time < to)
5302                                         to = max_time;
5303                         }
5304                         return (to);
5305                 }
5306                 return (0);
5307         }
5308         if (rsm->r_flags & RACK_ACKED) {
5309                 rsm = rack_find_lowest_rsm(rack);
5310                 if (rsm == NULL) {
5311                         /* No lowest? */
5312                         goto activate_rxt;
5313                 }
5314         }
5315         if (rack->sack_attack_disable) {
5316                 /*
5317                  * We don't want to do
5318                  * any TLP's if you are an attacker.
5319                  * Though if you are doing what
5320                  * is expected you may still have
5321                  * SACK-PASSED marks.
5322                  */
5323                 goto activate_rxt;
5324         }
5325         /* Convert from ms to usecs */
5326         if ((rsm->r_flags & RACK_SACK_PASSED) ||
5327             (rsm->r_flags & RACK_RWND_COLLAPSED) ||
5328             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5329                 if ((tp->t_flags & TF_SENTFIN) &&
5330                     ((tp->snd_max - tp->snd_una) == 1) &&
5331                     (rsm->r_flags & RACK_HAS_FIN)) {
5332                         /*
5333                          * We don't start a rack timer if all we have is a
5334                          * FIN outstanding.
5335                          */
5336                         goto activate_rxt;
5337                 }
5338                 if ((rack->use_rack_rr == 0) &&
5339                     (IN_FASTRECOVERY(tp->t_flags)) &&
5340                     (rack->rack_no_prr == 0) &&
5341                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5342                         /*
5343                          * We are not cheating, in recovery  and
5344                          * not enough ack's to yet get our next
5345                          * retransmission out.
5346                          *
5347                          * Note that classified attackers do not
5348                          * get to use the rack-cheat.
5349                          */
5350                         goto activate_tlp;
5351                 }
5352                 srtt = rack_grab_rtt(tp, rack);
5353                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5354                 idx = rsm->r_rtr_cnt - 1;
5355                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5356                 if (SEQ_GEQ(exp, cts)) {
5357                         to = exp - cts;
5358                         if (to < rack->r_ctl.rc_min_to) {
5359                                 to = rack->r_ctl.rc_min_to;
5360                                 if (rack->r_rr_config == 3)
5361                                         rack->rc_on_min_to = 1;
5362                         }
5363                 } else {
5364                         to = rack->r_ctl.rc_min_to;
5365                         if (rack->r_rr_config == 3)
5366                                 rack->rc_on_min_to = 1;
5367                 }
5368         } else {
5369                 /* Ok we need to do a TLP not RACK */
5370 activate_tlp:
5371                 if ((rack->rc_tlp_in_progress != 0) &&
5372                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5373                         /*
5374                          * The previous send was a TLP and we have sent
5375                          * N TLP's without sending new data.
5376                          */
5377                         goto activate_rxt;
5378                 }
5379                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5380                 if (rsm == NULL) {
5381                         /* We found no rsm to TLP with. */
5382                         goto activate_rxt;
5383                 }
5384                 if (rsm->r_flags & RACK_HAS_FIN) {
5385                         /* If its a FIN we dont do TLP */
5386                         rsm = NULL;
5387                         goto activate_rxt;
5388                 }
5389                 idx = rsm->r_rtr_cnt - 1;
5390                 time_since_sent = 0;
5391                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5392                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5393                 else
5394                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5395                 if (TSTMP_GT(cts, tstmp_touse))
5396                     time_since_sent = cts - tstmp_touse;
5397                 is_tlp_timer = 1;
5398                 if (tp->t_srtt) {
5399                         if ((rack->rc_srtt_measure_made == 0) &&
5400                             (tp->t_srtt == 1)) {
5401                                 /*
5402                                  * If another stack as run and set srtt to 1,
5403                                  * then the srtt was 0, so lets use the initial.
5404                                  */
5405                                 srtt = RACK_INITIAL_RTO;
5406                         } else {
5407                                 srtt_cur = tp->t_srtt;
5408                                 srtt = srtt_cur;
5409                         }
5410                 } else
5411                         srtt = RACK_INITIAL_RTO;
5412                 /*
5413                  * If the SRTT is not keeping up and the
5414                  * rack RTT has spiked we want to use
5415                  * the last RTT not the smoothed one.
5416                  */
5417                 if (rack_tlp_use_greater &&
5418                     tp->t_srtt &&
5419                     (srtt < rack_grab_rtt(tp, rack))) {
5420                         srtt = rack_grab_rtt(tp, rack);
5421                 }
5422                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5423                 if (thresh > time_since_sent) {
5424                         to = thresh - time_since_sent;
5425                 } else {
5426                         to = rack->r_ctl.rc_min_to;
5427                         rack_log_alt_to_to_cancel(rack,
5428                                                   thresh,               /* flex1 */
5429                                                   time_since_sent,      /* flex2 */
5430                                                   tstmp_touse,          /* flex3 */
5431                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5432                                                   (uint32_t)rsm->r_tim_lastsent[idx],
5433                                                   srtt,
5434                                                   idx, 99);
5435                 }
5436                 if (to < rack_tlp_min) {
5437                         to = rack_tlp_min;
5438                 }
5439                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5440                         /*
5441                          * If the TLP time works out to larger than the max
5442                          * RTO lets not do TLP.. just RTO.
5443                          */
5444                         goto activate_rxt;
5445                 }
5446         }
5447         if (is_tlp_timer == 0) {
5448                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5449         } else {
5450                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5451         }
5452         if (to == 0)
5453                 to = 1;
5454         return (to);
5455 }
5456
5457 static void
5458 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5459 {
5460         if (rack->rc_in_persist == 0) {
5461                 if (tp->t_flags & TF_GPUTINPROG) {
5462                         /*
5463                          * Stop the goodput now, the calling of the
5464                          * measurement function clears the flag.
5465                          */
5466                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5467                                                     RACK_QUALITY_PERSIST);
5468                 }
5469 #ifdef NETFLIX_SHARED_CWND
5470                 if (rack->r_ctl.rc_scw) {
5471                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5472                         rack->rack_scwnd_is_idle = 1;
5473                 }
5474 #endif
5475                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5476                 if (rack->r_ctl.rc_went_idle_time == 0)
5477                         rack->r_ctl.rc_went_idle_time = 1;
5478                 rack_timer_cancel(tp, rack, cts, __LINE__);
5479                 rack->r_ctl.persist_lost_ends = 0;
5480                 rack->probe_not_answered = 0;
5481                 rack->forced_ack = 0;
5482                 tp->t_rxtshift = 0;
5483                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5484                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5485                 rack->rc_in_persist = 1;
5486         }
5487 }
5488
5489 static void
5490 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5491 {
5492         if (tcp_in_hpts(rack->rc_inp)) {
5493                 tcp_hpts_remove(rack->rc_inp);
5494                 rack->r_ctl.rc_hpts_flags = 0;
5495         }
5496 #ifdef NETFLIX_SHARED_CWND
5497         if (rack->r_ctl.rc_scw) {
5498                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5499                 rack->rack_scwnd_is_idle = 0;
5500         }
5501 #endif
5502         if (rack->rc_gp_dyn_mul &&
5503             (rack->use_fixed_rate == 0) &&
5504             (rack->rc_always_pace)) {
5505                 /*
5506                  * Do we count this as if a probe-rtt just
5507                  * finished?
5508                  */
5509                 uint32_t time_idle, idle_min;
5510
5511                 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5512                 idle_min = rack_min_probertt_hold;
5513                 if (rack_probertt_gpsrtt_cnt_div) {
5514                         uint64_t extra;
5515                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5516                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5517                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5518                         idle_min += (uint32_t)extra;
5519                 }
5520                 if (time_idle >= idle_min) {
5521                         /* Yes, we count it as a probe-rtt. */
5522                         uint32_t us_cts;
5523
5524                         us_cts = tcp_get_usecs(NULL);
5525                         if (rack->in_probe_rtt == 0) {
5526                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5527                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5528                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5529                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5530                         } else {
5531                                 rack_exit_probertt(rack, us_cts);
5532                         }
5533                 }
5534         }
5535         rack->rc_in_persist = 0;
5536         rack->r_ctl.rc_went_idle_time = 0;
5537         tp->t_rxtshift = 0;
5538         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5539            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5540         rack->r_ctl.rc_agg_delayed = 0;
5541         rack->r_early = 0;
5542         rack->r_late = 0;
5543         rack->r_ctl.rc_agg_early = 0;
5544 }
5545
5546 static void
5547 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5548                    struct hpts_diag *diag, struct timeval *tv)
5549 {
5550         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5551                 union tcp_log_stackspecific log;
5552
5553                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5554                 log.u_bbr.flex1 = diag->p_nxt_slot;
5555                 log.u_bbr.flex2 = diag->p_cur_slot;
5556                 log.u_bbr.flex3 = diag->slot_req;
5557                 log.u_bbr.flex4 = diag->inp_hptsslot;
5558                 log.u_bbr.flex5 = diag->slot_remaining;
5559                 log.u_bbr.flex6 = diag->need_new_to;
5560                 log.u_bbr.flex7 = diag->p_hpts_active;
5561                 log.u_bbr.flex8 = diag->p_on_min_sleep;
5562                 /* Hijack other fields as needed */
5563                 log.u_bbr.epoch = diag->have_slept;
5564                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5565                 log.u_bbr.pkts_out = diag->co_ret;
5566                 log.u_bbr.applimited = diag->hpts_sleep_time;
5567                 log.u_bbr.delivered = diag->p_prev_slot;
5568                 log.u_bbr.inflight = diag->p_runningslot;
5569                 log.u_bbr.bw_inuse = diag->wheel_slot;
5570                 log.u_bbr.rttProp = diag->wheel_cts;
5571                 log.u_bbr.timeStamp = cts;
5572                 log.u_bbr.delRate = diag->maxslots;
5573                 log.u_bbr.cur_del_rate = diag->p_curtick;
5574                 log.u_bbr.cur_del_rate <<= 32;
5575                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5576                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5577                     &rack->rc_inp->inp_socket->so_rcv,
5578                     &rack->rc_inp->inp_socket->so_snd,
5579                     BBR_LOG_HPTSDIAG, 0,
5580                     0, &log, false, tv);
5581         }
5582
5583 }
5584
5585 static void
5586 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5587 {
5588         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5589                 union tcp_log_stackspecific log;
5590                 struct timeval tv;
5591
5592                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5593                 log.u_bbr.flex1 = sb->sb_flags;
5594                 log.u_bbr.flex2 = len;
5595                 log.u_bbr.flex3 = sb->sb_state;
5596                 log.u_bbr.flex8 = type;
5597                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5598                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5599                     &rack->rc_inp->inp_socket->so_rcv,
5600                     &rack->rc_inp->inp_socket->so_snd,
5601                     TCP_LOG_SB_WAKE, 0,
5602                     len, &log, false, &tv);
5603         }
5604 }
5605
5606 static void
5607 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5608       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5609 {
5610         struct hpts_diag diag;
5611         struct inpcb *inp;
5612         struct timeval tv;
5613         uint32_t delayed_ack = 0;
5614         uint32_t hpts_timeout;
5615         uint32_t entry_slot = slot;
5616         uint8_t stopped;
5617         uint32_t left = 0;
5618         uint32_t us_cts;
5619
5620         inp = tp->t_inpcb;
5621         if ((tp->t_state == TCPS_CLOSED) ||
5622             (tp->t_state == TCPS_LISTEN)) {
5623                 return;
5624         }
5625         if (tcp_in_hpts(inp)) {
5626                 /* Already on the pacer */
5627                 return;
5628         }
5629         stopped = rack->rc_tmr_stopped;
5630         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5631                 left = rack->r_ctl.rc_timer_exp - cts;
5632         }
5633         rack->r_ctl.rc_timer_exp = 0;
5634         rack->r_ctl.rc_hpts_flags = 0;
5635         us_cts = tcp_get_usecs(&tv);
5636         /* Now early/late accounting */
5637         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5638         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5639                 /*
5640                  * We have a early carry over set,
5641                  * we can always add more time so we
5642                  * can always make this compensation.
5643                  *
5644                  * Note if ack's are allowed to wake us do not
5645                  * penalize the next timer for being awoke
5646                  * by an ack aka the rc_agg_early (non-paced mode).
5647                  */
5648                 slot += rack->r_ctl.rc_agg_early;
5649                 rack->r_early = 0;
5650                 rack->r_ctl.rc_agg_early = 0;
5651         }
5652         if (rack->r_late) {
5653                 /*
5654                  * This is harder, we can
5655                  * compensate some but it
5656                  * really depends on what
5657                  * the current pacing time is.
5658                  */
5659                 if (rack->r_ctl.rc_agg_delayed >= slot) {
5660                         /*
5661                          * We can't compensate for it all.
5662                          * And we have to have some time
5663                          * on the clock. We always have a min
5664                          * 10 slots (10 x 10 i.e. 100 usecs).
5665                          */
5666                         if (slot <= HPTS_TICKS_PER_SLOT) {
5667                                 /* We gain delay */
5668                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5669                                 slot = HPTS_TICKS_PER_SLOT;
5670                         } else {
5671                                 /* We take off some */
5672                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5673                                 slot = HPTS_TICKS_PER_SLOT;
5674                         }
5675                 } else {
5676                         slot -= rack->r_ctl.rc_agg_delayed;
5677                         rack->r_ctl.rc_agg_delayed = 0;
5678                         /* Make sure we have 100 useconds at minimum */
5679                         if (slot < HPTS_TICKS_PER_SLOT) {
5680                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5681                                 slot = HPTS_TICKS_PER_SLOT;
5682                         }
5683                         if (rack->r_ctl.rc_agg_delayed == 0)
5684                                 rack->r_late = 0;
5685                 }
5686         }
5687         if (slot) {
5688                 /* We are pacing too */
5689                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5690         }
5691         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5692 #ifdef NETFLIX_EXP_DETECTION
5693         if (rack->sack_attack_disable &&
5694             (slot < tcp_sad_pacing_interval)) {
5695                 /*
5696                  * We have a potential attacker on
5697                  * the line. We have possibly some
5698                  * (or now) pacing time set. We want to
5699                  * slow down the processing of sacks by some
5700                  * amount (if it is an attacker). Set the default
5701                  * slot for attackers in place (unless the orginal
5702                  * interval is longer). Its stored in
5703                  * micro-seconds, so lets convert to msecs.
5704                  */
5705                 slot = tcp_sad_pacing_interval;
5706         }
5707 #endif
5708         if (tp->t_flags & TF_DELACK) {
5709                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5710                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5711         }
5712         if (delayed_ack && ((hpts_timeout == 0) ||
5713                             (delayed_ack < hpts_timeout)))
5714                 hpts_timeout = delayed_ack;
5715         else
5716                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5717         /*
5718          * If no timers are going to run and we will fall off the hptsi
5719          * wheel, we resort to a keep-alive timer if its configured.
5720          */
5721         if ((hpts_timeout == 0) &&
5722             (slot == 0)) {
5723                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5724                     (tp->t_state <= TCPS_CLOSING)) {
5725                         /*
5726                          * Ok we have no timer (persists, rack, tlp, rxt  or
5727                          * del-ack), we don't have segments being paced. So
5728                          * all that is left is the keepalive timer.
5729                          */
5730                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5731                                 /* Get the established keep-alive time */
5732                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5733                         } else {
5734                                 /*
5735                                  * Get the initial setup keep-alive time,
5736                                  * note that this is probably not going to
5737                                  * happen, since rack will be running a rxt timer
5738                                  * if a SYN of some sort is outstanding. It is
5739                                  * actually handled in rack_timeout_rxt().
5740                                  */
5741                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5742                         }
5743                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5744                         if (rack->in_probe_rtt) {
5745                                 /*
5746                                  * We want to instead not wake up a long time from
5747                                  * now but to wake up about the time we would
5748                                  * exit probe-rtt and initiate a keep-alive ack.
5749                                  * This will get us out of probe-rtt and update
5750                                  * our min-rtt.
5751                                  */
5752                                 hpts_timeout = rack_min_probertt_hold;
5753                         }
5754                 }
5755         }
5756         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5757             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5758                 /*
5759                  * RACK, TLP, persists and RXT timers all are restartable
5760                  * based on actions input .. i.e we received a packet (ack
5761                  * or sack) and that changes things (rw, or snd_una etc).
5762                  * Thus we can restart them with a new value. For
5763                  * keep-alive, delayed_ack we keep track of what was left
5764                  * and restart the timer with a smaller value.
5765                  */
5766                 if (left < hpts_timeout)
5767                         hpts_timeout = left;
5768         }
5769         if (hpts_timeout) {
5770                 /*
5771                  * Hack alert for now we can't time-out over 2,147,483
5772                  * seconds (a bit more than 596 hours), which is probably ok
5773                  * :).
5774                  */
5775                 if (hpts_timeout > 0x7ffffffe)
5776                         hpts_timeout = 0x7ffffffe;
5777                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5778         }
5779         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5780         if ((rack->gp_ready == 0) &&
5781             (rack->use_fixed_rate == 0) &&
5782             (hpts_timeout < slot) &&
5783             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5784                 /*
5785                  * We have no good estimate yet for the
5786                  * old clunky burst mitigation or the
5787                  * real pacing. And the tlp or rxt is smaller
5788                  * than the pacing calculation. Lets not
5789                  * pace that long since we know the calculation
5790                  * so far is not accurate.
5791                  */
5792                 slot = hpts_timeout;
5793         }
5794         /**
5795          * Turn off all the flags for queuing by default. The
5796          * flags have important meanings to what happens when
5797          * LRO interacts with the transport. Most likely (by default now)
5798          * mbuf_queueing and ack compression are on. So the transport
5799          * has a couple of flags that control what happens (if those
5800          * are not on then these flags won't have any effect since it
5801          * won't go through the queuing LRO path).
5802          *
5803          * INP_MBUF_QUEUE_READY - This flags says that I am busy
5804          *                        pacing output, so don't disturb. But
5805          *                        it also means LRO can wake me if there
5806          *                        is a SACK arrival.
5807          *
5808          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5809          *                       with the above flag (QUEUE_READY) and
5810          *                       when present it says don't even wake me
5811          *                       if a SACK arrives.
5812          *
5813          * The idea behind these flags is that if we are pacing we
5814          * set the MBUF_QUEUE_READY and only get woken up if
5815          * a SACK arrives (which could change things) or if
5816          * our pacing timer expires. If, however, we have a rack
5817          * timer running, then we don't even want a sack to wake
5818          * us since the rack timer has to expire before we can send.
5819          *
5820          * Other cases should usually have none of the flags set
5821          * so LRO can call into us.
5822          */
5823         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5824         if (slot) {
5825                 rack->r_ctl.rc_last_output_to = us_cts + slot;
5826                 /*
5827                  * A pacing timer (slot) is being set, in
5828                  * such a case we cannot send (we are blocked by
5829                  * the timer). So lets tell LRO that it should not
5830                  * wake us unless there is a SACK. Note this only
5831                  * will be effective if mbuf queueing is on or
5832                  * compressed acks are being processed.
5833                  */
5834                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5835                 /*
5836                  * But wait if we have a Rack timer running
5837                  * even a SACK should not disturb us (with
5838                  * the exception of r_rr_config 3).
5839                  */
5840                 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5841                     (rack->r_rr_config != 3))
5842                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5843                 if (rack->rc_ack_can_sendout_data) {
5844                         /*
5845                          * Ahh but wait, this is that special case
5846                          * where the pacing timer can be disturbed
5847                          * backout the changes (used for non-paced
5848                          * burst limiting).
5849                          */
5850                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5851                 }
5852                 if ((rack->use_rack_rr) &&
5853                     (rack->r_rr_config < 2) &&
5854                     ((hpts_timeout) && (hpts_timeout < slot))) {
5855                         /*
5856                          * Arrange for the hpts to kick back in after the
5857                          * t-o if the t-o does not cause a send.
5858                          */
5859                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5860                                                    __LINE__, &diag);
5861                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5862                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5863                 } else {
5864                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
5865                                                    __LINE__, &diag);
5866                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5867                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5868                 }
5869         } else if (hpts_timeout) {
5870                 /*
5871                  * With respect to inp_flags2 here, lets let any new acks wake
5872                  * us up here. Since we are not pacing (no pacing timer), output
5873                  * can happen so we should let it. If its a Rack timer, then any inbound
5874                  * packet probably won't change the sending (we will be blocked)
5875                  * but it may change the prr stats so letting it in (the set defaults
5876                  * at the start of this block) are good enough.
5877                  */
5878                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5879                                            __LINE__, &diag);
5880                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5881                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5882         } else {
5883                 /* No timer starting */
5884 #ifdef INVARIANTS
5885                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5886                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5887                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5888                 }
5889 #endif
5890         }
5891         rack->rc_tmr_stopped = 0;
5892         if (slot)
5893                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5894 }
5895
5896 /*
5897  * RACK Timer, here we simply do logging and house keeping.
5898  * the normal rack_output() function will call the
5899  * appropriate thing to check if we need to do a RACK retransmit.
5900  * We return 1, saying don't proceed with rack_output only
5901  * when all timers have been stopped (destroyed PCB?).
5902  */
5903 static int
5904 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5905 {
5906         /*
5907          * This timer simply provides an internal trigger to send out data.
5908          * The check_recovery_mode call will see if there are needed
5909          * retransmissions, if so we will enter fast-recovery. The output
5910          * call may or may not do the same thing depending on sysctl
5911          * settings.
5912          */
5913         struct rack_sendmap *rsm;
5914
5915         if (tp->t_timers->tt_flags & TT_STOPPED) {
5916                 return (1);
5917         }
5918         counter_u64_add(rack_to_tot, 1);
5919         if (rack->r_state && (rack->r_state != tp->t_state))
5920                 rack_set_state(tp, rack);
5921         rack->rc_on_min_to = 0;
5922         rsm = rack_check_recovery_mode(tp, cts);
5923         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5924         if (rsm) {
5925                 rack->r_ctl.rc_resend = rsm;
5926                 rack->r_timer_override = 1;
5927                 if (rack->use_rack_rr) {
5928                         /*
5929                          * Don't accumulate extra pacing delay
5930                          * we are allowing the rack timer to
5931                          * over-ride pacing i.e. rrr takes precedence
5932                          * if the pacing interval is longer than the rrr
5933                          * time (in other words we get the min pacing
5934                          * time versus rrr pacing time).
5935                          */
5936                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5937                 }
5938         }
5939         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5940         if (rsm == NULL) {
5941                 /* restart a timer and return 1 */
5942                 rack_start_hpts_timer(rack, tp, cts,
5943                                       0, 0, 0);
5944                 return (1);
5945         }
5946         return (0);
5947 }
5948
5949 static void
5950 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5951 {
5952         if (rsm->m->m_len > rsm->orig_m_len) {
5953                 /*
5954                  * Mbuf grew, caused by sbcompress, our offset does
5955                  * not change.
5956                  */
5957                 rsm->orig_m_len = rsm->m->m_len;
5958         } else if (rsm->m->m_len < rsm->orig_m_len) {
5959                 /*
5960                  * Mbuf shrank, trimmed off the top by an ack, our
5961                  * offset changes.
5962                  */
5963                 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
5964                 rsm->orig_m_len = rsm->m->m_len;
5965         }
5966 }
5967
5968 static void
5969 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
5970 {
5971         struct mbuf *m;
5972         uint32_t soff;
5973
5974         if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
5975                 /* Fix up the orig_m_len and possibly the mbuf offset */
5976                 rack_adjust_orig_mlen(src_rsm);
5977         }
5978         m = src_rsm->m;
5979         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
5980         while (soff >= m->m_len) {
5981                 /* Move out past this mbuf */
5982                 soff -= m->m_len;
5983                 m = m->m_next;
5984                 KASSERT((m != NULL),
5985                         ("rsm:%p nrsm:%p hit at soff:%u null m",
5986                          src_rsm, rsm, soff));
5987         }
5988         rsm->m = m;
5989         rsm->soff = soff;
5990         rsm->orig_m_len = m->m_len;
5991 }
5992
5993 static __inline void
5994 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5995                struct rack_sendmap *rsm, uint32_t start)
5996 {
5997         int idx;
5998
5999         nrsm->r_start = start;
6000         nrsm->r_end = rsm->r_end;
6001         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6002         nrsm->r_flags = rsm->r_flags;
6003         nrsm->r_dupack = rsm->r_dupack;
6004         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6005         nrsm->r_rtr_bytes = 0;
6006         nrsm->r_fas = rsm->r_fas;
6007         rsm->r_end = nrsm->r_start;
6008         nrsm->r_just_ret = rsm->r_just_ret;
6009         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6010                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6011         }
6012         /* Now if we have SYN flag we keep it on the left edge */
6013         if (nrsm->r_flags & RACK_HAS_SYN)
6014                 nrsm->r_flags &= ~RACK_HAS_SYN;
6015         /* Now if we have a FIN flag we keep it on the right edge */
6016         if (rsm->r_flags & RACK_HAS_FIN)
6017                 rsm->r_flags &= ~RACK_HAS_FIN;
6018         /* Push bit must go to the right edge as well */
6019         if (rsm->r_flags & RACK_HAD_PUSH)
6020                 rsm->r_flags &= ~RACK_HAD_PUSH;
6021         /* Clone over the state of the hw_tls flag */
6022         nrsm->r_hw_tls = rsm->r_hw_tls;
6023         /*
6024          * Now we need to find nrsm's new location in the mbuf chain
6025          * we basically calculate a new offset, which is soff +
6026          * how much is left in original rsm. Then we walk out the mbuf
6027          * chain to find the righ position, it may be the same mbuf
6028          * or maybe not.
6029          */
6030         KASSERT(((rsm->m != NULL) ||
6031                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6032                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6033         if (rsm->m)
6034                 rack_setup_offset_for_rsm(rsm, nrsm);
6035 }
6036
6037 static struct rack_sendmap *
6038 rack_merge_rsm(struct tcp_rack *rack,
6039                struct rack_sendmap *l_rsm,
6040                struct rack_sendmap *r_rsm)
6041 {
6042         /*
6043          * We are merging two ack'd RSM's,
6044          * the l_rsm is on the left (lower seq
6045          * values) and the r_rsm is on the right
6046          * (higher seq value). The simplest way
6047          * to merge these is to move the right
6048          * one into the left. I don't think there
6049          * is any reason we need to try to find
6050          * the oldest (or last oldest retransmitted).
6051          */
6052 #ifdef INVARIANTS
6053         struct rack_sendmap *rm;
6054 #endif
6055         rack_log_map_chg(rack->rc_tp, rack, NULL,
6056                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6057         l_rsm->r_end = r_rsm->r_end;
6058         if (l_rsm->r_dupack < r_rsm->r_dupack)
6059                 l_rsm->r_dupack = r_rsm->r_dupack;
6060         if (r_rsm->r_rtr_bytes)
6061                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6062         if (r_rsm->r_in_tmap) {
6063                 /* This really should not happen */
6064                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6065                 r_rsm->r_in_tmap = 0;
6066         }
6067
6068         /* Now the flags */
6069         if (r_rsm->r_flags & RACK_HAS_FIN)
6070                 l_rsm->r_flags |= RACK_HAS_FIN;
6071         if (r_rsm->r_flags & RACK_TLP)
6072                 l_rsm->r_flags |= RACK_TLP;
6073         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6074                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6075         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6076             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6077                 /*
6078                  * If both are app-limited then let the
6079                  * free lower the count. If right is app
6080                  * limited and left is not, transfer.
6081                  */
6082                 l_rsm->r_flags |= RACK_APP_LIMITED;
6083                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
6084                 if (r_rsm == rack->r_ctl.rc_first_appl)
6085                         rack->r_ctl.rc_first_appl = l_rsm;
6086         }
6087 #ifndef INVARIANTS
6088         (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6089 #else
6090         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6091         if (rm != r_rsm) {
6092                 panic("removing head in rack:%p rsm:%p rm:%p",
6093                       rack, r_rsm, rm);
6094         }
6095 #endif
6096         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6097                 /* Transfer the split limit to the map we free */
6098                 r_rsm->r_limit_type = l_rsm->r_limit_type;
6099                 l_rsm->r_limit_type = 0;
6100         }
6101         rack_free(rack, r_rsm);
6102         return (l_rsm);
6103 }
6104
6105 /*
6106  * TLP Timer, here we simply setup what segment we want to
6107  * have the TLP expire on, the normal rack_output() will then
6108  * send it out.
6109  *
6110  * We return 1, saying don't proceed with rack_output only
6111  * when all timers have been stopped (destroyed PCB?).
6112  */
6113 static int
6114 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6115 {
6116         /*
6117          * Tail Loss Probe.
6118          */
6119         struct rack_sendmap *rsm = NULL;
6120 #ifdef INVARIANTS
6121         struct rack_sendmap *insret;
6122 #endif
6123         struct socket *so;
6124         uint32_t amm;
6125         uint32_t out, avail;
6126         int collapsed_win = 0;
6127
6128         if (tp->t_timers->tt_flags & TT_STOPPED) {
6129                 return (1);
6130         }
6131         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6132                 /* Its not time yet */
6133                 return (0);
6134         }
6135         if (ctf_progress_timeout_check(tp, true)) {
6136                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6137                 return (-ETIMEDOUT);    /* tcp_drop() */
6138         }
6139         /*
6140          * A TLP timer has expired. We have been idle for 2 rtts. So we now
6141          * need to figure out how to force a full MSS segment out.
6142          */
6143         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6144         rack->r_ctl.retran_during_recovery = 0;
6145         rack->r_ctl.dsack_byte_cnt = 0;
6146         counter_u64_add(rack_tlp_tot, 1);
6147         if (rack->r_state && (rack->r_state != tp->t_state))
6148                 rack_set_state(tp, rack);
6149         so = tp->t_inpcb->inp_socket;
6150         avail = sbavail(&so->so_snd);
6151         out = tp->snd_max - tp->snd_una;
6152         if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
6153                 /* special case, we need a retransmission */
6154                 collapsed_win = 1;
6155                 goto need_retran;
6156         }
6157         if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6158                 rack->r_ctl.dsack_persist--;
6159                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6160                         rack->r_ctl.num_dsack = 0;
6161                 }
6162                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6163         }
6164         if ((tp->t_flags & TF_GPUTINPROG) &&
6165             (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6166                 /*
6167                  * If this is the second in a row
6168                  * TLP and we are doing a measurement
6169                  * its time to abandon the measurement.
6170                  * Something is likely broken on
6171                  * the clients network and measuring a
6172                  * broken network does us no good.
6173                  */
6174                 tp->t_flags &= ~TF_GPUTINPROG;
6175                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6176                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6177                                            tp->gput_seq,
6178                                            0, 0, 18, __LINE__, NULL, 0);
6179         }
6180         /*
6181          * Check our send oldest always settings, and if
6182          * there is an oldest to send jump to the need_retran.
6183          */
6184         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6185                 goto need_retran;
6186
6187         if (avail > out) {
6188                 /* New data is available */
6189                 amm = avail - out;
6190                 if (amm > ctf_fixed_maxseg(tp)) {
6191                         amm = ctf_fixed_maxseg(tp);
6192                         if ((amm + out) > tp->snd_wnd) {
6193                                 /* We are rwnd limited */
6194                                 goto need_retran;
6195                         }
6196                 } else if (amm < ctf_fixed_maxseg(tp)) {
6197                         /* not enough to fill a MTU */
6198                         goto need_retran;
6199                 }
6200                 if (IN_FASTRECOVERY(tp->t_flags)) {
6201                         /* Unlikely */
6202                         if (rack->rack_no_prr == 0) {
6203                                 if (out + amm <= tp->snd_wnd) {
6204                                         rack->r_ctl.rc_prr_sndcnt = amm;
6205                                         rack->r_ctl.rc_tlp_new_data = amm;
6206                                         rack_log_to_prr(rack, 4, 0, __LINE__);
6207                                 }
6208                         } else
6209                                 goto need_retran;
6210                 } else {
6211                         /* Set the send-new override */
6212                         if (out + amm <= tp->snd_wnd)
6213                                 rack->r_ctl.rc_tlp_new_data = amm;
6214                         else
6215                                 goto need_retran;
6216                 }
6217                 rack->r_ctl.rc_tlpsend = NULL;
6218                 counter_u64_add(rack_tlp_newdata, 1);
6219                 goto send;
6220         }
6221 need_retran:
6222         /*
6223          * Ok we need to arrange the last un-acked segment to be re-sent, or
6224          * optionally the first un-acked segment.
6225          */
6226         if (collapsed_win == 0) {
6227                 if (rack_always_send_oldest)
6228                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6229                 else {
6230                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6231                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6232                                 rsm = rack_find_high_nonack(rack, rsm);
6233                         }
6234                 }
6235                 if (rsm == NULL) {
6236 #ifdef TCP_BLACKBOX
6237                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6238 #endif
6239                         goto out;
6240                 }
6241         } else {
6242                 /*
6243                  * We must find the last segment
6244                  * that was acceptable by the client.
6245                  */
6246                 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6247                         if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6248                                 /* Found one */
6249                                 break;
6250                         }
6251                 }
6252                 if (rsm == NULL) {
6253                         /* None? if so send the first */
6254                         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6255                         if (rsm == NULL) {
6256 #ifdef TCP_BLACKBOX
6257                                 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6258 #endif
6259                                 goto out;
6260                         }
6261                 }
6262         }
6263         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6264                 /*
6265                  * We need to split this the last segment in two.
6266                  */
6267                 struct rack_sendmap *nrsm;
6268
6269                 nrsm = rack_alloc_full_limit(rack);
6270                 if (nrsm == NULL) {
6271                         /*
6272                          * No memory to split, we will just exit and punt
6273                          * off to the RXT timer.
6274                          */
6275                         goto out;
6276                 }
6277                 rack_clone_rsm(rack, nrsm, rsm,
6278                                (rsm->r_end - ctf_fixed_maxseg(tp)));
6279                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6280 #ifndef INVARIANTS
6281                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6282 #else
6283                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6284                 if (insret != NULL) {
6285                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6286                               nrsm, insret, rack, rsm);
6287                 }
6288 #endif
6289                 if (rsm->r_in_tmap) {
6290                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6291                         nrsm->r_in_tmap = 1;
6292                 }
6293                 rsm = nrsm;
6294         }
6295         rack->r_ctl.rc_tlpsend = rsm;
6296 send:
6297         /* Make sure output path knows we are doing a TLP */
6298         *doing_tlp = 1;
6299         rack->r_timer_override = 1;
6300         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6301         return (0);
6302 out:
6303         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6304         return (0);
6305 }
6306
6307 /*
6308  * Delayed ack Timer, here we simply need to setup the
6309  * ACK_NOW flag and remove the DELACK flag. From there
6310  * the output routine will send the ack out.
6311  *
6312  * We only return 1, saying don't proceed, if all timers
6313  * are stopped (destroyed PCB?).
6314  */
6315 static int
6316 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6317 {
6318         if (tp->t_timers->tt_flags & TT_STOPPED) {
6319                 return (1);
6320         }
6321         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6322         tp->t_flags &= ~TF_DELACK;
6323         tp->t_flags |= TF_ACKNOW;
6324         KMOD_TCPSTAT_INC(tcps_delack);
6325         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6326         return (0);
6327 }
6328
6329 /*
6330  * Persists timer, here we simply send the
6331  * same thing as a keepalive will.
6332  * the one byte send.
6333  *
6334  * We only return 1, saying don't proceed, if all timers
6335  * are stopped (destroyed PCB?).
6336  */
6337 static int
6338 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6339 {
6340         struct tcptemp *t_template;
6341 #ifdef INVARIANTS
6342         struct inpcb *inp = tp->t_inpcb;
6343 #endif
6344         int32_t retval = 1;
6345
6346         if (tp->t_timers->tt_flags & TT_STOPPED) {
6347                 return (1);
6348         }
6349         if (rack->rc_in_persist == 0)
6350                 return (0);
6351         if (ctf_progress_timeout_check(tp, false)) {
6352                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6353                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6354                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6355                 return (-ETIMEDOUT);    /* tcp_drop() */
6356         }
6357         KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6358         /*
6359          * Persistence timer into zero window. Force a byte to be output, if
6360          * possible.
6361          */
6362         KMOD_TCPSTAT_INC(tcps_persisttimeo);
6363         /*
6364          * Hack: if the peer is dead/unreachable, we do not time out if the
6365          * window is closed.  After a full backoff, drop the connection if
6366          * the idle time (no responses to probes) reaches the maximum
6367          * backoff that we would use if retransmitting.
6368          */
6369         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6370             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6371              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6372                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6373                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6374                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6375                 retval = -ETIMEDOUT;    /* tcp_drop() */
6376                 goto out;
6377         }
6378         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6379             tp->snd_una == tp->snd_max)
6380                 rack_exit_persist(tp, rack, cts);
6381         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6382         /*
6383          * If the user has closed the socket then drop a persisting
6384          * connection after a much reduced timeout.
6385          */
6386         if (tp->t_state > TCPS_CLOSE_WAIT &&
6387             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6388                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6389                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6390                 counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6391                 retval = -ETIMEDOUT;    /* tcp_drop() */
6392                 goto out;
6393         }
6394         t_template = tcpip_maketemplate(rack->rc_inp);
6395         if (t_template) {
6396                 /* only set it if we were answered */
6397                 if (rack->forced_ack == 0) {
6398                         rack->forced_ack = 1;
6399                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6400                 } else {
6401                         rack->probe_not_answered = 1;
6402                         counter_u64_add(rack_persists_loss, 1);
6403                         rack->r_ctl.persist_lost_ends++;
6404                 }
6405                 counter_u64_add(rack_persists_sends, 1);
6406                 tcp_respond(tp, t_template->tt_ipgen,
6407                             &t_template->tt_t, (struct mbuf *)NULL,
6408                             tp->rcv_nxt, tp->snd_una - 1, 0);
6409                 /* This sends an ack */
6410                 if (tp->t_flags & TF_DELACK)
6411                         tp->t_flags &= ~TF_DELACK;
6412                 free(t_template, M_TEMP);
6413         }
6414         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6415                 tp->t_rxtshift++;
6416 out:
6417         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6418         rack_start_hpts_timer(rack, tp, cts,
6419                               0, 0, 0);
6420         return (retval);
6421 }
6422
6423 /*
6424  * If a keepalive goes off, we had no other timers
6425  * happening. We always return 1 here since this
6426  * routine either drops the connection or sends
6427  * out a segment with respond.
6428  */
6429 static int
6430 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6431 {
6432         struct tcptemp *t_template;
6433         struct inpcb *inp;
6434
6435         if (tp->t_timers->tt_flags & TT_STOPPED) {
6436                 return (1);
6437         }
6438         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6439         inp = tp->t_inpcb;
6440         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6441         /*
6442          * Keep-alive timer went off; send something or drop connection if
6443          * idle for too long.
6444          */
6445         KMOD_TCPSTAT_INC(tcps_keeptimeo);
6446         if (tp->t_state < TCPS_ESTABLISHED)
6447                 goto dropit;
6448         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6449             tp->t_state <= TCPS_CLOSING) {
6450                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6451                         goto dropit;
6452                 /*
6453                  * Send a packet designed to force a response if the peer is
6454                  * up and reachable: either an ACK if the connection is
6455                  * still alive, or an RST if the peer has closed the
6456                  * connection due to timeout or reboot. Using sequence
6457                  * number tp->snd_una-1 causes the transmitted zero-length
6458                  * segment to lie outside the receive window; by the
6459                  * protocol spec, this requires the correspondent TCP to
6460                  * respond.
6461                  */
6462                 KMOD_TCPSTAT_INC(tcps_keepprobe);
6463                 t_template = tcpip_maketemplate(inp);
6464                 if (t_template) {
6465                         if (rack->forced_ack == 0) {
6466                                 rack->forced_ack = 1;
6467                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6468                         } else {
6469                                 rack->probe_not_answered = 1;
6470                         }
6471                         tcp_respond(tp, t_template->tt_ipgen,
6472                             &t_template->tt_t, (struct mbuf *)NULL,
6473                             tp->rcv_nxt, tp->snd_una - 1, 0);
6474                         free(t_template, M_TEMP);
6475                 }
6476         }
6477         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6478         return (1);
6479 dropit:
6480         KMOD_TCPSTAT_INC(tcps_keepdrops);
6481         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6482         return (-ETIMEDOUT);    /* tcp_drop() */
6483 }
6484
6485 /*
6486  * Retransmit helper function, clear up all the ack
6487  * flags and take care of important book keeping.
6488  */
6489 static void
6490 rack_remxt_tmr(struct tcpcb *tp)
6491 {
6492         /*
6493          * The retransmit timer went off, all sack'd blocks must be
6494          * un-acked.
6495          */
6496         struct rack_sendmap *rsm, *trsm = NULL;
6497         struct tcp_rack *rack;
6498
6499         rack = (struct tcp_rack *)tp->t_fb_ptr;
6500         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6501         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6502         if (rack->r_state && (rack->r_state != tp->t_state))
6503                 rack_set_state(tp, rack);
6504         /*
6505          * Ideally we would like to be able to
6506          * mark SACK-PASS on anything not acked here.
6507          *
6508          * However, if we do that we would burst out
6509          * all that data 1ms apart. This would be unwise,
6510          * so for now we will just let the normal rxt timer
6511          * and tlp timer take care of it.
6512          *
6513          * Also we really need to stick them back in sequence
6514          * order. This way we send in the proper order and any
6515          * sacks that come floating in will "re-ack" the data.
6516          * To do this we zap the tmap with an INIT and then
6517          * walk through and place every rsm in the RB tree
6518          * back in its seq ordered place.
6519          */
6520         TAILQ_INIT(&rack->r_ctl.rc_tmap);
6521         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6522                 rsm->r_dupack = 0;
6523                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6524                 /* We must re-add it back to the tlist */
6525                 if (trsm == NULL) {
6526                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6527                 } else {
6528                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6529                 }
6530                 rsm->r_in_tmap = 1;
6531                 trsm = rsm;
6532                 if (rsm->r_flags & RACK_ACKED)
6533                         rsm->r_flags |= RACK_WAS_ACKED;
6534                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
6535                 rsm->r_flags |= RACK_MUST_RXT;
6536         }
6537         /* Clear the count (we just un-acked them) */
6538         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6539         rack->r_ctl.rc_sacked = 0;
6540         rack->r_ctl.rc_sacklast = NULL;
6541         rack->r_ctl.rc_agg_delayed = 0;
6542         rack->r_early = 0;
6543         rack->r_ctl.rc_agg_early = 0;
6544         rack->r_late = 0;
6545         /* Clear the tlp rtx mark */
6546         rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6547         if (rack->r_ctl.rc_resend != NULL)
6548                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6549         rack->r_ctl.rc_prr_sndcnt = 0;
6550         rack_log_to_prr(rack, 6, 0, __LINE__);
6551         rack->r_timer_override = 1;
6552         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6553 #ifdef NETFLIX_EXP_DETECTION
6554             || (rack->sack_attack_disable != 0)
6555 #endif
6556                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6557                 /*
6558                  * For non-sack customers new data
6559                  * needs to go out as retransmits until
6560                  * we retransmit up to snd_max.
6561                  */
6562                 rack->r_must_retran = 1;
6563                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6564                                                 rack->r_ctl.rc_sacked);
6565         }
6566         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6567 }
6568
6569 static void
6570 rack_convert_rtts(struct tcpcb *tp)
6571 {
6572         if (tp->t_srtt > 1) {
6573                 uint32_t val, frac;
6574
6575                 val = tp->t_srtt >> TCP_RTT_SHIFT;
6576                 frac = tp->t_srtt & 0x1f;
6577                 tp->t_srtt = TICKS_2_USEC(val);
6578                 /*
6579                  * frac is the fractional part of the srtt (if any)
6580                  * but its in ticks and every bit represents
6581                  * 1/32nd of a hz.
6582                  */
6583                 if (frac) {
6584                         if (hz == 1000) {
6585                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6586                         } else {
6587                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6588                         }
6589                         tp->t_srtt += frac;
6590                 }
6591         }
6592         if (tp->t_rttvar) {
6593                 uint32_t val, frac;
6594
6595                 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6596                 frac = tp->t_rttvar & 0x1f;
6597                 tp->t_rttvar = TICKS_2_USEC(val);
6598                 /*
6599                  * frac is the fractional part of the srtt (if any)
6600                  * but its in ticks and every bit represents
6601                  * 1/32nd of a hz.
6602                  */
6603                 if (frac) {
6604                         if (hz == 1000) {
6605                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6606                         } else {
6607                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6608                         }
6609                         tp->t_rttvar += frac;
6610                 }
6611         }
6612         tp->t_rxtcur = RACK_REXMTVAL(tp);
6613         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6614                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6615         }
6616         if (tp->t_rxtcur > rack_rto_max) {
6617                 tp->t_rxtcur = rack_rto_max;
6618         }
6619 }
6620
6621 static void
6622 rack_cc_conn_init(struct tcpcb *tp)
6623 {
6624         struct tcp_rack *rack;
6625         uint32_t srtt;
6626
6627         rack = (struct tcp_rack *)tp->t_fb_ptr;
6628         srtt = tp->t_srtt;
6629         cc_conn_init(tp);
6630         /*
6631          * Now convert to rack's internal format,
6632          * if required.
6633          */
6634         if ((srtt == 0) && (tp->t_srtt != 0))
6635                 rack_convert_rtts(tp);
6636         /*
6637          * We want a chance to stay in slowstart as
6638          * we create a connection. TCP spec says that
6639          * initially ssthresh is infinite. For our
6640          * purposes that is the snd_wnd.
6641          */
6642         if (tp->snd_ssthresh < tp->snd_wnd) {
6643                 tp->snd_ssthresh = tp->snd_wnd;
6644         }
6645         /*
6646          * We also want to assure a IW worth of
6647          * data can get inflight.
6648          */
6649         if (rc_init_window(rack) < tp->snd_cwnd)
6650                 tp->snd_cwnd = rc_init_window(rack);
6651 }
6652
6653 /*
6654  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6655  * we will setup to retransmit the lowest seq number outstanding.
6656  */
6657 static int
6658 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6659 {
6660         int32_t rexmt;
6661         int32_t retval = 0;
6662         bool isipv6;
6663
6664         if (tp->t_timers->tt_flags & TT_STOPPED) {
6665                 return (1);
6666         }
6667         if ((tp->t_flags & TF_GPUTINPROG) &&
6668             (tp->t_rxtshift)) {
6669                 /*
6670                  * We have had a second timeout
6671                  * measurements on successive rxt's are not profitable.
6672                  * It is unlikely to be of any use (the network is
6673                  * broken or the client went away).
6674                  */
6675                 tp->t_flags &= ~TF_GPUTINPROG;
6676                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6677                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6678                                            tp->gput_seq,
6679                                            0, 0, 18, __LINE__, NULL, 0);
6680         }
6681         if (ctf_progress_timeout_check(tp, false)) {
6682                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6683                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6684                 return (-ETIMEDOUT);    /* tcp_drop() */
6685         }
6686         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6687         rack->r_ctl.retran_during_recovery = 0;
6688         rack->rc_ack_required = 1;
6689         rack->r_ctl.dsack_byte_cnt = 0;
6690         if (IN_FASTRECOVERY(tp->t_flags))
6691                 tp->t_flags |= TF_WASFRECOVERY;
6692         else
6693                 tp->t_flags &= ~TF_WASFRECOVERY;
6694         if (IN_CONGRECOVERY(tp->t_flags))
6695                 tp->t_flags |= TF_WASCRECOVERY;
6696         else
6697                 tp->t_flags &= ~TF_WASCRECOVERY;
6698         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6699             (tp->snd_una == tp->snd_max)) {
6700                 /* Nothing outstanding .. nothing to do */
6701                 return (0);
6702         }
6703         if (rack->r_ctl.dsack_persist) {
6704                 rack->r_ctl.dsack_persist--;
6705                 if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6706                         rack->r_ctl.num_dsack = 0;
6707                 }
6708                 rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6709         }
6710         /*
6711          * Rack can only run one timer  at a time, so we cannot
6712          * run a KEEPINIT (gating SYN sending) and a retransmit
6713          * timer for the SYN. So if we are in a front state and
6714          * have a KEEPINIT timer we need to check the first transmit
6715          * against now to see if we have exceeded the KEEPINIT time
6716          * (if one is set).
6717          */
6718         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6719             (TP_KEEPINIT(tp) != 0)) {
6720                 struct rack_sendmap *rsm;
6721
6722                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6723                 if (rsm) {
6724                         /* Ok we have something outstanding to test keepinit with */
6725                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6726                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6727                                 /* We have exceeded the KEEPINIT time */
6728                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6729                                 goto drop_it;
6730                         }
6731                 }
6732         }
6733         /*
6734          * Retransmission timer went off.  Message has not been acked within
6735          * retransmit interval.  Back off to a longer retransmit interval
6736          * and retransmit one segment.
6737          */
6738         rack_remxt_tmr(tp);
6739         if ((rack->r_ctl.rc_resend == NULL) ||
6740             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6741                 /*
6742                  * If the rwnd collapsed on
6743                  * the one we are retransmitting
6744                  * it does not count against the
6745                  * rxt count.
6746                  */
6747                 tp->t_rxtshift++;
6748         }
6749         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6750                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6751 drop_it:
6752                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6753                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6754                 /* XXXGL: previously t_softerror was casted to uint16_t */
6755                 MPASS(tp->t_softerror >= 0);
6756                 retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6757                 goto out;       /* tcp_drop() */
6758         }
6759         if (tp->t_state == TCPS_SYN_SENT) {
6760                 /*
6761                  * If the SYN was retransmitted, indicate CWND to be limited
6762                  * to 1 segment in cc_conn_init().
6763                  */
6764                 tp->snd_cwnd = 1;
6765         } else if (tp->t_rxtshift == 1) {
6766                 /*
6767                  * first retransmit; record ssthresh and cwnd so they can be
6768                  * recovered if this turns out to be a "bad" retransmit. A
6769                  * retransmit is considered "bad" if an ACK for this segment
6770                  * is received within RTT/2 interval; the assumption here is
6771                  * that the ACK was already in flight.  See "On Estimating
6772                  * End-to-End Network Path Properties" by Allman and Paxson
6773                  * for more details.
6774                  */
6775                 tp->snd_cwnd_prev = tp->snd_cwnd;
6776                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6777                 tp->snd_recover_prev = tp->snd_recover;
6778                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6779                 tp->t_flags |= TF_PREVVALID;
6780         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6781                 tp->t_flags &= ~TF_PREVVALID;
6782         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6783         if ((tp->t_state == TCPS_SYN_SENT) ||
6784             (tp->t_state == TCPS_SYN_RECEIVED))
6785                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6786         else
6787                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6788
6789         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6790            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6791         /*
6792          * We enter the path for PLMTUD if connection is established or, if
6793          * connection is FIN_WAIT_1 status, reason for the last is that if
6794          * amount of data we send is very small, we could send it in couple
6795          * of packets and process straight to FIN. In that case we won't
6796          * catch ESTABLISHED state.
6797          */
6798 #ifdef INET6
6799         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6800 #else
6801         isipv6 = false;
6802 #endif
6803         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6804             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6805             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6806             ((tp->t_state == TCPS_ESTABLISHED) ||
6807             (tp->t_state == TCPS_FIN_WAIT_1))) {
6808                 /*
6809                  * Idea here is that at each stage of mtu probe (usually,
6810                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6811                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6812                  * should take care of that.
6813                  */
6814                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6815                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6816                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6817                     tp->t_rxtshift % 2 == 0)) {
6818                         /*
6819                          * Enter Path MTU Black-hole Detection mechanism: -
6820                          * Disable Path MTU Discovery (IP "DF" bit). -
6821                          * Reduce MTU to lower value than what we negotiated
6822                          * with peer.
6823                          */
6824                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6825                                 /* Record that we may have found a black hole. */
6826                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6827                                 /* Keep track of previous MSS. */
6828                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6829                         }
6830
6831                         /*
6832                          * Reduce the MSS to blackhole value or to the
6833                          * default in an attempt to retransmit.
6834                          */
6835 #ifdef INET6
6836                         if (isipv6 &&
6837                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6838                                 /* Use the sysctl tuneable blackhole MSS. */
6839                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6840                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6841                         } else if (isipv6) {
6842                                 /* Use the default MSS. */
6843                                 tp->t_maxseg = V_tcp_v6mssdflt;
6844                                 /*
6845                                  * Disable Path MTU Discovery when we switch
6846                                  * to minmss.
6847                                  */
6848                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6849                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6850                         }
6851 #endif
6852 #if defined(INET6) && defined(INET)
6853                         else
6854 #endif
6855 #ifdef INET
6856                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6857                                 /* Use the sysctl tuneable blackhole MSS. */
6858                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6859                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6860                         } else {
6861                                 /* Use the default MSS. */
6862                                 tp->t_maxseg = V_tcp_mssdflt;
6863                                 /*
6864                                  * Disable Path MTU Discovery when we switch
6865                                  * to minmss.
6866                                  */
6867                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6868                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6869                         }
6870 #endif
6871                 } else {
6872                         /*
6873                          * If further retransmissions are still unsuccessful
6874                          * with a lowered MTU, maybe this isn't a blackhole
6875                          * and we restore the previous MSS and blackhole
6876                          * detection flags. The limit '6' is determined by
6877                          * giving each probe stage (1448, 1188, 524) 2
6878                          * chances to recover.
6879                          */
6880                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6881                             (tp->t_rxtshift >= 6)) {
6882                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6883                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6884                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6885                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6886                         }
6887                 }
6888         }
6889         /*
6890          * Disable RFC1323 and SACK if we haven't got any response to
6891          * our third SYN to work-around some broken terminal servers
6892          * (most of which have hopefully been retired) that have bad VJ
6893          * header compression code which trashes TCP segments containing
6894          * unknown-to-them TCP options.
6895          */
6896         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6897             (tp->t_rxtshift == 3))
6898                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6899         /*
6900          * If we backed off this far, our srtt estimate is probably bogus.
6901          * Clobber it so we'll take the next rtt measurement as our srtt;
6902          * move the current srtt into rttvar to keep the current retransmit
6903          * times until then.
6904          */
6905         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6906 #ifdef INET6
6907                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6908                         in6_losing(tp->t_inpcb);
6909                 else
6910 #endif
6911                         in_losing(tp->t_inpcb);
6912                 tp->t_rttvar += tp->t_srtt;
6913                 tp->t_srtt = 0;
6914         }
6915         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6916         tp->snd_recover = tp->snd_max;
6917         tp->t_flags |= TF_ACKNOW;
6918         tp->t_rtttime = 0;
6919         rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
6920 out:
6921         return (retval);
6922 }
6923
6924 static int
6925 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6926 {
6927         int32_t ret = 0;
6928         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6929
6930         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6931             (tp->t_flags & TF_GPUTINPROG)) {
6932                 /*
6933                  * We have a goodput in progress
6934                  * and we have entered a late state.
6935                  * Do we have enough data in the sb
6936                  * to handle the GPUT request?
6937                  */
6938                 uint32_t bytes;
6939
6940                 bytes = tp->gput_ack - tp->gput_seq;
6941                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
6942                         bytes += tp->gput_seq - tp->snd_una;
6943                 if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
6944                         /*
6945                          * There are not enough bytes in the socket
6946                          * buffer that have been sent to cover this
6947                          * measurement. Cancel it.
6948                          */
6949                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6950                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
6951                                                    tp->gput_seq,
6952                                                    0, 0, 18, __LINE__, NULL, 0);
6953                         tp->t_flags &= ~TF_GPUTINPROG;
6954                 }
6955         }
6956         if (timers == 0) {
6957                 return (0);
6958         }
6959         if (tp->t_state == TCPS_LISTEN) {
6960                 /* no timers on listen sockets */
6961                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6962                         return (0);
6963                 return (1);
6964         }
6965         if ((timers & PACE_TMR_RACK) &&
6966             rack->rc_on_min_to) {
6967                 /*
6968                  * For the rack timer when we
6969                  * are on a min-timeout (which means rrr_conf = 3)
6970                  * we don't want to check the timer. It may
6971                  * be going off for a pace and thats ok we
6972                  * want to send the retransmit (if its ready).
6973                  *
6974                  * If its on a normal rack timer (non-min) then
6975                  * we will check if its expired.
6976                  */
6977                 goto skip_time_check;
6978         }
6979         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6980                 uint32_t left;
6981
6982                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6983                         ret = -1;
6984                         rack_log_to_processing(rack, cts, ret, 0);
6985                         return (0);
6986                 }
6987                 if (hpts_calling == 0) {
6988                         /*
6989                          * A user send or queued mbuf (sack) has called us? We
6990                          * return 0 and let the pacing guards
6991                          * deal with it if they should or
6992                          * should not cause a send.
6993                          */
6994                         ret = -2;
6995                         rack_log_to_processing(rack, cts, ret, 0);
6996                         return (0);
6997                 }
6998                 /*
6999                  * Ok our timer went off early and we are not paced false
7000                  * alarm, go back to sleep.
7001                  */
7002                 ret = -3;
7003                 left = rack->r_ctl.rc_timer_exp - cts;
7004                 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
7005                 rack_log_to_processing(rack, cts, ret, left);
7006                 return (1);
7007         }
7008 skip_time_check:
7009         rack->rc_tmr_stopped = 0;
7010         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
7011         if (timers & PACE_TMR_DELACK) {
7012                 ret = rack_timeout_delack(tp, rack, cts);
7013         } else if (timers & PACE_TMR_RACK) {
7014                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
7015                 rack->r_fast_output = 0;
7016                 ret = rack_timeout_rack(tp, rack, cts);
7017         } else if (timers & PACE_TMR_TLP) {
7018                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
7019                 ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
7020         } else if (timers & PACE_TMR_RXT) {
7021                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
7022                 rack->r_fast_output = 0;
7023                 ret = rack_timeout_rxt(tp, rack, cts);
7024         } else if (timers & PACE_TMR_PERSIT) {
7025                 ret = rack_timeout_persist(tp, rack, cts);
7026         } else if (timers & PACE_TMR_KEEP) {
7027                 ret = rack_timeout_keepalive(tp, rack, cts);
7028         }
7029         rack_log_to_processing(rack, cts, ret, timers);
7030         return (ret);
7031 }
7032
7033 static void
7034 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
7035 {
7036         struct timeval tv;
7037         uint32_t us_cts, flags_on_entry;
7038         uint8_t hpts_removed = 0;
7039
7040         flags_on_entry = rack->r_ctl.rc_hpts_flags;
7041         us_cts = tcp_get_usecs(&tv);
7042         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7043             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7044              ((tp->snd_max - tp->snd_una) == 0))) {
7045                 tcp_hpts_remove(rack->rc_inp);
7046                 hpts_removed = 1;
7047                 /* If we were not delayed cancel out the flag. */
7048                 if ((tp->snd_max - tp->snd_una) == 0)
7049                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7050                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7051         }
7052         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7053                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7054                 if (tcp_in_hpts(rack->rc_inp) &&
7055                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7056                         /*
7057                          * Canceling timer's when we have no output being
7058                          * paced. We also must remove ourselves from the
7059                          * hpts.
7060                          */
7061                         tcp_hpts_remove(rack->rc_inp);
7062                         hpts_removed = 1;
7063                 }
7064                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7065         }
7066         if (hpts_removed == 0)
7067                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7068 }
7069
7070 static void
7071 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7072 {
7073         return;
7074 }
7075
7076 static int
7077 rack_stopall(struct tcpcb *tp)
7078 {
7079         struct tcp_rack *rack;
7080         rack = (struct tcp_rack *)tp->t_fb_ptr;
7081         rack->t_timers_stopped = 1;
7082         return (0);
7083 }
7084
7085 static void
7086 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7087 {
7088         return;
7089 }
7090
7091 static int
7092 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7093 {
7094         return (0);
7095 }
7096
7097 static void
7098 rack_stop_all_timers(struct tcpcb *tp)
7099 {
7100         struct tcp_rack *rack;
7101
7102         /*
7103          * Assure no timers are running.
7104          */
7105         if (tcp_timer_active(tp, TT_PERSIST)) {
7106                 /* We enter in persists, set the flag appropriately */
7107                 rack = (struct tcp_rack *)tp->t_fb_ptr;
7108                 rack->rc_in_persist = 1;
7109         }
7110         tcp_timer_suspend(tp, TT_PERSIST);
7111         tcp_timer_suspend(tp, TT_REXMT);
7112         tcp_timer_suspend(tp, TT_KEEP);
7113         tcp_timer_suspend(tp, TT_DELACK);
7114 }
7115
7116 static void
7117 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7118     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7119 {
7120         int32_t idx;
7121
7122         rsm->r_rtr_cnt++;
7123         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7124         rsm->r_dupack = 0;
7125         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7126                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7127                 rsm->r_flags |= RACK_OVERMAX;
7128         }
7129         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7130                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7131                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7132         }
7133         idx = rsm->r_rtr_cnt - 1;
7134         rsm->r_tim_lastsent[idx] = ts;
7135         /*
7136          * Here we don't add in the len of send, since its already
7137          * in snduna <->snd_max.
7138          */
7139         rsm->r_fas = ctf_flight_size(rack->rc_tp,
7140                                      rack->r_ctl.rc_sacked);
7141         if (rsm->r_flags & RACK_ACKED) {
7142                 /* Problably MTU discovery messing with us */
7143                 rsm->r_flags &= ~RACK_ACKED;
7144                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7145         }
7146         if (rsm->r_in_tmap) {
7147                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7148                 rsm->r_in_tmap = 0;
7149         }
7150         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7151         rsm->r_in_tmap = 1;
7152         /* Take off the must retransmit flag, if its on */
7153         if (rsm->r_flags & RACK_MUST_RXT) {
7154                 if (rack->r_must_retran)
7155                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
7156                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
7157                         /*
7158                          * We have retransmitted all we need. Clear
7159                          * any must retransmit flags.
7160                          */
7161                         rack->r_must_retran = 0;
7162                         rack->r_ctl.rc_out_at_rto = 0;
7163                 }
7164                 rsm->r_flags &= ~RACK_MUST_RXT;
7165         }
7166         if (rsm->r_flags & RACK_SACK_PASSED) {
7167                 /* We have retransmitted due to the SACK pass */
7168                 rsm->r_flags &= ~RACK_SACK_PASSED;
7169                 rsm->r_flags |= RACK_WAS_SACKPASS;
7170         }
7171 }
7172
7173 static uint32_t
7174 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7175     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7176 {
7177         /*
7178          * We (re-)transmitted starting at rsm->r_start for some length
7179          * (possibly less than r_end.
7180          */
7181         struct rack_sendmap *nrsm;
7182 #ifdef INVARIANTS
7183         struct rack_sendmap *insret;
7184 #endif
7185         uint32_t c_end;
7186         int32_t len;
7187
7188         len = *lenp;
7189         c_end = rsm->r_start + len;
7190         if (SEQ_GEQ(c_end, rsm->r_end)) {
7191                 /*
7192                  * We retransmitted the whole piece or more than the whole
7193                  * slopping into the next rsm.
7194                  */
7195                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7196                 if (c_end == rsm->r_end) {
7197                         *lenp = 0;
7198                         return (0);
7199                 } else {
7200                         int32_t act_len;
7201
7202                         /* Hangs over the end return whats left */
7203                         act_len = rsm->r_end - rsm->r_start;
7204                         *lenp = (len - act_len);
7205                         return (rsm->r_end);
7206                 }
7207                 /* We don't get out of this block. */
7208         }
7209         /*
7210          * Here we retransmitted less than the whole thing which means we
7211          * have to split this into what was transmitted and what was not.
7212          */
7213         nrsm = rack_alloc_full_limit(rack);
7214         if (nrsm == NULL) {
7215                 /*
7216                  * We can't get memory, so lets not proceed.
7217                  */
7218                 *lenp = 0;
7219                 return (0);
7220         }
7221         /*
7222          * So here we are going to take the original rsm and make it what we
7223          * retransmitted. nrsm will be the tail portion we did not
7224          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7225          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7226          * 1, 6 and the new piece will be 6, 11.
7227          */
7228         rack_clone_rsm(rack, nrsm, rsm, c_end);
7229         nrsm->r_dupack = 0;
7230         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7231 #ifndef INVARIANTS
7232         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7233 #else
7234         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7235         if (insret != NULL) {
7236                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7237                       nrsm, insret, rack, rsm);
7238         }
7239 #endif
7240         if (rsm->r_in_tmap) {
7241                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7242                 nrsm->r_in_tmap = 1;
7243         }
7244         rsm->r_flags &= (~RACK_HAS_FIN);
7245         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7246         /* Log a split of rsm into rsm and nrsm */
7247         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7248         *lenp = 0;
7249         return (0);
7250 }
7251
7252 static void
7253 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7254                 uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
7255                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7256 {
7257         struct tcp_rack *rack;
7258         struct rack_sendmap *rsm, *nrsm, fe;
7259 #ifdef INVARIANTS
7260         struct rack_sendmap *insret;
7261 #endif
7262         register uint32_t snd_max, snd_una;
7263
7264         /*
7265          * Add to the RACK log of packets in flight or retransmitted. If
7266          * there is a TS option we will use the TS echoed, if not we will
7267          * grab a TS.
7268          *
7269          * Retransmissions will increment the count and move the ts to its
7270          * proper place. Note that if options do not include TS's then we
7271          * won't be able to effectively use the ACK for an RTT on a retran.
7272          *
7273          * Notes about r_start and r_end. Lets consider a send starting at
7274          * sequence 1 for 10 bytes. In such an example the r_start would be
7275          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7276          * This means that r_end is actually the first sequence for the next
7277          * slot (11).
7278          *
7279          */
7280         /*
7281          * If err is set what do we do XXXrrs? should we not add the thing?
7282          * -- i.e. return if err != 0 or should we pretend we sent it? --
7283          * i.e. proceed with add ** do this for now.
7284          */
7285         INP_WLOCK_ASSERT(tp->t_inpcb);
7286         if (err)
7287                 /*
7288                  * We don't log errors -- we could but snd_max does not
7289                  * advance in this case either.
7290                  */
7291                 return;
7292
7293         if (th_flags & TH_RST) {
7294                 /*
7295                  * We don't log resets and we return immediately from
7296                  * sending
7297                  */
7298                 return;
7299         }
7300         rack = (struct tcp_rack *)tp->t_fb_ptr;
7301         snd_una = tp->snd_una;
7302         snd_max = tp->snd_max;
7303         if (th_flags & (TH_SYN | TH_FIN)) {
7304                 /*
7305                  * The call to rack_log_output is made before bumping
7306                  * snd_max. This means we can record one extra byte on a SYN
7307                  * or FIN if seq_out is adding more on and a FIN is present
7308                  * (and we are not resending).
7309                  */
7310                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7311                         len++;
7312                 if (th_flags & TH_FIN)
7313                         len++;
7314                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7315                         /*
7316                          * The add/update as not been done for the FIN/SYN
7317                          * yet.
7318                          */
7319                         snd_max = tp->snd_nxt;
7320                 }
7321         }
7322         if (SEQ_LEQ((seq_out + len), snd_una)) {
7323                 /* Are sending an old segment to induce an ack (keep-alive)? */
7324                 return;
7325         }
7326         if (SEQ_LT(seq_out, snd_una)) {
7327                 /* huh? should we panic? */
7328                 uint32_t end;
7329
7330                 end = seq_out + len;
7331                 seq_out = snd_una;
7332                 if (SEQ_GEQ(end, seq_out))
7333                         len = end - seq_out;
7334                 else
7335                         len = 0;
7336         }
7337         if (len == 0) {
7338                 /* We don't log zero window probes */
7339                 return;
7340         }
7341         if (IN_FASTRECOVERY(tp->t_flags)) {
7342                 rack->r_ctl.rc_prr_out += len;
7343         }
7344         /* First question is it a retransmission or new? */
7345         if (seq_out == snd_max) {
7346                 /* Its new */
7347 again:
7348                 rsm = rack_alloc(rack);
7349                 if (rsm == NULL) {
7350                         /*
7351                          * Hmm out of memory and the tcb got destroyed while
7352                          * we tried to wait.
7353                          */
7354                         return;
7355                 }
7356                 if (th_flags & TH_FIN) {
7357                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7358                 } else {
7359                         rsm->r_flags = add_flag;
7360                 }
7361                 if (hw_tls)
7362                         rsm->r_hw_tls = 1;
7363                 rsm->r_tim_lastsent[0] = cts;
7364                 rsm->r_rtr_cnt = 1;
7365                 rsm->r_rtr_bytes = 0;
7366                 if (th_flags & TH_SYN) {
7367                         /* The data space is one beyond snd_una */
7368                         rsm->r_flags |= RACK_HAS_SYN;
7369                 }
7370                 rsm->r_start = seq_out;
7371                 rsm->r_end = rsm->r_start + len;
7372                 rsm->r_dupack = 0;
7373                 /*
7374                  * save off the mbuf location that
7375                  * sndmbuf_noadv returned (which is
7376                  * where we started copying from)..
7377                  */
7378                 rsm->m = s_mb;
7379                 rsm->soff = s_moff;
7380                 /*
7381                  * Here we do add in the len of send, since its not yet
7382                  * reflected in in snduna <->snd_max
7383                  */
7384                 rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7385                                               rack->r_ctl.rc_sacked) +
7386                               (rsm->r_end - rsm->r_start));
7387                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7388                 if (rsm->m) {
7389                         if (rsm->m->m_len <= rsm->soff) {
7390                                 /*
7391                                  * XXXrrs Question, will this happen?
7392                                  *
7393                                  * If sbsndptr is set at the correct place
7394                                  * then s_moff should always be somewhere
7395                                  * within rsm->m. But if the sbsndptr was
7396                                  * off then that won't be true. If it occurs
7397                                  * we need to walkout to the correct location.
7398                                  */
7399                                 struct mbuf *lm;
7400
7401                                 lm = rsm->m;
7402                                 while (lm->m_len <= rsm->soff) {
7403                                         rsm->soff -= lm->m_len;
7404                                         lm = lm->m_next;
7405                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7406                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7407                                 }
7408                                 rsm->m = lm;
7409                         }
7410                         rsm->orig_m_len = rsm->m->m_len;
7411                 } else
7412                         rsm->orig_m_len = 0;
7413                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7414                 /* Log a new rsm */
7415                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7416 #ifndef INVARIANTS
7417                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7418 #else
7419                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7420                 if (insret != NULL) {
7421                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7422                               nrsm, insret, rack, rsm);
7423                 }
7424 #endif
7425                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7426                 rsm->r_in_tmap = 1;
7427                 /*
7428                  * Special case detection, is there just a single
7429                  * packet outstanding when we are not in recovery?
7430                  *
7431                  * If this is true mark it so.
7432                  */
7433                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7434                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7435                         struct rack_sendmap *prsm;
7436
7437                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7438                         if (prsm)
7439                                 prsm->r_one_out_nr = 1;
7440                 }
7441                 return;
7442         }
7443         /*
7444          * If we reach here its a retransmission and we need to find it.
7445          */
7446         memset(&fe, 0, sizeof(fe));
7447 more:
7448         if (hintrsm && (hintrsm->r_start == seq_out)) {
7449                 rsm = hintrsm;
7450                 hintrsm = NULL;
7451         } else {
7452                 /* No hints sorry */
7453                 rsm = NULL;
7454         }
7455         if ((rsm) && (rsm->r_start == seq_out)) {
7456                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7457                 if (len == 0) {
7458                         return;
7459                 } else {
7460                         goto more;
7461                 }
7462         }
7463         /* Ok it was not the last pointer go through it the hard way. */
7464 refind:
7465         fe.r_start = seq_out;
7466         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7467         if (rsm) {
7468                 if (rsm->r_start == seq_out) {
7469                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7470                         if (len == 0) {
7471                                 return;
7472                         } else {
7473                                 goto refind;
7474                         }
7475                 }
7476                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7477                         /* Transmitted within this piece */
7478                         /*
7479                          * Ok we must split off the front and then let the
7480                          * update do the rest
7481                          */
7482                         nrsm = rack_alloc_full_limit(rack);
7483                         if (nrsm == NULL) {
7484                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7485                                 return;
7486                         }
7487                         /*
7488                          * copy rsm to nrsm and then trim the front of rsm
7489                          * to not include this part.
7490                          */
7491                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7492                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7493 #ifndef INVARIANTS
7494                         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7495 #else
7496                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7497                         if (insret != NULL) {
7498                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7499                                       nrsm, insret, rack, rsm);
7500                         }
7501 #endif
7502                         if (rsm->r_in_tmap) {
7503                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7504                                 nrsm->r_in_tmap = 1;
7505                         }
7506                         rsm->r_flags &= (~RACK_HAS_FIN);
7507                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7508                         if (len == 0) {
7509                                 return;
7510                         } else if (len > 0)
7511                                 goto refind;
7512                 }
7513         }
7514         /*
7515          * Hmm not found in map did they retransmit both old and on into the
7516          * new?
7517          */
7518         if (seq_out == tp->snd_max) {
7519                 goto again;
7520         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7521 #ifdef INVARIANTS
7522                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7523                        seq_out, len, tp->snd_una, tp->snd_max);
7524                 printf("Starting Dump of all rack entries\n");
7525                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7526                         printf("rsm:%p start:%u end:%u\n",
7527                                rsm, rsm->r_start, rsm->r_end);
7528                 }
7529                 printf("Dump complete\n");
7530                 panic("seq_out not found rack:%p tp:%p",
7531                       rack, tp);
7532 #endif
7533         } else {
7534 #ifdef INVARIANTS
7535                 /*
7536                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7537                  * flag)
7538                  */
7539                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7540                       seq_out, len, tp->snd_max, tp);
7541 #endif
7542         }
7543 }
7544
7545 /*
7546  * Record one of the RTT updates from an ack into
7547  * our sample structure.
7548  */
7549
7550 static void
7551 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7552                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7553 {
7554         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7555             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7556                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7557         }
7558         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7559             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7560                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7561         }
7562         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7563             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7564                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7565             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7566                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7567         }
7568         if ((confidence == 1) &&
7569             ((rsm == NULL) ||
7570              (rsm->r_just_ret) ||
7571              (rsm->r_one_out_nr &&
7572               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7573                 /*
7574                  * If the rsm had a just return
7575                  * hit it then we can't trust the
7576                  * rtt measurement for buffer deterimination
7577                  * Note that a confidence of 2, indicates
7578                  * SACK'd which overrides the r_just_ret or
7579                  * the r_one_out_nr. If it was a CUM-ACK and
7580                  * we had only two outstanding, but get an
7581                  * ack for only 1. Then that also lowers our
7582                  * confidence.
7583                  */
7584                 confidence = 0;
7585         }
7586         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7587             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7588                 if (rack->r_ctl.rack_rs.confidence == 0) {
7589                         /*
7590                          * We take anything with no current confidence
7591                          * saved.
7592                          */
7593                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7594                         rack->r_ctl.rack_rs.confidence = confidence;
7595                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7596                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7597                         /*
7598                          * Once we have a confident number,
7599                          * we can update it with a smaller
7600                          * value since this confident number
7601                          * may include the DSACK time until
7602                          * the next segment (the second one) arrived.
7603                          */
7604                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7605                         rack->r_ctl.rack_rs.confidence = confidence;
7606                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7607                 }
7608         }
7609         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7610         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7611         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7612         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7613 }
7614
7615 /*
7616  * Collect new round-trip time estimate
7617  * and update averages and current timeout.
7618  */
7619 static void
7620 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7621 {
7622         int32_t delta;
7623         int32_t rtt;
7624
7625         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7626                 /* No valid sample */
7627                 return;
7628         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7629                 /* We are to use the lowest RTT seen in a single ack */
7630                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7631         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7632                 /* We are to use the highest RTT seen in a single ack */
7633                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7634         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7635                 /* We are to use the average RTT seen in a single ack */
7636                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7637                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7638         } else {
7639 #ifdef INVARIANTS
7640                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7641 #endif
7642                 return;
7643         }
7644         if (rtt == 0)
7645                 rtt = 1;
7646         if (rack->rc_gp_rtt_set == 0) {
7647                 /*
7648                  * With no RTT we have to accept
7649                  * even one we are not confident of.
7650                  */
7651                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7652                 rack->rc_gp_rtt_set = 1;
7653         } else if (rack->r_ctl.rack_rs.confidence) {
7654                 /* update the running gp srtt */
7655                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7656                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7657         }
7658         if (rack->r_ctl.rack_rs.confidence) {
7659                 /*
7660                  * record the low and high for highly buffered path computation,
7661                  * we only do this if we are confident (not a retransmission).
7662                  */
7663                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7664                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7665                 }
7666                 if (rack->rc_highly_buffered == 0) {
7667                         /*
7668                          * Currently once we declare a path has
7669                          * highly buffered there is no going
7670                          * back, which may be a problem...
7671                          */
7672                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7673                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7674                                                      rack->r_ctl.rc_highest_us_rtt,
7675                                                      rack->r_ctl.rc_lowest_us_rtt,
7676                                                      RACK_RTTS_SEEHBP);
7677                                 rack->rc_highly_buffered = 1;
7678                         }
7679                 }
7680         }
7681         if ((rack->r_ctl.rack_rs.confidence) ||
7682             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7683                 /*
7684                  * If we are highly confident of it <or> it was
7685                  * never retransmitted we accept it as the last us_rtt.
7686                  */
7687                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7688                 /* The lowest rtt can be set if its was not retransmited */
7689                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7690                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7691                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7692                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7693                 }
7694         }
7695         rack = (struct tcp_rack *)tp->t_fb_ptr;
7696         if (tp->t_srtt != 0) {
7697                 /*
7698                  * We keep a simple srtt in microseconds, like our rtt
7699                  * measurement. We don't need to do any tricks with shifting
7700                  * etc. Instead we just add in 1/8th of the new measurement
7701                  * and subtract out 1/8 of the old srtt. We do the same with
7702                  * the variance after finding the absolute value of the
7703                  * difference between this sample and the current srtt.
7704                  */
7705                 delta = tp->t_srtt - rtt;
7706                 /* Take off 1/8th of the current sRTT */
7707                 tp->t_srtt -= (tp->t_srtt >> 3);
7708                 /* Add in 1/8th of the new RTT just measured */
7709                 tp->t_srtt += (rtt >> 3);
7710                 if (tp->t_srtt <= 0)
7711                         tp->t_srtt = 1;
7712                 /* Now lets make the absolute value of the variance */
7713                 if (delta < 0)
7714                         delta = -delta;
7715                 /* Subtract out 1/8th */
7716                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7717                 /* Add in 1/8th of the new variance we just saw */
7718                 tp->t_rttvar += (delta >> 3);
7719                 if (tp->t_rttvar <= 0)
7720                         tp->t_rttvar = 1;
7721                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7722                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7723         } else {
7724                 /*
7725                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7726                  * variance to half the rtt (so our first retransmit happens
7727                  * at 3*rtt).
7728                  */
7729                 tp->t_srtt = rtt;
7730                 tp->t_rttvar = rtt >> 1;
7731                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7732         }
7733         rack->rc_srtt_measure_made = 1;
7734         KMOD_TCPSTAT_INC(tcps_rttupdated);
7735         tp->t_rttupdated++;
7736 #ifdef STATS
7737         if (rack_stats_gets_ms_rtt == 0) {
7738                 /* Send in the microsecond rtt used for rxt timeout purposes */
7739                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7740         } else if (rack_stats_gets_ms_rtt == 1) {
7741                 /* Send in the millisecond rtt used for rxt timeout purposes */
7742                 int32_t ms_rtt;
7743
7744                 /* Round up */
7745                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7746                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7747         } else if (rack_stats_gets_ms_rtt == 2) {
7748                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7749                 int32_t ms_rtt;
7750
7751                 /* Round up */
7752                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7753                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7754         }  else {
7755                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7756                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7757         }
7758
7759 #endif
7760         /*
7761          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7762          * way we do the smoothing, srtt and rttvar will each average +1/2
7763          * tick of bias.  When we compute the retransmit timer, we want 1/2
7764          * tick of rounding and 1 extra tick because of +-1/2 tick
7765          * uncertainty in the firing of the timer.  The bias will give us
7766          * exactly the 1.5 tick we need.  But, because the bias is
7767          * statistical, we have to test that we don't drop below the minimum
7768          * feasible timer (which is 2 ticks).
7769          */
7770         tp->t_rxtshift = 0;
7771         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7772                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7773         rack_log_rtt_sample(rack, rtt);
7774         tp->t_softerror = 0;
7775 }
7776
7777
7778 static void
7779 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7780 {
7781         /*
7782          * Apply to filter the inbound us-rtt at us_cts.
7783          */
7784         uint32_t old_rtt;
7785
7786         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7787         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7788                                us_rtt, us_cts);
7789         if (old_rtt > us_rtt) {
7790                 /* We just hit a new lower rtt time */
7791                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7792                                      __LINE__, RACK_RTTS_NEWRTT);
7793                 /*
7794                  * Only count it if its lower than what we saw within our
7795                  * calculated range.
7796                  */
7797                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7798                         if (rack_probertt_lower_within &&
7799                             rack->rc_gp_dyn_mul &&
7800                             (rack->use_fixed_rate == 0) &&
7801                             (rack->rc_always_pace)) {
7802                                 /*
7803                                  * We are seeing a new lower rtt very close
7804                                  * to the time that we would have entered probe-rtt.
7805                                  * This is probably due to the fact that a peer flow
7806                                  * has entered probe-rtt. Lets go in now too.
7807                                  */
7808                                 uint32_t val;
7809
7810                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7811                                 val /= 100;
7812                                 if ((rack->in_probe_rtt == 0)  &&
7813                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7814                                         rack_enter_probertt(rack, us_cts);
7815                                 }
7816                         }
7817                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7818                 }
7819         }
7820 }
7821
7822 static int
7823 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7824     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7825 {
7826         uint32_t us_rtt;
7827         int32_t i, all;
7828         uint32_t t, len_acked;
7829
7830         if ((rsm->r_flags & RACK_ACKED) ||
7831             (rsm->r_flags & RACK_WAS_ACKED))
7832                 /* Already done */
7833                 return (0);
7834         if (rsm->r_no_rtt_allowed) {
7835                 /* Not allowed */
7836                 return (0);
7837         }
7838         if (ack_type == CUM_ACKED) {
7839                 if (SEQ_GT(th_ack, rsm->r_end)) {
7840                         len_acked = rsm->r_end - rsm->r_start;
7841                         all = 1;
7842                 } else {
7843                         len_acked = th_ack - rsm->r_start;
7844                         all = 0;
7845                 }
7846         } else {
7847                 len_acked = rsm->r_end - rsm->r_start;
7848                 all = 0;
7849         }
7850         if (rsm->r_rtr_cnt == 1) {
7851
7852                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7853                 if ((int)t <= 0)
7854                         t = 1;
7855                 if (!tp->t_rttlow || tp->t_rttlow > t)
7856                         tp->t_rttlow = t;
7857                 if (!rack->r_ctl.rc_rack_min_rtt ||
7858                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7859                         rack->r_ctl.rc_rack_min_rtt = t;
7860                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7861                                 rack->r_ctl.rc_rack_min_rtt = 1;
7862                         }
7863                 }
7864                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7865                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7866                 else
7867                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7868                 if (us_rtt == 0)
7869                         us_rtt = 1;
7870                 if (CC_ALGO(tp)->rttsample != NULL) {
7871                         /* Kick the RTT to the CC */
7872                         CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
7873                 }
7874                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7875                 if (ack_type == SACKED) {
7876                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7877                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7878                 } else {
7879                         /*
7880                          * We need to setup what our confidence
7881                          * is in this ack.
7882                          *
7883                          * If the rsm was app limited and it is
7884                          * less than a mss in length (the end
7885                          * of the send) then we have a gap. If we
7886                          * were app limited but say we were sending
7887                          * multiple MSS's then we are more confident
7888                          * int it.
7889                          *
7890                          * When we are not app-limited then we see if
7891                          * the rsm is being included in the current
7892                          * measurement, we tell this by the app_limited_needs_set
7893                          * flag.
7894                          *
7895                          * Note that being cwnd blocked is not applimited
7896                          * as well as the pacing delay between packets which
7897                          * are sending only 1 or 2 MSS's also will show up
7898                          * in the RTT. We probably need to examine this algorithm
7899                          * a bit more and enhance it to account for the delay
7900                          * between rsm's. We could do that by saving off the
7901                          * pacing delay of each rsm (in an rsm) and then
7902                          * factoring that in somehow though for now I am
7903                          * not sure how :)
7904                          */
7905                         int calc_conf = 0;
7906
7907                         if (rsm->r_flags & RACK_APP_LIMITED) {
7908                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7909                                         calc_conf = 0;
7910                                 else
7911                                         calc_conf = 1;
7912                         } else if (rack->app_limited_needs_set == 0) {
7913                                 calc_conf = 1;
7914                         } else {
7915                                 calc_conf = 0;
7916                         }
7917                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7918                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7919                                             calc_conf, rsm, rsm->r_rtr_cnt);
7920                 }
7921                 if ((rsm->r_flags & RACK_TLP) &&
7922                     (!IN_FASTRECOVERY(tp->t_flags))) {
7923                         /* Segment was a TLP and our retrans matched */
7924                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7925                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
7926                         }
7927                 }
7928                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7929                         /* New more recent rack_tmit_time */
7930                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7931                         rack->rc_rack_rtt = t;
7932                 }
7933                 return (1);
7934         }
7935         /*
7936          * We clear the soft/rxtshift since we got an ack.
7937          * There is no assurance we will call the commit() function
7938          * so we need to clear these to avoid incorrect handling.
7939          */
7940         tp->t_rxtshift = 0;
7941         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7942                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7943         tp->t_softerror = 0;
7944         if (to && (to->to_flags & TOF_TS) &&
7945             (ack_type == CUM_ACKED) &&
7946             (to->to_tsecr) &&
7947             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7948                 /*
7949                  * Now which timestamp does it match? In this block the ACK
7950                  * must be coming from a previous transmission.
7951                  */
7952                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7953                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7954                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7955                                 if ((int)t <= 0)
7956                                         t = 1;
7957                                 if (CC_ALGO(tp)->rttsample != NULL) {
7958                                         /*
7959                                          * Kick the RTT to the CC, here
7960                                          * we lie a bit in that we know the
7961                                          * retransmission is correct even though
7962                                          * we retransmitted. This is because
7963                                          * we match the timestamps.
7964                                          */
7965                                         if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
7966                                                 us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
7967                                         else
7968                                                 us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
7969                                         CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
7970                                 }
7971                                 if ((i + 1) < rsm->r_rtr_cnt) {
7972                                         /*
7973                                          * The peer ack'd from our previous
7974                                          * transmission. We have a spurious
7975                                          * retransmission and thus we dont
7976                                          * want to update our rack_rtt.
7977                                          *
7978                                          * Hmm should there be a CC revert here?
7979                                          *
7980                                          */
7981                                         return (0);
7982                                 }
7983                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7984                                         tp->t_rttlow = t;
7985                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7986                                         rack->r_ctl.rc_rack_min_rtt = t;
7987                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7988                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7989                                         }
7990                                 }
7991                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7992                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7993                                         /* New more recent rack_tmit_time */
7994                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7995                                         rack->rc_rack_rtt = t;
7996                                 }
7997                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7998                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7999                                                     rsm->r_rtr_cnt);
8000                                 return (1);
8001                         }
8002                 }
8003                 goto ts_not_found;
8004         } else {
8005                 /*
8006                  * Ok its a SACK block that we retransmitted. or a windows
8007                  * machine without timestamps. We can tell nothing from the
8008                  * time-stamp since its not there or the time the peer last
8009                  * recieved a segment that moved forward its cum-ack point.
8010                  */
8011 ts_not_found:
8012                 i = rsm->r_rtr_cnt - 1;
8013                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8014                 if ((int)t <= 0)
8015                         t = 1;
8016                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8017                         /*
8018                          * We retransmitted and the ack came back in less
8019                          * than the smallest rtt we have observed. We most
8020                          * likely did an improper retransmit as outlined in
8021                          * 6.2 Step 2 point 2 in the rack-draft so we
8022                          * don't want to update our rack_rtt. We in
8023                          * theory (in future) might want to think about reverting our
8024                          * cwnd state but we won't for now.
8025                          */
8026                         return (0);
8027                 } else if (rack->r_ctl.rc_rack_min_rtt) {
8028                         /*
8029                          * We retransmitted it and the retransmit did the
8030                          * job.
8031                          */
8032                         if (!rack->r_ctl.rc_rack_min_rtt ||
8033                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8034                                 rack->r_ctl.rc_rack_min_rtt = t;
8035                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
8036                                         rack->r_ctl.rc_rack_min_rtt = 1;
8037                                 }
8038                         }
8039                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
8040                                 /* New more recent rack_tmit_time */
8041                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
8042                                 rack->rc_rack_rtt = t;
8043                         }
8044                         return (1);
8045                 }
8046         }
8047         return (0);
8048 }
8049
8050 /*
8051  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
8052  */
8053 static void
8054 rack_log_sack_passed(struct tcpcb *tp,
8055     struct tcp_rack *rack, struct rack_sendmap *rsm)
8056 {
8057         struct rack_sendmap *nrsm;
8058
8059         nrsm = rsm;
8060         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
8061             rack_head, r_tnext) {
8062                 if (nrsm == rsm) {
8063                         /* Skip orginal segment he is acked */
8064                         continue;
8065                 }
8066                 if (nrsm->r_flags & RACK_ACKED) {
8067                         /*
8068                          * Skip ack'd segments, though we
8069                          * should not see these, since tmap
8070                          * should not have ack'd segments.
8071                          */
8072                         continue;
8073                 }
8074                 if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
8075                         /*
8076                          * If the peer dropped the rwnd on
8077                          * these then we don't worry about them.
8078                          */
8079                         continue;
8080                 }
8081                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8082                         /*
8083                          * We found one that is already marked
8084                          * passed, we have been here before and
8085                          * so all others below this are marked.
8086                          */
8087                         break;
8088                 }
8089                 nrsm->r_flags |= RACK_SACK_PASSED;
8090                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8091         }
8092 }
8093
8094 static void
8095 rack_need_set_test(struct tcpcb *tp,
8096                    struct tcp_rack *rack,
8097                    struct rack_sendmap *rsm,
8098                    tcp_seq th_ack,
8099                    int line,
8100                    int use_which)
8101 {
8102
8103         if ((tp->t_flags & TF_GPUTINPROG) &&
8104             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8105                 /*
8106                  * We were app limited, and this ack
8107                  * butts up or goes beyond the point where we want
8108                  * to start our next measurement. We need
8109                  * to record the new gput_ts as here and
8110                  * possibly update the start sequence.
8111                  */
8112                 uint32_t seq, ts;
8113
8114                 if (rsm->r_rtr_cnt > 1) {
8115                         /*
8116                          * This is a retransmit, can we
8117                          * really make any assessment at this
8118                          * point?  We are not really sure of
8119                          * the timestamp, is it this or the
8120                          * previous transmission?
8121                          *
8122                          * Lets wait for something better that
8123                          * is not retransmitted.
8124                          */
8125                         return;
8126                 }
8127                 seq = tp->gput_seq;
8128                 ts = tp->gput_ts;
8129                 rack->app_limited_needs_set = 0;
8130                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8131                 /* Do we start at a new end? */
8132                 if ((use_which == RACK_USE_BEG) &&
8133                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8134                         /*
8135                          * When we get an ACK that just eats
8136                          * up some of the rsm, we set RACK_USE_BEG
8137                          * since whats at r_start (i.e. th_ack)
8138                          * is left unacked and thats where the
8139                          * measurement not starts.
8140                          */
8141                         tp->gput_seq = rsm->r_start;
8142                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8143                 }
8144                 if ((use_which == RACK_USE_END) &&
8145                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8146                             /*
8147                              * We use the end when the cumack
8148                              * is moving forward and completely
8149                              * deleting the rsm passed so basically
8150                              * r_end holds th_ack.
8151                              *
8152                              * For SACK's we also want to use the end
8153                              * since this piece just got sacked and
8154                              * we want to target anything after that
8155                              * in our measurement.
8156                              */
8157                             tp->gput_seq = rsm->r_end;
8158                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8159                 }
8160                 if (use_which == RACK_USE_END_OR_THACK) {
8161                         /*
8162                          * special case for ack moving forward,
8163                          * not a sack, we need to move all the
8164                          * way up to where this ack cum-ack moves
8165                          * to.
8166                          */
8167                         if (SEQ_GT(th_ack, rsm->r_end))
8168                                 tp->gput_seq = th_ack;
8169                         else
8170                                 tp->gput_seq = rsm->r_end;
8171                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8172                 }
8173                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8174                         /*
8175                          * We moved beyond this guy's range, re-calculate
8176                          * the new end point.
8177                          */
8178                         if (rack->rc_gp_filled == 0) {
8179                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8180                         } else {
8181                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8182                         }
8183                 }
8184                 /*
8185                  * We are moving the goal post, we may be able to clear the
8186                  * measure_saw_probe_rtt flag.
8187                  */
8188                 if ((rack->in_probe_rtt == 0) &&
8189                     (rack->measure_saw_probe_rtt) &&
8190                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8191                         rack->measure_saw_probe_rtt = 0;
8192                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8193                                            seq, tp->gput_seq, 0, 5, line, NULL, 0);
8194                 if (rack->rc_gp_filled &&
8195                     ((tp->gput_ack - tp->gput_seq) <
8196                      max(rc_init_window(rack), (MIN_GP_WIN *
8197                                                 ctf_fixed_maxseg(tp))))) {
8198                         uint32_t ideal_amount;
8199
8200                         ideal_amount = rack_get_measure_window(tp, rack);
8201                         if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8202                                 /*
8203                                  * There is no sense of continuing this measurement
8204                                  * because its too small to gain us anything we
8205                                  * trust. Skip it and that way we can start a new
8206                                  * measurement quicker.
8207                                  */
8208                                 tp->t_flags &= ~TF_GPUTINPROG;
8209                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8210                                                            0, 0, 0, 6, __LINE__, NULL, 0);
8211                         } else {
8212                                 /*
8213                                  * Reset the window further out.
8214                                  */
8215                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8216                         }
8217                 }
8218         }
8219 }
8220
8221 static inline int
8222 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8223 {
8224         if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8225                 /* Behind our TLP definition or right at */
8226                 return (0);
8227         }
8228         if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8229                 /* The start is beyond or right at our end of TLP definition */
8230                 return (0);
8231         }
8232         /* It has to be a sub-part of the original TLP recorded */
8233         return (1);
8234 }
8235
8236
8237 static uint32_t
8238 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8239                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8240 {
8241         uint32_t start, end, changed = 0;
8242         struct rack_sendmap stack_map;
8243         struct rack_sendmap *rsm, *nrsm, fe, *prev, *next;
8244 #ifdef INVARIANTS
8245         struct rack_sendmap *insret;
8246 #endif
8247         int32_t used_ref = 1;
8248         int moved = 0;
8249
8250         start = sack->start;
8251         end = sack->end;
8252         rsm = *prsm;
8253         memset(&fe, 0, sizeof(fe));
8254 do_rest_ofb:
8255         if ((rsm == NULL) ||
8256             (SEQ_LT(end, rsm->r_start)) ||
8257             (SEQ_GEQ(start, rsm->r_end)) ||
8258             (SEQ_LT(start, rsm->r_start))) {
8259                 /*
8260                  * We are not in the right spot,
8261                  * find the correct spot in the tree.
8262                  */
8263                 used_ref = 0;
8264                 fe.r_start = start;
8265                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8266                 moved++;
8267         }
8268         if (rsm == NULL) {
8269                 /* TSNH */
8270                 goto out;
8271         }
8272         /* Ok we have an ACK for some piece of this rsm */
8273         if (rsm->r_start != start) {
8274                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8275                         /*
8276                          * Before any splitting or hookery is
8277                          * done is it a TLP of interest i.e. rxt?
8278                          */
8279                         if ((rsm->r_flags & RACK_TLP) &&
8280                             (rsm->r_rtr_cnt > 1)) {
8281                                 /*
8282                                  * We are splitting a rxt TLP, check
8283                                  * if we need to save off the start/end
8284                                  */
8285                                 if (rack->rc_last_tlp_acked_set &&
8286                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8287                                         /*
8288                                          * We already turned this on since we are inside
8289                                          * the previous one was a partially sack now we
8290                                          * are getting another one (maybe all of it).
8291                                          *
8292                                          */
8293                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8294                                         /*
8295                                          * Lets make sure we have all of it though.
8296                                          */
8297                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8298                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8299                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8300                                                                      rack->r_ctl.last_tlp_acked_end);
8301                                         }
8302                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8303                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8304                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8305                                                                      rack->r_ctl.last_tlp_acked_end);
8306                                         }
8307                                 } else {
8308                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8309                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8310                                         rack->rc_last_tlp_past_cumack = 0;
8311                                         rack->rc_last_tlp_acked_set = 1;
8312                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8313                                 }
8314                         }
8315                         /**
8316                          * Need to split this in two pieces the before and after,
8317                          * the before remains in the map, the after must be
8318                          * added. In other words we have:
8319                          * rsm        |--------------|
8320                          * sackblk        |------->
8321                          * rsm will become
8322                          *     rsm    |---|
8323                          * and nrsm will be  the sacked piece
8324                          *     nrsm       |----------|
8325                          *
8326                          * But before we start down that path lets
8327                          * see if the sack spans over on top of
8328                          * the next guy and it is already sacked.
8329                          *
8330                          */
8331                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8332                         if (next && (next->r_flags & RACK_ACKED) &&
8333                             SEQ_GEQ(end, next->r_start)) {
8334                                 /**
8335                                  * So the next one is already acked, and
8336                                  * we can thus by hookery use our stack_map
8337                                  * to reflect the piece being sacked and
8338                                  * then adjust the two tree entries moving
8339                                  * the start and ends around. So we start like:
8340                                  *  rsm     |------------|             (not-acked)
8341                                  *  next                 |-----------| (acked)
8342                                  *  sackblk        |-------->
8343                                  *  We want to end like so:
8344                                  *  rsm     |------|                   (not-acked)
8345                                  *  next           |-----------------| (acked)
8346                                  *  nrsm           |-----|
8347                                  * Where nrsm is a temporary stack piece we
8348                                  * use to update all the gizmos.
8349                                  */
8350                                 /* Copy up our fudge block */
8351                                 nrsm = &stack_map;
8352                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8353                                 /* Now adjust our tree blocks */
8354                                 rsm->r_end = start;
8355                                 next->r_start = start;
8356                                 /* Now we must adjust back where next->m is */
8357                                 rack_setup_offset_for_rsm(rsm, next);
8358
8359                                 /* We don't need to adjust rsm, it did not change */
8360                                 /* Clear out the dup ack count of the remainder */
8361                                 rsm->r_dupack = 0;
8362                                 rsm->r_just_ret = 0;
8363                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8364                                 /* Now lets make sure our fudge block is right */
8365                                 nrsm->r_start = start;
8366                                 /* Now lets update all the stats and such */
8367                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8368                                 if (rack->app_limited_needs_set)
8369                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8370                                 changed += (nrsm->r_end - nrsm->r_start);
8371                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8372                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8373                                         rack->r_ctl.rc_reorder_ts = cts;
8374                                 }
8375                                 /*
8376                                  * Now we want to go up from rsm (the
8377                                  * one left un-acked) to the next one
8378                                  * in the tmap. We do this so when
8379                                  * we walk backwards we include marking
8380                                  * sack-passed on rsm (The one passed in
8381                                  * is skipped since it is generally called
8382                                  * on something sacked before removing it
8383                                  * from the tmap).
8384                                  */
8385                                 if (rsm->r_in_tmap) {
8386                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8387                                         /*
8388                                          * Now that we have the next
8389                                          * one walk backwards from there.
8390                                          */
8391                                         if (nrsm && nrsm->r_in_tmap)
8392                                                 rack_log_sack_passed(tp, rack, nrsm);
8393                                 }
8394                                 /* Now are we done? */
8395                                 if (SEQ_LT(end, next->r_end) ||
8396                                     (end == next->r_end)) {
8397                                         /* Done with block */
8398                                         goto out;
8399                                 }
8400                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8401                                 counter_u64_add(rack_sack_used_next_merge, 1);
8402                                 /* Postion for the next block */
8403                                 start = next->r_end;
8404                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8405                                 if (rsm == NULL)
8406                                         goto out;
8407                         } else {
8408                                 /**
8409                                  * We can't use any hookery here, so we
8410                                  * need to split the map. We enter like
8411                                  * so:
8412                                  *  rsm      |--------|
8413                                  *  sackblk       |----->
8414                                  * We will add the new block nrsm and
8415                                  * that will be the new portion, and then
8416                                  * fall through after reseting rsm. So we
8417                                  * split and look like this:
8418                                  *  rsm      |----|
8419                                  *  sackblk       |----->
8420                                  *  nrsm          |---|
8421                                  * We then fall through reseting
8422                                  * rsm to nrsm, so the next block
8423                                  * picks it up.
8424                                  */
8425                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8426                                 if (nrsm == NULL) {
8427                                         /*
8428                                          * failed XXXrrs what can we do but loose the sack
8429                                          * info?
8430                                          */
8431                                         goto out;
8432                                 }
8433                                 counter_u64_add(rack_sack_splits, 1);
8434                                 rack_clone_rsm(rack, nrsm, rsm, start);
8435                                 rsm->r_just_ret = 0;
8436 #ifndef INVARIANTS
8437                                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8438 #else
8439                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8440                                 if (insret != NULL) {
8441                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8442                                               nrsm, insret, rack, rsm);
8443                                 }
8444 #endif
8445                                 if (rsm->r_in_tmap) {
8446                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8447                                         nrsm->r_in_tmap = 1;
8448                                 }
8449                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8450                                 rsm->r_flags &= (~RACK_HAS_FIN);
8451                                 /* Position us to point to the new nrsm that starts the sack blk */
8452                                 rsm = nrsm;
8453                         }
8454                 } else {
8455                         /* Already sacked this piece */
8456                         counter_u64_add(rack_sack_skipped_acked, 1);
8457                         moved++;
8458                         if (end == rsm->r_end) {
8459                                 /* Done with block */
8460                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8461                                 goto out;
8462                         } else if (SEQ_LT(end, rsm->r_end)) {
8463                                 /* A partial sack to a already sacked block */
8464                                 moved++;
8465                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8466                                 goto out;
8467                         } else {
8468                                 /*
8469                                  * The end goes beyond this guy
8470                                  * reposition the start to the
8471                                  * next block.
8472                                  */
8473                                 start = rsm->r_end;
8474                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8475                                 if (rsm == NULL)
8476                                         goto out;
8477                         }
8478                 }
8479         }
8480         if (SEQ_GEQ(end, rsm->r_end)) {
8481                 /**
8482                  * The end of this block is either beyond this guy or right
8483                  * at this guy. I.e.:
8484                  *  rsm ---                 |-----|
8485                  *  end                     |-----|
8486                  *  <or>
8487                  *  end                     |---------|
8488                  */
8489                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8490                         /*
8491                          * Is it a TLP of interest?
8492                          */
8493                         if ((rsm->r_flags & RACK_TLP) &&
8494                             (rsm->r_rtr_cnt > 1)) {
8495                                 /*
8496                                  * We are splitting a rxt TLP, check
8497                                  * if we need to save off the start/end
8498                                  */
8499                                 if (rack->rc_last_tlp_acked_set &&
8500                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8501                                         /*
8502                                          * We already turned this on since we are inside
8503                                          * the previous one was a partially sack now we
8504                                          * are getting another one (maybe all of it).
8505                                          */
8506                                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8507                                         /*
8508                                          * Lets make sure we have all of it though.
8509                                          */
8510                                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8511                                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8512                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8513                                                                      rack->r_ctl.last_tlp_acked_end);
8514                                         }
8515                                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8516                                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8517                                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8518                                                                      rack->r_ctl.last_tlp_acked_end);
8519                                         }
8520                                 } else {
8521                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8522                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8523                                         rack->rc_last_tlp_past_cumack = 0;
8524                                         rack->rc_last_tlp_acked_set = 1;
8525                                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8526                                 }
8527                         }
8528                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8529                         changed += (rsm->r_end - rsm->r_start);
8530                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8531                         if (rsm->r_in_tmap) /* should be true */
8532                                 rack_log_sack_passed(tp, rack, rsm);
8533                         /* Is Reordering occuring? */
8534                         if (rsm->r_flags & RACK_SACK_PASSED) {
8535                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8536                                 rack->r_ctl.rc_reorder_ts = cts;
8537                         }
8538                         if (rack->app_limited_needs_set)
8539                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8540                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8541                         rsm->r_flags |= RACK_ACKED;
8542                         if (rsm->r_in_tmap) {
8543                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8544                                 rsm->r_in_tmap = 0;
8545                         }
8546                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8547                 } else {
8548                         counter_u64_add(rack_sack_skipped_acked, 1);
8549                         moved++;
8550                 }
8551                 if (end == rsm->r_end) {
8552                         /* This block only - done, setup for next */
8553                         goto out;
8554                 }
8555                 /*
8556                  * There is more not coverend by this rsm move on
8557                  * to the next block in the RB tree.
8558                  */
8559                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8560                 start = rsm->r_end;
8561                 rsm = nrsm;
8562                 if (rsm == NULL)
8563                         goto out;
8564                 goto do_rest_ofb;
8565         }
8566         /**
8567          * The end of this sack block is smaller than
8568          * our rsm i.e.:
8569          *  rsm ---                 |-----|
8570          *  end                     |--|
8571          */
8572         if ((rsm->r_flags & RACK_ACKED) == 0) {
8573                 /*
8574                  * Is it a TLP of interest?
8575                  */
8576                 if ((rsm->r_flags & RACK_TLP) &&
8577                     (rsm->r_rtr_cnt > 1)) {
8578                         /*
8579                          * We are splitting a rxt TLP, check
8580                          * if we need to save off the start/end
8581                          */
8582                         if (rack->rc_last_tlp_acked_set &&
8583                             (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8584                                 /*
8585                                  * We already turned this on since we are inside
8586                                  * the previous one was a partially sack now we
8587                                  * are getting another one (maybe all of it).
8588                                  */
8589                                 rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8590                                 /*
8591                                  * Lets make sure we have all of it though.
8592                                  */
8593                                 if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8594                                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8595                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8596                                                              rack->r_ctl.last_tlp_acked_end);
8597                                 }
8598                                 if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8599                                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8600                                         rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8601                                                              rack->r_ctl.last_tlp_acked_end);
8602                                 }
8603                         } else {
8604                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8605                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8606                                 rack->rc_last_tlp_past_cumack = 0;
8607                                 rack->rc_last_tlp_acked_set = 1;
8608                                 rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8609                         }
8610                 }
8611                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8612                 if (prev &&
8613                     (prev->r_flags & RACK_ACKED)) {
8614                         /**
8615                          * Goal, we want the right remainder of rsm to shrink
8616                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8617                          * We want to expand prev to go all the way
8618                          * to prev->r_end <- end.
8619                          * so in the tree we have before:
8620                          *   prev     |--------|         (acked)
8621                          *   rsm               |-------| (non-acked)
8622                          *   sackblk           |-|
8623                          * We churn it so we end up with
8624                          *   prev     |----------|       (acked)
8625                          *   rsm                 |-----| (non-acked)
8626                          *   nrsm              |-| (temporary)
8627                          *
8628                          * Note if either prev/rsm is a TLP we don't
8629                          * do this.
8630                          */
8631                         nrsm = &stack_map;
8632                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8633                         prev->r_end = end;
8634                         rsm->r_start = end;
8635                         /* Now adjust nrsm (stack copy) to be
8636                          * the one that is the small
8637                          * piece that was "sacked".
8638                          */
8639                         nrsm->r_end = end;
8640                         rsm->r_dupack = 0;
8641                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8642                         /*
8643                          * Now that the rsm has had its start moved forward
8644                          * lets go ahead and get its new place in the world.
8645                          */
8646                         rack_setup_offset_for_rsm(prev, rsm);
8647                         /*
8648                          * Now nrsm is our new little piece
8649                          * that is acked (which was merged
8650                          * to prev). Update the rtt and changed
8651                          * based on that. Also check for reordering.
8652                          */
8653                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8654                         if (rack->app_limited_needs_set)
8655                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8656                         changed += (nrsm->r_end - nrsm->r_start);
8657                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8658                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8659                                 rack->r_ctl.rc_reorder_ts = cts;
8660                         }
8661                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8662                         rsm = prev;
8663                         counter_u64_add(rack_sack_used_prev_merge, 1);
8664                 } else {
8665                         /**
8666                          * This is the case where our previous
8667                          * block is not acked either, so we must
8668                          * split the block in two.
8669                          */
8670                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8671                         if (nrsm == NULL) {
8672                                 /* failed rrs what can we do but loose the sack info? */
8673                                 goto out;
8674                         }
8675                         if ((rsm->r_flags & RACK_TLP) &&
8676                             (rsm->r_rtr_cnt > 1)) {
8677                                 /*
8678                                  * We are splitting a rxt TLP, check
8679                                  * if we need to save off the start/end
8680                                  */
8681                                 if (rack->rc_last_tlp_acked_set &&
8682                                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8683                                             /*
8684                                              * We already turned this on since this block is inside
8685                                              * the previous one was a partially sack now we
8686                                              * are getting another one (maybe all of it).
8687                                              */
8688                                             rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8689                                             /*
8690                                              * Lets make sure we have all of it though.
8691                                              */
8692                                             if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8693                                                     rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8694                                                     rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8695                                                                          rack->r_ctl.last_tlp_acked_end);
8696                                             }
8697                                             if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8698                                                     rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8699                                                     rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8700                                                                          rack->r_ctl.last_tlp_acked_end);
8701                                             }
8702                                     } else {
8703                                             rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8704                                             rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8705                                             rack->rc_last_tlp_acked_set = 1;
8706                                             rack->rc_last_tlp_past_cumack = 0;
8707                                             rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8708                                     }
8709                         }
8710                         /**
8711                          * In this case nrsm becomes
8712                          * nrsm->r_start = end;
8713                          * nrsm->r_end = rsm->r_end;
8714                          * which is un-acked.
8715                          * <and>
8716                          * rsm->r_end = nrsm->r_start;
8717                          * i.e. the remaining un-acked
8718                          * piece is left on the left
8719                          * hand side.
8720                          *
8721                          * So we start like this
8722                          * rsm      |----------| (not acked)
8723                          * sackblk  |---|
8724                          * build it so we have
8725                          * rsm      |---|         (acked)
8726                          * nrsm         |------|  (not acked)
8727                          */
8728                         counter_u64_add(rack_sack_splits, 1);
8729                         rack_clone_rsm(rack, nrsm, rsm, end);
8730                         rsm->r_flags &= (~RACK_HAS_FIN);
8731                         rsm->r_just_ret = 0;
8732 #ifndef INVARIANTS
8733                         (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8734 #else
8735                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8736                         if (insret != NULL) {
8737                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8738                                       nrsm, insret, rack, rsm);
8739                         }
8740 #endif
8741                         if (rsm->r_in_tmap) {
8742                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8743                                 nrsm->r_in_tmap = 1;
8744                         }
8745                         nrsm->r_dupack = 0;
8746                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8747                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8748                         changed += (rsm->r_end - rsm->r_start);
8749                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8750                         if (rsm->r_in_tmap) /* should be true */
8751                                 rack_log_sack_passed(tp, rack, rsm);
8752                         /* Is Reordering occuring? */
8753                         if (rsm->r_flags & RACK_SACK_PASSED) {
8754                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8755                                 rack->r_ctl.rc_reorder_ts = cts;
8756                         }
8757                         if (rack->app_limited_needs_set)
8758                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8759                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8760                         rsm->r_flags |= RACK_ACKED;
8761                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8762                         if (rsm->r_in_tmap) {
8763                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8764                                 rsm->r_in_tmap = 0;
8765                         }
8766                 }
8767         } else if (start != end){
8768                 /*
8769                  * The block was already acked.
8770                  */
8771                 counter_u64_add(rack_sack_skipped_acked, 1);
8772                 moved++;
8773         }
8774 out:
8775         if (rsm &&
8776             ((rsm->r_flags & RACK_TLP) == 0) &&
8777             (rsm->r_flags & RACK_ACKED)) {
8778                 /*
8779                  * Now can we merge where we worked
8780                  * with either the previous or
8781                  * next block?
8782                  */
8783                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8784                 while (next) {
8785                         if (next->r_flags & RACK_TLP)
8786                                 break;
8787                         if (next->r_flags & RACK_ACKED) {
8788                         /* yep this and next can be merged */
8789                                 rsm = rack_merge_rsm(rack, rsm, next);
8790                                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8791                         } else
8792                                 break;
8793                 }
8794                 /* Now what about the previous? */
8795                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8796                 while (prev) {
8797                         if (prev->r_flags & RACK_TLP)
8798                                 break;
8799                         if (prev->r_flags & RACK_ACKED) {
8800                                 /* yep the previous and this can be merged */
8801                                 rsm = rack_merge_rsm(rack, prev, rsm);
8802                                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8803                         } else
8804                                 break;
8805                 }
8806         }
8807         if (used_ref == 0) {
8808                 counter_u64_add(rack_sack_proc_all, 1);
8809         } else {
8810                 counter_u64_add(rack_sack_proc_short, 1);
8811         }
8812         /* Save off the next one for quick reference. */
8813         if (rsm)
8814                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8815         else
8816                 nrsm = NULL;
8817         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8818         /* Pass back the moved. */
8819         *moved_two = moved;
8820         return (changed);
8821 }
8822
8823 static void inline
8824 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8825 {
8826         struct rack_sendmap *tmap;
8827
8828         tmap = NULL;
8829         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8830                 /* Its no longer sacked, mark it so */
8831                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8832 #ifdef INVARIANTS
8833                 if (rsm->r_in_tmap) {
8834                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8835                               rack, rsm, rsm->r_flags);
8836                 }
8837 #endif
8838                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8839                 /* Rebuild it into our tmap */
8840                 if (tmap == NULL) {
8841                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8842                         tmap = rsm;
8843                 } else {
8844                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8845                         tmap = rsm;
8846                 }
8847                 tmap->r_in_tmap = 1;
8848                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8849         }
8850         /*
8851          * Now lets possibly clear the sack filter so we start
8852          * recognizing sacks that cover this area.
8853          */
8854         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8855
8856 }
8857
8858 static void
8859 rack_do_decay(struct tcp_rack *rack)
8860 {
8861         struct timeval res;
8862
8863 #define timersub(tvp, uvp, vvp)                                         \
8864         do {                                                            \
8865                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8866                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8867                 if ((vvp)->tv_usec < 0) {                               \
8868                         (vvp)->tv_sec--;                                \
8869                         (vvp)->tv_usec += 1000000;                      \
8870                 }                                                       \
8871         } while (0)
8872
8873         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8874 #undef timersub
8875
8876         rack->r_ctl.input_pkt++;
8877         if ((rack->rc_in_persist) ||
8878             (res.tv_sec >= 1) ||
8879             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8880                 /*
8881                  * Check for decay of non-SAD,
8882                  * we want all SAD detection metrics to
8883                  * decay 1/4 per second (or more) passed.
8884                  */
8885 #ifdef NETFLIX_EXP_DETECTION
8886                 uint32_t pkt_delta;
8887
8888                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8889 #endif
8890                 /* Update our saved tracking values */
8891                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8892                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8893                 /* Now do we escape without decay? */
8894 #ifdef NETFLIX_EXP_DETECTION
8895                 if (rack->rc_in_persist ||
8896                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8897                     (pkt_delta < tcp_sad_low_pps)){
8898                         /*
8899                          * We don't decay idle connections
8900                          * or ones that have a low input pps.
8901                          */
8902                         return;
8903                 }
8904                 /* Decay the counters */
8905                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8906                                                         tcp_sad_decay_val);
8907                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8908                                                          tcp_sad_decay_val);
8909                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8910                                                                tcp_sad_decay_val);
8911                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8912                                                                 tcp_sad_decay_val);
8913 #endif
8914         }
8915 }
8916
8917 static void
8918 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8919 {
8920         struct rack_sendmap *rsm;
8921 #ifdef INVARIANTS
8922         struct rack_sendmap *rm;
8923 #endif
8924
8925         /*
8926          * The ACK point is advancing to th_ack, we must drop off
8927          * the packets in the rack log and calculate any eligble
8928          * RTT's.
8929          */
8930         rack->r_wanted_output = 1;
8931
8932         /* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
8933         if ((rack->rc_last_tlp_acked_set == 1)&&
8934             (rack->rc_last_tlp_past_cumack == 1) &&
8935             (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
8936                 /*
8937                  * We have reached the point where our last rack
8938                  * tlp retransmit sequence is ahead of the cum-ack.
8939                  * This can only happen when the cum-ack moves all
8940                  * the way around (its been a full 2^^31+1 bytes
8941                  * or more since we sent a retransmitted TLP). Lets
8942                  * turn off the valid flag since its not really valid.
8943                  *
8944                  * Note since sack's also turn on this event we have
8945                  * a complication, we have to wait to age it out until
8946                  * the cum-ack is by the TLP before checking which is
8947                  * what the next else clause does.
8948                  */
8949                 rack_log_dsack_event(rack, 9, __LINE__,
8950                                      rack->r_ctl.last_tlp_acked_start,
8951                                      rack->r_ctl.last_tlp_acked_end);
8952                 rack->rc_last_tlp_acked_set = 0;
8953                 rack->rc_last_tlp_past_cumack = 0;
8954         } else if ((rack->rc_last_tlp_acked_set == 1) &&
8955                    (rack->rc_last_tlp_past_cumack == 0) &&
8956                    (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
8957                 /*
8958                  * It is safe to start aging TLP's out.
8959                  */
8960                 rack->rc_last_tlp_past_cumack = 1;
8961         }
8962         /* We do the same for the tlp send seq as well */
8963         if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8964             (rack->rc_last_sent_tlp_past_cumack == 1) &&
8965             (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
8966                 rack_log_dsack_event(rack, 9, __LINE__,
8967                                      rack->r_ctl.last_sent_tlp_seq,
8968                                      (rack->r_ctl.last_sent_tlp_seq +
8969                                       rack->r_ctl.last_sent_tlp_len));
8970                 rack->rc_last_sent_tlp_seq_valid = 0;
8971                 rack->rc_last_sent_tlp_past_cumack = 0;
8972         } else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8973                    (rack->rc_last_sent_tlp_past_cumack == 0) &&
8974                    (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
8975                 /*
8976                  * It is safe to start aging TLP's send.
8977                  */
8978                 rack->rc_last_sent_tlp_past_cumack = 1;
8979         }
8980 more:
8981         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8982         if (rsm == NULL) {
8983                 if ((th_ack - 1) == tp->iss) {
8984                         /*
8985                          * For the SYN incoming case we will not
8986                          * have called tcp_output for the sending of
8987                          * the SYN, so there will be no map. All
8988                          * other cases should probably be a panic.
8989                          */
8990                         return;
8991                 }
8992                 if (tp->t_flags & TF_SENTFIN) {
8993                         /* if we sent a FIN we often will not have map */
8994                         return;
8995                 }
8996 #ifdef INVARIANTS
8997                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8998                       tp,
8999                       tp->t_state, th_ack, rack,
9000                       tp->snd_una, tp->snd_max, tp->snd_nxt);
9001 #endif
9002                 return;
9003         }
9004         if (SEQ_LT(th_ack, rsm->r_start)) {
9005                 /* Huh map is missing this */
9006 #ifdef INVARIANTS
9007                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
9008                        rsm->r_start,
9009                        th_ack, tp->t_state, rack->r_state);
9010 #endif
9011                 return;
9012         }
9013         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
9014
9015         /* Now was it a retransmitted TLP? */
9016         if ((rsm->r_flags & RACK_TLP) &&
9017             (rsm->r_rtr_cnt > 1)) {
9018                 /*
9019                  * Yes, this rsm was a TLP and retransmitted, remember that
9020                  * since if a DSACK comes back on this we don't want
9021                  * to think of it as a reordered segment. This may
9022                  * get updated again with possibly even other TLPs
9023                  * in flight, but thats ok. Only when we don't send
9024                  * a retransmitted TLP for 1/2 the sequences space
9025                  * will it get turned off (above).
9026                  */
9027                 if (rack->rc_last_tlp_acked_set &&
9028                     (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9029                         /*
9030                          * We already turned this on since the end matches,
9031                          * the previous one was a partially ack now we
9032                          * are getting another one (maybe all of it).
9033                          */
9034                         rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9035                         /*
9036                          * Lets make sure we have all of it though.
9037                          */
9038                         if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9039                                 rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9040                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9041                                                      rack->r_ctl.last_tlp_acked_end);
9042                         }
9043                         if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9044                                 rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9045                                 rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9046                                                      rack->r_ctl.last_tlp_acked_end);
9047                         }
9048                 } else {
9049                         rack->rc_last_tlp_past_cumack = 1;
9050                         rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9051                         rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9052                         rack->rc_last_tlp_acked_set = 1;
9053                         rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9054                 }
9055         }
9056         /* Now do we consume the whole thing? */
9057         if (SEQ_GEQ(th_ack, rsm->r_end)) {
9058                 /* Its all consumed. */
9059                 uint32_t left;
9060                 uint8_t newly_acked;
9061
9062                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
9063                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
9064                 rsm->r_rtr_bytes = 0;
9065                 /* Record the time of highest cumack sent */
9066                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9067 #ifndef INVARIANTS
9068                 (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
9069 #else
9070                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
9071                 if (rm != rsm) {
9072                         panic("removing head in rack:%p rsm:%p rm:%p",
9073                               rack, rsm, rm);
9074                 }
9075 #endif
9076                 if (rsm->r_in_tmap) {
9077                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9078                         rsm->r_in_tmap = 0;
9079                 }
9080                 newly_acked = 1;
9081                 if (rsm->r_flags & RACK_ACKED) {
9082                         /*
9083                          * It was acked on the scoreboard -- remove
9084                          * it from total
9085                          */
9086                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
9087                         newly_acked = 0;
9088                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
9089                         /*
9090                          * There are segments ACKED on the
9091                          * scoreboard further up. We are seeing
9092                          * reordering.
9093                          */
9094                         rsm->r_flags &= ~RACK_SACK_PASSED;
9095                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9096                         rsm->r_flags |= RACK_ACKED;
9097                         rack->r_ctl.rc_reorder_ts = cts;
9098                         if (rack->r_ent_rec_ns) {
9099                                 /*
9100                                  * We have sent no more, and we saw an sack
9101                                  * then ack arrive.
9102                                  */
9103                                 rack->r_might_revert = 1;
9104                         }
9105                 }
9106                 if ((rsm->r_flags & RACK_TO_REXT) &&
9107                     (tp->t_flags & TF_RCVD_TSTMP) &&
9108                     (to->to_flags & TOF_TS) &&
9109                     (to->to_tsecr != 0) &&
9110                     (tp->t_flags & TF_PREVVALID)) {
9111                         /*
9112                          * We can use the timestamp to see
9113                          * if this retransmission was from the
9114                          * first transmit. If so we made a mistake.
9115                          */
9116                         tp->t_flags &= ~TF_PREVVALID;
9117                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9118                                 /* The first transmit is what this ack is for */
9119                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
9120                         }
9121                 }
9122                 left = th_ack - rsm->r_end;
9123                 if (rack->app_limited_needs_set && newly_acked)
9124                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9125                 /* Free back to zone */
9126                 rack_free(rack, rsm);
9127                 if (left) {
9128                         goto more;
9129                 }
9130                 /* Check for reneging */
9131                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9132                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9133                         /*
9134                          * The peer has moved snd_una up to
9135                          * the edge of this send, i.e. one
9136                          * that it had previously acked. The only
9137                          * way that can be true if the peer threw
9138                          * away data (space issues) that it had
9139                          * previously sacked (else it would have
9140                          * given us snd_una up to (rsm->r_end).
9141                          * We need to undo the acked markings here.
9142                          *
9143                          * Note we have to look to make sure th_ack is
9144                          * our rsm->r_start in case we get an old ack
9145                          * where th_ack is behind snd_una.
9146                          */
9147                         rack_peer_reneges(rack, rsm, th_ack);
9148                 }
9149                 return;
9150         }
9151         if (rsm->r_flags & RACK_ACKED) {
9152                 /*
9153                  * It was acked on the scoreboard -- remove it from
9154                  * total for the part being cum-acked.
9155                  */
9156                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9157         }
9158         /*
9159          * Clear the dup ack count for
9160          * the piece that remains.
9161          */
9162         rsm->r_dupack = 0;
9163         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9164         if (rsm->r_rtr_bytes) {
9165                 /*
9166                  * It was retransmitted adjust the
9167                  * sack holes for what was acked.
9168                  */
9169                 int ack_am;
9170
9171                 ack_am = (th_ack - rsm->r_start);
9172                 if (ack_am >= rsm->r_rtr_bytes) {
9173                         rack->r_ctl.rc_holes_rxt -= ack_am;
9174                         rsm->r_rtr_bytes -= ack_am;
9175                 }
9176         }
9177         /*
9178          * Update where the piece starts and record
9179          * the time of send of highest cumack sent.
9180          */
9181         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9182         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9183         /* Now we need to move our offset forward too */
9184         if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9185                 /* Fix up the orig_m_len and possibly the mbuf offset */
9186                 rack_adjust_orig_mlen(rsm);
9187         }
9188         rsm->soff += (th_ack - rsm->r_start);
9189         rsm->r_start = th_ack;
9190         /* Now do we need to move the mbuf fwd too? */
9191         if (rsm->m) {
9192                 while (rsm->soff >= rsm->m->m_len) {
9193                         rsm->soff -= rsm->m->m_len;
9194                         rsm->m = rsm->m->m_next;
9195                         KASSERT((rsm->m != NULL),
9196                                 (" nrsm:%p hit at soff:%u null m",
9197                                  rsm, rsm->soff));
9198                 }
9199                 rsm->orig_m_len = rsm->m->m_len;
9200         }
9201         if (rack->app_limited_needs_set)
9202                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9203 }
9204
9205 static void
9206 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9207 {
9208         struct rack_sendmap *rsm;
9209         int sack_pass_fnd = 0;
9210
9211         if (rack->r_might_revert) {
9212                 /*
9213                  * Ok we have reordering, have not sent anything, we
9214                  * might want to revert the congestion state if nothing
9215                  * further has SACK_PASSED on it. Lets check.
9216                  *
9217                  * We also get here when we have DSACKs come in for
9218                  * all the data that we FR'd. Note that a rxt or tlp
9219                  * timer clears this from happening.
9220                  */
9221
9222                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9223                         if (rsm->r_flags & RACK_SACK_PASSED) {
9224                                 sack_pass_fnd = 1;
9225                                 break;
9226                         }
9227                 }
9228                 if (sack_pass_fnd == 0) {
9229                         /*
9230                          * We went into recovery
9231                          * incorrectly due to reordering!
9232                          */
9233                         int orig_cwnd;
9234
9235                         rack->r_ent_rec_ns = 0;
9236                         orig_cwnd = tp->snd_cwnd;
9237                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9238                         tp->snd_recover = tp->snd_una;
9239                         rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
9240                         EXIT_RECOVERY(tp->t_flags);
9241                 }
9242                 rack->r_might_revert = 0;
9243         }
9244 }
9245
9246 #ifdef NETFLIX_EXP_DETECTION
9247 static void
9248 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9249 {
9250         if ((rack->do_detection || tcp_force_detection) &&
9251             tcp_sack_to_ack_thresh &&
9252             tcp_sack_to_move_thresh &&
9253             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9254                 /*
9255                  * We have thresholds set to find
9256                  * possible attackers and disable sack.
9257                  * Check them.
9258                  */
9259                 uint64_t ackratio, moveratio, movetotal;
9260
9261                 /* Log detecting */
9262                 rack_log_sad(rack, 1);
9263                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
9264                 ackratio *= (uint64_t)(1000);
9265                 if (rack->r_ctl.ack_count)
9266                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9267                 else {
9268                         /* We really should not hit here */
9269                         ackratio = 1000;
9270                 }
9271                 if ((rack->sack_attack_disable == 0) &&
9272                     (ackratio > rack_highest_sack_thresh_seen))
9273                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9274                 movetotal = rack->r_ctl.sack_moved_extra;
9275                 movetotal += rack->r_ctl.sack_noextra_move;
9276                 moveratio = rack->r_ctl.sack_moved_extra;
9277                 moveratio *= (uint64_t)1000;
9278                 if (movetotal)
9279                         moveratio /= movetotal;
9280                 else {
9281                         /* No moves, thats pretty good */
9282                         moveratio = 0;
9283                 }
9284                 if ((rack->sack_attack_disable == 0) &&
9285                     (moveratio > rack_highest_move_thresh_seen))
9286                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
9287                 if (rack->sack_attack_disable == 0) {
9288                         if ((ackratio > tcp_sack_to_ack_thresh) &&
9289                             (moveratio > tcp_sack_to_move_thresh)) {
9290                                 /* Disable sack processing */
9291                                 rack->sack_attack_disable = 1;
9292                                 if (rack->r_rep_attack == 0) {
9293                                         rack->r_rep_attack = 1;
9294                                         counter_u64_add(rack_sack_attacks_detected, 1);
9295                                 }
9296                                 if (tcp_attack_on_turns_on_logging) {
9297                                         /*
9298                                          * Turn on logging, used for debugging
9299                                          * false positives.
9300                                          */
9301                                         rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9302                                 }
9303                                 /* Clamp the cwnd at flight size */
9304                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9305                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9306                                 rack_log_sad(rack, 2);
9307                         }
9308                 } else {
9309                         /* We are sack-disabled check for false positives */
9310                         if ((ackratio <= tcp_restoral_thresh) ||
9311                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9312                                 rack->sack_attack_disable = 0;
9313                                 rack_log_sad(rack, 3);
9314                                 /* Restart counting */
9315                                 rack->r_ctl.sack_count = 0;
9316                                 rack->r_ctl.sack_moved_extra = 0;
9317                                 rack->r_ctl.sack_noextra_move = 1;
9318                                 rack->r_ctl.ack_count = max(1,
9319                                       (bytes_this_ack / segsiz));
9320
9321                                 if (rack->r_rep_reverse == 0) {
9322                                         rack->r_rep_reverse = 1;
9323                                         counter_u64_add(rack_sack_attacks_reversed, 1);
9324                                 }
9325                                 /* Restore the cwnd */
9326                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9327                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9328                         }
9329                 }
9330         }
9331 }
9332 #endif
9333
9334 static int
9335 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9336 {
9337
9338         uint32_t am, l_end;
9339         int was_tlp = 0;
9340
9341         if (SEQ_GT(end, start))
9342                 am = end - start;
9343         else
9344                 am = 0;
9345         if ((rack->rc_last_tlp_acked_set ) &&
9346             (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9347             (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9348                 /*
9349                  * The DSACK is because of a TLP which we don't
9350                  * do anything with the reordering window over since
9351                  * it was not reordering that caused the DSACK but
9352                  * our previous retransmit TLP.
9353                  */
9354                 rack_log_dsack_event(rack, 7, __LINE__, start, end);
9355                 was_tlp = 1;
9356                 goto skip_dsack_round;
9357         }
9358         if (rack->rc_last_sent_tlp_seq_valid) {
9359                 l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9360                 if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9361                     (SEQ_LEQ(end, l_end))) {
9362                         /*
9363                          * This dsack is from the last sent TLP, ignore it
9364                          * for reordering purposes.
9365                          */
9366                         rack_log_dsack_event(rack, 7, __LINE__, start, end);
9367                         was_tlp = 1;
9368                         goto skip_dsack_round;
9369                 }
9370         }
9371         if (rack->rc_dsack_round_seen == 0) {
9372                 rack->rc_dsack_round_seen = 1;
9373                 rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9374                 rack->r_ctl.num_dsack++;
9375                 rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */
9376                 rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9377         }
9378 skip_dsack_round:
9379         /*
9380          * We keep track of how many DSACK blocks we get
9381          * after a recovery incident.
9382          */
9383         rack->r_ctl.dsack_byte_cnt += am;
9384         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9385             rack->r_ctl.retran_during_recovery &&
9386             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9387                 /*
9388                  * False recovery most likely culprit is reordering. If
9389                  * nothing else is missing we need to revert.
9390                  */
9391                 rack->r_might_revert = 1;
9392                 rack_handle_might_revert(rack->rc_tp, rack);
9393                 rack->r_might_revert = 0;
9394                 rack->r_ctl.retran_during_recovery = 0;
9395                 rack->r_ctl.dsack_byte_cnt = 0;
9396         }
9397         return (was_tlp);
9398 }
9399
9400 static uint32_t
9401 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
9402 {
9403         return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
9404 }
9405
9406 static int32_t
9407 rack_compute_pipe(struct tcpcb *tp)
9408 {
9409         return ((int32_t)do_rack_compute_pipe(tp,
9410                                               (struct tcp_rack *)tp->t_fb_ptr,
9411                                               tp->snd_una));
9412 }
9413
9414 static void
9415 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9416 {
9417         /* Deal with changed and PRR here (in recovery only) */
9418         uint32_t pipe, snd_una;
9419
9420         rack->r_ctl.rc_prr_delivered += changed;
9421
9422         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9423                 /*
9424                  * It is all outstanding, we are application limited
9425                  * and thus we don't need more room to send anything.
9426                  * Note we use tp->snd_una here and not th_ack because
9427                  * the data as yet not been cut from the sb.
9428                  */
9429                 rack->r_ctl.rc_prr_sndcnt = 0;
9430                 return;
9431         }
9432         /* Compute prr_sndcnt */
9433         if (SEQ_GT(tp->snd_una, th_ack)) {
9434                 snd_una = tp->snd_una;
9435         } else {
9436                 snd_una = th_ack;
9437         }
9438         pipe = do_rack_compute_pipe(tp, rack, snd_una);
9439         if (pipe > tp->snd_ssthresh) {
9440                 long sndcnt;
9441
9442                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9443                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9444                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9445                 else {
9446                         rack->r_ctl.rc_prr_sndcnt = 0;
9447                         rack_log_to_prr(rack, 9, 0, __LINE__);
9448                         sndcnt = 0;
9449                 }
9450                 sndcnt++;
9451                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9452                         sndcnt -= rack->r_ctl.rc_prr_out;
9453                 else
9454                         sndcnt = 0;
9455                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9456                 rack_log_to_prr(rack, 10, 0, __LINE__);
9457         } else {
9458                 uint32_t limit;
9459
9460                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9461                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9462                 else
9463                         limit = 0;
9464                 if (changed > limit)
9465                         limit = changed;
9466                 limit += ctf_fixed_maxseg(tp);
9467                 if (tp->snd_ssthresh > pipe) {
9468                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9469                         rack_log_to_prr(rack, 11, 0, __LINE__);
9470                 } else {
9471                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9472                         rack_log_to_prr(rack, 12, 0, __LINE__);
9473                 }
9474         }
9475 }
9476
9477 static void
9478 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9479 {
9480         uint32_t changed;
9481         struct tcp_rack *rack;
9482         struct rack_sendmap *rsm;
9483         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9484         register uint32_t th_ack;
9485         int32_t i, j, k, num_sack_blks = 0;
9486         uint32_t cts, acked, ack_point;
9487         int loop_start = 0, moved_two = 0;
9488         uint32_t tsused;
9489
9490
9491         INP_WLOCK_ASSERT(tp->t_inpcb);
9492         if (tcp_get_flags(th) & TH_RST) {
9493                 /* We don't log resets */
9494                 return;
9495         }
9496         rack = (struct tcp_rack *)tp->t_fb_ptr;
9497         cts = tcp_get_usecs(NULL);
9498         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9499         changed = 0;
9500         th_ack = th->th_ack;
9501         if (rack->sack_attack_disable == 0)
9502                 rack_do_decay(rack);
9503         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9504                 /*
9505                  * You only get credit for
9506                  * MSS and greater (and you get extra
9507                  * credit for larger cum-ack moves).
9508                  */
9509                 int ac;
9510
9511                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9512                 rack->r_ctl.ack_count += ac;
9513                 counter_u64_add(rack_ack_total, ac);
9514         }
9515         if (rack->r_ctl.ack_count > 0xfff00000) {
9516                 /*
9517                  * reduce the number to keep us under
9518                  * a uint32_t.
9519                  */
9520                 rack->r_ctl.ack_count /= 2;
9521                 rack->r_ctl.sack_count /= 2;
9522         }
9523         if (SEQ_GT(th_ack, tp->snd_una)) {
9524                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9525                 tp->t_acktime = ticks;
9526         }
9527         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9528                 changed = th_ack - rsm->r_start;
9529         if (changed) {
9530                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9531         }
9532         if ((to->to_flags & TOF_SACK) == 0) {
9533                 /* We are done nothing left and no sack. */
9534                 rack_handle_might_revert(tp, rack);
9535                 /*
9536                  * For cases where we struck a dup-ack
9537                  * with no SACK, add to the changes so
9538                  * PRR will work right.
9539                  */
9540                 if (dup_ack_struck && (changed == 0)) {
9541                         changed += ctf_fixed_maxseg(rack->rc_tp);
9542                 }
9543                 goto out;
9544         }
9545         /* Sack block processing */
9546         if (SEQ_GT(th_ack, tp->snd_una))
9547                 ack_point = th_ack;
9548         else
9549                 ack_point = tp->snd_una;
9550         for (i = 0; i < to->to_nsacks; i++) {
9551                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9552                       &sack, sizeof(sack));
9553                 sack.start = ntohl(sack.start);
9554                 sack.end = ntohl(sack.end);
9555                 if (SEQ_GT(sack.end, sack.start) &&
9556                     SEQ_GT(sack.start, ack_point) &&
9557                     SEQ_LT(sack.start, tp->snd_max) &&
9558                     SEQ_GT(sack.end, ack_point) &&
9559                     SEQ_LEQ(sack.end, tp->snd_max)) {
9560                         sack_blocks[num_sack_blks] = sack;
9561                         num_sack_blks++;
9562                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9563                            SEQ_LEQ(sack.end, th_ack)) {
9564                         int was_tlp;
9565
9566                         was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9567                         /*
9568                          * Its a D-SACK block.
9569                          */
9570                         tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9571                 }
9572         }
9573         if (rack->rc_dsack_round_seen) {
9574                 /* Is the dsack roound over? */
9575                 if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9576                         /* Yes it is */
9577                         rack->rc_dsack_round_seen = 0;
9578                         rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9579                 }
9580         }
9581         /*
9582          * Sort the SACK blocks so we can update the rack scoreboard with
9583          * just one pass.
9584          */
9585         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9586                                          num_sack_blks, th->th_ack);
9587         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9588         if (num_sack_blks == 0) {
9589                 /* Nothing to sack (DSACKs?) */
9590                 goto out_with_totals;
9591         }
9592         if (num_sack_blks < 2) {
9593                 /* Only one, we don't need to sort */
9594                 goto do_sack_work;
9595         }
9596         /* Sort the sacks */
9597         for (i = 0; i < num_sack_blks; i++) {
9598                 for (j = i + 1; j < num_sack_blks; j++) {
9599                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9600                                 sack = sack_blocks[i];
9601                                 sack_blocks[i] = sack_blocks[j];
9602                                 sack_blocks[j] = sack;
9603                         }
9604                 }
9605         }
9606         /*
9607          * Now are any of the sack block ends the same (yes some
9608          * implementations send these)?
9609          */
9610 again:
9611         if (num_sack_blks == 0)
9612                 goto out_with_totals;
9613         if (num_sack_blks > 1) {
9614                 for (i = 0; i < num_sack_blks; i++) {
9615                         for (j = i + 1; j < num_sack_blks; j++) {
9616                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9617                                         /*
9618                                          * Ok these two have the same end we
9619                                          * want the smallest end and then
9620                                          * throw away the larger and start
9621                                          * again.
9622                                          */
9623                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9624                                                 /*
9625                                                  * The second block covers
9626                                                  * more area use that
9627                                                  */
9628                                                 sack_blocks[i].start = sack_blocks[j].start;
9629                                         }
9630                                         /*
9631                                          * Now collapse out the dup-sack and
9632                                          * lower the count
9633                                          */
9634                                         for (k = (j + 1); k < num_sack_blks; k++) {
9635                                                 sack_blocks[j].start = sack_blocks[k].start;
9636                                                 sack_blocks[j].end = sack_blocks[k].end;
9637                                                 j++;
9638                                         }
9639                                         num_sack_blks--;
9640                                         goto again;
9641                                 }
9642                         }
9643                 }
9644         }
9645 do_sack_work:
9646         /*
9647          * First lets look to see if
9648          * we have retransmitted and
9649          * can use the transmit next?
9650          */
9651         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9652         if (rsm &&
9653             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9654             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9655                 /*
9656                  * We probably did the FR and the next
9657                  * SACK in continues as we would expect.
9658                  */
9659                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9660                 if (acked) {
9661                         rack->r_wanted_output = 1;
9662                         changed += acked;
9663                 }
9664                 if (num_sack_blks == 1) {
9665                         /*
9666                          * This is what we would expect from
9667                          * a normal implementation to happen
9668                          * after we have retransmitted the FR,
9669                          * i.e the sack-filter pushes down
9670                          * to 1 block and the next to be retransmitted
9671                          * is the sequence in the sack block (has more
9672                          * are acked). Count this as ACK'd data to boost
9673                          * up the chances of recovering any false positives.
9674                          */
9675                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9676                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9677                         counter_u64_add(rack_express_sack, 1);
9678                         if (rack->r_ctl.ack_count > 0xfff00000) {
9679                                 /*
9680                                  * reduce the number to keep us under
9681                                  * a uint32_t.
9682                                  */
9683                                 rack->r_ctl.ack_count /= 2;
9684                                 rack->r_ctl.sack_count /= 2;
9685                         }
9686                         goto out_with_totals;
9687                 } else {
9688                         /*
9689                          * Start the loop through the
9690                          * rest of blocks, past the first block.
9691                          */
9692                         moved_two = 0;
9693                         loop_start = 1;
9694                 }
9695         }
9696         /* Its a sack of some sort */
9697         rack->r_ctl.sack_count++;
9698         if (rack->r_ctl.sack_count > 0xfff00000) {
9699                 /*
9700                  * reduce the number to keep us under
9701                  * a uint32_t.
9702                  */
9703                 rack->r_ctl.ack_count /= 2;
9704                 rack->r_ctl.sack_count /= 2;
9705         }
9706         counter_u64_add(rack_sack_total, 1);
9707         if (rack->sack_attack_disable) {
9708                 /* An attacker disablement is in place */
9709                 if (num_sack_blks > 1) {
9710                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9711                         rack->r_ctl.sack_moved_extra++;
9712                         counter_u64_add(rack_move_some, 1);
9713                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9714                                 rack->r_ctl.sack_moved_extra /= 2;
9715                                 rack->r_ctl.sack_noextra_move /= 2;
9716                         }
9717                 }
9718                 goto out;
9719         }
9720         rsm = rack->r_ctl.rc_sacklast;
9721         for (i = loop_start; i < num_sack_blks; i++) {
9722                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9723                 if (acked) {
9724                         rack->r_wanted_output = 1;
9725                         changed += acked;
9726                 }
9727                 if (moved_two) {
9728                         /*
9729                          * If we did not get a SACK for at least a MSS and
9730                          * had to move at all, or if we moved more than our
9731                          * threshold, it counts against the "extra" move.
9732                          */
9733                         rack->r_ctl.sack_moved_extra += moved_two;
9734                         counter_u64_add(rack_move_some, 1);
9735                 } else {
9736                         /*
9737                          * else we did not have to move
9738                          * any more than we would expect.
9739                          */
9740                         rack->r_ctl.sack_noextra_move++;
9741                         counter_u64_add(rack_move_none, 1);
9742                 }
9743                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9744                         /*
9745                          * If the SACK was not a full MSS then
9746                          * we add to sack_count the number of
9747                          * MSS's (or possibly more than
9748                          * a MSS if its a TSO send) we had to skip by.
9749                          */
9750                         rack->r_ctl.sack_count += moved_two;
9751                         counter_u64_add(rack_sack_total, moved_two);
9752                 }
9753                 /*
9754                  * Now we need to setup for the next
9755                  * round. First we make sure we won't
9756                  * exceed the size of our uint32_t on
9757                  * the various counts, and then clear out
9758                  * moved_two.
9759                  */
9760                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9761                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9762                         rack->r_ctl.sack_moved_extra /= 2;
9763                         rack->r_ctl.sack_noextra_move /= 2;
9764                 }
9765                 if (rack->r_ctl.sack_count > 0xfff00000) {
9766                         rack->r_ctl.ack_count /= 2;
9767                         rack->r_ctl.sack_count /= 2;
9768                 }
9769                 moved_two = 0;
9770         }
9771 out_with_totals:
9772         if (num_sack_blks > 1) {
9773                 /*
9774                  * You get an extra stroke if
9775                  * you have more than one sack-blk, this
9776                  * could be where we are skipping forward
9777                  * and the sack-filter is still working, or
9778                  * it could be an attacker constantly
9779                  * moving us.
9780                  */
9781                 rack->r_ctl.sack_moved_extra++;
9782                 counter_u64_add(rack_move_some, 1);
9783         }
9784 out:
9785 #ifdef NETFLIX_EXP_DETECTION
9786         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9787 #endif
9788         if (changed) {
9789                 /* Something changed cancel the rack timer */
9790                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9791         }
9792         tsused = tcp_get_usecs(NULL);
9793         rsm = tcp_rack_output(tp, rack, tsused);
9794         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9795             rsm &&
9796             ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
9797                 /* Enter recovery */
9798                 entered_recovery = 1;
9799                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9800                 /*
9801                  * When we enter recovery we need to assure we send
9802                  * one packet.
9803                  */
9804                 if (rack->rack_no_prr == 0) {
9805                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9806                         rack_log_to_prr(rack, 8, 0, __LINE__);
9807                 }
9808                 rack->r_timer_override = 1;
9809                 rack->r_early = 0;
9810                 rack->r_ctl.rc_agg_early = 0;
9811         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9812                    rsm &&
9813                    (rack->r_rr_config == 3)) {
9814                 /*
9815                  * Assure we can output and we get no
9816                  * remembered pace time except the retransmit.
9817                  */
9818                 rack->r_timer_override = 1;
9819                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9820                 rack->r_ctl.rc_resend = rsm;
9821         }
9822         if (IN_FASTRECOVERY(tp->t_flags) &&
9823             (rack->rack_no_prr == 0) &&
9824             (entered_recovery == 0)) {
9825                 rack_update_prr(tp, rack, changed, th_ack);
9826                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9827                      ((tcp_in_hpts(rack->rc_inp) == 0) &&
9828                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9829                         /*
9830                          * If you are pacing output you don't want
9831                          * to override.
9832                          */
9833                         rack->r_early = 0;
9834                         rack->r_ctl.rc_agg_early = 0;
9835                         rack->r_timer_override = 1;
9836                 }
9837         }
9838 }
9839
9840 static void
9841 rack_strike_dupack(struct tcp_rack *rack)
9842 {
9843         struct rack_sendmap *rsm;
9844
9845         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9846         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9847                 rsm = TAILQ_NEXT(rsm, r_tnext);
9848                 if (rsm->r_flags & RACK_MUST_RXT) {
9849                         /* Sendmap entries that are marked to
9850                          * be retransmitted do not need dupack's
9851                          * struck. We get these marks for a number
9852                          * of reasons (rxt timeout with no sack,
9853                          * mtu change, or rwnd collapses). When
9854                          * these events occur, we know we must retransmit
9855                          * them and mark the sendmap entries. Dupack counting
9856                          * is not needed since we are already set to retransmit
9857                          * it as soon as we can.
9858                          */
9859                         continue;
9860                 }
9861         }
9862         if (rsm && (rsm->r_dupack < 0xff)) {
9863                 rsm->r_dupack++;
9864                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9865                         struct timeval tv;
9866                         uint32_t cts;
9867                         /*
9868                          * Here we see if we need to retransmit. For
9869                          * a SACK type connection if enough time has passed
9870                          * we will get a return of the rsm. For a non-sack
9871                          * connection we will get the rsm returned if the
9872                          * dupack value is 3 or more.
9873                          */
9874                         cts = tcp_get_usecs(&tv);
9875                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9876                         if (rack->r_ctl.rc_resend != NULL) {
9877                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9878                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9879                                                          rack->rc_tp->snd_una, __LINE__);
9880                                 }
9881                                 rack->r_wanted_output = 1;
9882                                 rack->r_timer_override = 1;
9883                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9884                         }
9885                 } else {
9886                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9887                 }
9888         }
9889 }
9890
9891 static void
9892 rack_check_bottom_drag(struct tcpcb *tp,
9893                        struct tcp_rack *rack,
9894                        struct socket *so, int32_t acked)
9895 {
9896         uint32_t segsiz, minseg;
9897
9898         segsiz = ctf_fixed_maxseg(tp);
9899         minseg = segsiz;
9900
9901         if (tp->snd_max == tp->snd_una) {
9902                 /*
9903                  * We are doing dynamic pacing and we are way
9904                  * under. Basically everything got acked while
9905                  * we were still waiting on the pacer to expire.
9906                  *
9907                  * This means we need to boost the b/w in
9908                  * addition to any earlier boosting of
9909                  * the multiplier.
9910                  */
9911                 rack->rc_dragged_bottom = 1;
9912                 rack_validate_multipliers_at_or_above100(rack);
9913                 /*
9914                  * Lets use the segment bytes acked plus
9915                  * the lowest RTT seen as the basis to
9916                  * form a b/w estimate. This will be off
9917                  * due to the fact that the true estimate
9918                  * should be around 1/2 the time of the RTT
9919                  * but we can settle for that.
9920                  */
9921                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9922                     acked) {
9923                         uint64_t bw, calc_bw, rtt;
9924
9925                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9926                         if (rtt == 0) {
9927                                 /* no us sample is there a ms one? */
9928                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9929                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9930                                 } else {
9931                                         goto no_measurement;
9932                                 }
9933                         }
9934                         bw = acked;
9935                         calc_bw = bw * 1000000;
9936                         calc_bw /= rtt;
9937                         if (rack->r_ctl.last_max_bw &&
9938                             (rack->r_ctl.last_max_bw < calc_bw)) {
9939                                 /*
9940                                  * If we have a last calculated max bw
9941                                  * enforce it.
9942                                  */
9943                                 calc_bw = rack->r_ctl.last_max_bw;
9944                         }
9945                         /* now plop it in */
9946                         if (rack->rc_gp_filled == 0) {
9947                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9948                                         /*
9949                                          * If we have no measurement
9950                                          * don't let us set in more than
9951                                          * 1.2Mbps. If we are still too
9952                                          * low after pacing with this we
9953                                          * will hopefully have a max b/w
9954                                          * available to sanity check things.
9955                                          */
9956                                         calc_bw = ONE_POINT_TWO_MEG;
9957                                 }
9958                                 rack->r_ctl.rc_rtt_diff = 0;
9959                                 rack->r_ctl.gp_bw = calc_bw;
9960                                 rack->rc_gp_filled = 1;
9961                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9962                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9963                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9964                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9965                                 rack->r_ctl.rc_rtt_diff = 0;
9966                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9967                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9968                                 rack->r_ctl.gp_bw = calc_bw;
9969                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9970                         } else
9971                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9972                         if ((rack->gp_ready == 0) &&
9973                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9974                                 /* We have enough measurements now */
9975                                 rack->gp_ready = 1;
9976                                 rack_set_cc_pacing(rack);
9977                                 if (rack->defer_options)
9978                                         rack_apply_deferred_options(rack);
9979                         }
9980                         /*
9981                          * For acks over 1mss we do a extra boost to simulate
9982                          * where we would get 2 acks (we want 110 for the mul).
9983                          */
9984                         if (acked > segsiz)
9985                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9986                 } else {
9987                         /*
9988                          * zero rtt possibly?, settle for just an old increase.
9989                          */
9990 no_measurement:
9991                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9992                 }
9993         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9994                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9995                                                minseg)) &&
9996                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9997                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9998                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9999                     (segsiz * rack_req_segs))) {
10000                 /*
10001                  * We are doing dynamic GP pacing and
10002                  * we have everything except 1MSS or less
10003                  * bytes left out. We are still pacing away.
10004                  * And there is data that could be sent, This
10005                  * means we are inserting delayed ack time in
10006                  * our measurements because we are pacing too slow.
10007                  */
10008                 rack_validate_multipliers_at_or_above100(rack);
10009                 rack->rc_dragged_bottom = 1;
10010                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
10011         }
10012 }
10013
10014
10015
10016 static void
10017 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
10018 {
10019         /*
10020          * The fast output path is enabled and we
10021          * have moved the cumack forward. Lets see if
10022          * we can expand forward the fast path length by
10023          * that amount. What we would ideally like to
10024          * do is increase the number of bytes in the
10025          * fast path block (left_to_send) by the
10026          * acked amount. However we have to gate that
10027          * by two factors:
10028          * 1) The amount outstanding and the rwnd of the peer
10029          *    (i.e. we don't want to exceed the rwnd of the peer).
10030          *    <and>
10031          * 2) The amount of data left in the socket buffer (i.e.
10032          *    we can't send beyond what is in the buffer).
10033          *
10034          * Note that this does not take into account any increase
10035          * in the cwnd. We will only extend the fast path by
10036          * what was acked.
10037          */
10038         uint32_t new_total, gating_val;
10039
10040         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
10041         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
10042                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
10043         if (new_total <= gating_val) {
10044                 /* We can increase left_to_send by the acked amount */
10045                 counter_u64_add(rack_extended_rfo, 1);
10046                 rack->r_ctl.fsb.left_to_send = new_total;
10047                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
10048                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
10049                          rack, rack->r_ctl.fsb.left_to_send,
10050                          sbavail(&rack->rc_inp->inp_socket->so_snd),
10051                          (tp->snd_max - tp->snd_una)));
10052
10053         }
10054 }
10055
10056 static void
10057 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
10058 {
10059         /*
10060          * Here any sendmap entry that points to the
10061          * beginning mbuf must be adjusted to the correct
10062          * offset. This must be called with:
10063          * 1) The socket buffer locked
10064          * 2) snd_una adjusted to its new postion.
10065          *
10066          * Note that (2) implies rack_ack_received has also
10067          * been called.
10068          *
10069          * We grab the first mbuf in the socket buffer and
10070          * then go through the front of the sendmap, recalculating
10071          * the stored offset for any sendmap entry that has
10072          * that mbuf. We must use the sb functions to do this
10073          * since its possible an add was done has well as
10074          * the subtraction we may have just completed. This should
10075          * not be a penalty though, since we just referenced the sb
10076          * to go in and trim off the mbufs that we freed (of course
10077          * there will be a penalty for the sendmap references though).
10078          */
10079         struct mbuf *m;
10080         struct rack_sendmap *rsm;
10081
10082         SOCKBUF_LOCK_ASSERT(sb);
10083         m = sb->sb_mb;
10084         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
10085         if ((rsm == NULL) || (m == NULL)) {
10086                 /* Nothing outstanding */
10087                 return;
10088         }
10089         while (rsm->m && (rsm->m == m)) {
10090                 /* one to adjust */
10091 #ifdef INVARIANTS
10092                 struct mbuf *tm;
10093                 uint32_t soff;
10094
10095                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
10096                 if (rsm->orig_m_len != m->m_len) {
10097                         rack_adjust_orig_mlen(rsm);
10098                 }
10099                 if (rsm->soff != soff) {
10100                         /*
10101                          * This is not a fatal error, we anticipate it
10102                          * might happen (the else code), so we count it here
10103                          * so that under invariant we can see that it really
10104                          * does happen.
10105                          */
10106                         counter_u64_add(rack_adjust_map_bw, 1);
10107                 }
10108                 rsm->m = tm;
10109                 rsm->soff = soff;
10110                 if (tm)
10111                         rsm->orig_m_len = rsm->m->m_len;
10112                 else
10113                         rsm->orig_m_len = 0;
10114 #else
10115                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
10116                 if (rsm->m)
10117                         rsm->orig_m_len = rsm->m->m_len;
10118                 else
10119                         rsm->orig_m_len = 0;
10120 #endif
10121                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
10122                               rsm);
10123                 if (rsm == NULL)
10124                         break;
10125         }
10126 }
10127
10128 /*
10129  * Return value of 1, we do not need to call rack_process_data().
10130  * return value of 0, rack_process_data can be called.
10131  * For ret_val if its 0 the TCP is locked, if its non-zero
10132  * its unlocked and probably unsafe to touch the TCB.
10133  */
10134 static int
10135 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10136     struct tcpcb *tp, struct tcpopt *to,
10137     uint32_t tiwin, int32_t tlen,
10138     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10139 {
10140         int32_t ourfinisacked = 0;
10141         int32_t nsegs, acked_amount;
10142         int32_t acked;
10143         struct mbuf *mfree;
10144         struct tcp_rack *rack;
10145         int32_t under_pacing = 0;
10146         int32_t recovery = 0;
10147
10148         rack = (struct tcp_rack *)tp->t_fb_ptr;
10149         if (SEQ_GT(th->th_ack, tp->snd_max)) {
10150                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10151                                       &rack->r_ctl.challenge_ack_ts,
10152                                       &rack->r_ctl.challenge_ack_cnt);
10153                 rack->r_wanted_output = 1;
10154                 return (1);
10155         }
10156         if (rack->gp_ready &&
10157             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10158                 under_pacing = 1;
10159         }
10160         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10161                 int in_rec, dup_ack_struck = 0;
10162
10163                 in_rec = IN_FASTRECOVERY(tp->t_flags);
10164                 if (rack->rc_in_persist) {
10165                         tp->t_rxtshift = 0;
10166                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10167                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10168                 }
10169                 if ((th->th_ack == tp->snd_una) &&
10170                     (tiwin == tp->snd_wnd) &&
10171                     ((to->to_flags & TOF_SACK) == 0)) {
10172                         rack_strike_dupack(rack);
10173                         dup_ack_struck = 1;
10174                 }
10175                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10176         }
10177         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10178                 /*
10179                  * Old ack, behind (or duplicate to) the last one rcv'd
10180                  * Note: We mark reordering is occuring if its
10181                  * less than and we have not closed our window.
10182                  */
10183                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10184                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10185                 }
10186                 return (0);
10187         }
10188         /*
10189          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10190          * something we sent.
10191          */
10192         if (tp->t_flags & TF_NEEDSYN) {
10193                 /*
10194                  * T/TCP: Connection was half-synchronized, and our SYN has
10195                  * been ACK'd (so connection is now fully synchronized).  Go
10196                  * to non-starred state, increment snd_una for ACK of SYN,
10197                  * and check if we can do window scaling.
10198                  */
10199                 tp->t_flags &= ~TF_NEEDSYN;
10200                 tp->snd_una++;
10201                 /* Do window scaling? */
10202                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10203                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10204                         tp->rcv_scale = tp->request_r_scale;
10205                         /* Send window already scaled. */
10206                 }
10207         }
10208         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10209         INP_WLOCK_ASSERT(tp->t_inpcb);
10210
10211         acked = BYTES_THIS_ACK(tp, th);
10212         if (acked) {
10213                 /*
10214                  * Any time we move the cum-ack forward clear
10215                  * keep-alive tied probe-not-answered. The
10216                  * persists clears its own on entry.
10217                  */
10218                 rack->probe_not_answered = 0;
10219         }
10220         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10221         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10222         /*
10223          * If we just performed our first retransmit, and the ACK arrives
10224          * within our recovery window, then it was a mistake to do the
10225          * retransmit in the first place.  Recover our original cwnd and
10226          * ssthresh, and proceed to transmit where we left off.
10227          */
10228         if ((tp->t_flags & TF_PREVVALID) &&
10229             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10230                 tp->t_flags &= ~TF_PREVVALID;
10231                 if (tp->t_rxtshift == 1 &&
10232                     (int)(ticks - tp->t_badrxtwin) < 0)
10233                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10234         }
10235         if (acked) {
10236                 /* assure we are not backed off */
10237                 tp->t_rxtshift = 0;
10238                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10239                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10240                 rack->rc_tlp_in_progress = 0;
10241                 rack->r_ctl.rc_tlp_cnt_out = 0;
10242                 /*
10243                  * If it is the RXT timer we want to
10244                  * stop it, so we can restart a TLP.
10245                  */
10246                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10247                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10248 #ifdef NETFLIX_HTTP_LOGGING
10249                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10250 #endif
10251         }
10252         /*
10253          * If we have a timestamp reply, update smoothed round trip time. If
10254          * no timestamp is present but transmit timer is running and timed
10255          * sequence number was acked, update smoothed round trip time. Since
10256          * we now have an rtt measurement, cancel the timer backoff (cf.,
10257          * Phil Karn's retransmit alg.). Recompute the initial retransmit
10258          * timer.
10259          *
10260          * Some boxes send broken timestamp replies during the SYN+ACK
10261          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10262          * and blow up the retransmit timer.
10263          */
10264         /*
10265          * If all outstanding data is acked, stop retransmit timer and
10266          * remember to restart (more output or persist). If there is more
10267          * data to be acked, restart retransmit timer, using current
10268          * (possibly backed-off) value.
10269          */
10270         if (acked == 0) {
10271                 if (ofia)
10272                         *ofia = ourfinisacked;
10273                 return (0);
10274         }
10275         if (IN_RECOVERY(tp->t_flags)) {
10276                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10277                     (SEQ_LT(th->th_ack, tp->snd_max))) {
10278                         tcp_rack_partialack(tp);
10279                 } else {
10280                         rack_post_recovery(tp, th->th_ack);
10281                         recovery = 1;
10282                 }
10283         }
10284         /*
10285          * Let the congestion control algorithm update congestion control
10286          * related information. This typically means increasing the
10287          * congestion window.
10288          */
10289         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10290         SOCKBUF_LOCK(&so->so_snd);
10291         acked_amount = min(acked, (int)sbavail(&so->so_snd));
10292         tp->snd_wnd -= acked_amount;
10293         mfree = sbcut_locked(&so->so_snd, acked_amount);
10294         if ((sbused(&so->so_snd) == 0) &&
10295             (acked > acked_amount) &&
10296             (tp->t_state >= TCPS_FIN_WAIT_1) &&
10297             (tp->t_flags & TF_SENTFIN)) {
10298                 /*
10299                  * We must be sure our fin
10300                  * was sent and acked (we can be
10301                  * in FIN_WAIT_1 without having
10302                  * sent the fin).
10303                  */
10304                 ourfinisacked = 1;
10305         }
10306         tp->snd_una = th->th_ack;
10307         if (acked_amount && sbavail(&so->so_snd))
10308                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10309         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10310         /* NB: sowwakeup_locked() does an implicit unlock. */
10311         sowwakeup_locked(so);
10312         m_freem(mfree);
10313         if (SEQ_GT(tp->snd_una, tp->snd_recover))
10314                 tp->snd_recover = tp->snd_una;
10315
10316         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10317                 tp->snd_nxt = tp->snd_una;
10318         }
10319         if (under_pacing &&
10320             (rack->use_fixed_rate == 0) &&
10321             (rack->in_probe_rtt == 0) &&
10322             rack->rc_gp_dyn_mul &&
10323             rack->rc_always_pace) {
10324                 /* Check if we are dragging bottom */
10325                 rack_check_bottom_drag(tp, rack, so, acked);
10326         }
10327         if (tp->snd_una == tp->snd_max) {
10328                 /* Nothing left outstanding */
10329                 tp->t_flags &= ~TF_PREVVALID;
10330                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10331                 rack->r_ctl.retran_during_recovery = 0;
10332                 rack->r_ctl.dsack_byte_cnt = 0;
10333                 if (rack->r_ctl.rc_went_idle_time == 0)
10334                         rack->r_ctl.rc_went_idle_time = 1;
10335                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10336                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10337                         tp->t_acktime = 0;
10338                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10339                 /* Set need output so persist might get set */
10340                 rack->r_wanted_output = 1;
10341                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10342                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10343                     (sbavail(&so->so_snd) == 0) &&
10344                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10345                         /*
10346                          * The socket was gone and the
10347                          * peer sent data (now or in the past), time to
10348                          * reset him.
10349                          */
10350                         *ret_val = 1;
10351                         /* tcp_close will kill the inp pre-log the Reset */
10352                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10353                         tp = tcp_close(tp);
10354                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10355                         return (1);
10356                 }
10357         }
10358         if (ofia)
10359                 *ofia = ourfinisacked;
10360         return (0);
10361 }
10362
10363
10364 static void
10365 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
10366                   int dir, uint32_t flags, struct rack_sendmap *rsm)
10367 {
10368         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
10369                 union tcp_log_stackspecific log;
10370                 struct timeval tv;
10371
10372                 memset(&log, 0, sizeof(log));
10373                 log.u_bbr.flex1 = cnt;
10374                 log.u_bbr.flex2 = split;
10375                 log.u_bbr.flex3 = out;
10376                 log.u_bbr.flex4 = line;
10377                 log.u_bbr.flex5 = rack->r_must_retran;
10378                 log.u_bbr.flex6 = flags;
10379                 log.u_bbr.flex7 = rack->rc_has_collapsed;
10380                 log.u_bbr.flex8 = dir;  /*
10381                                          * 1 is collapsed, 0 is uncollapsed,
10382                                          * 2 is log of a rsm being marked, 3 is a split.
10383                                          */
10384                 if (rsm == NULL)
10385                         log.u_bbr.rttProp = 0;
10386                 else
10387                         log.u_bbr.rttProp = (uint64_t)rsm;
10388                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
10389                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10390                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
10391                     &rack->rc_inp->inp_socket->so_rcv,
10392                     &rack->rc_inp->inp_socket->so_snd,
10393                     TCP_RACK_LOG_COLLAPSE, 0,
10394                     0, &log, false, &tv);
10395         }
10396 }
10397
10398 static void
10399 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, int line)
10400 {
10401         /*
10402          * Here all we do is mark the collapsed point and set the flag.
10403          * This may happen again and again, but there is no
10404          * sense splitting our map until we know where the
10405          * peer finally lands in the collapse.
10406          */
10407         rack_trace_point(rack, RACK_TP_COLLAPSED_WND);
10408         if ((rack->rc_has_collapsed == 0) ||
10409             (rack->r_ctl.last_collapse_point != (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)))
10410                 counter_u64_add(rack_collapsed_win_seen, 1);
10411         rack->r_ctl.last_collapse_point = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10412         rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
10413         rack->rc_has_collapsed = 1;
10414         rack->r_collapse_point_valid = 1;
10415         rack_log_collapse(rack, 0, 0, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
10416 }
10417
10418 static void
10419 rack_un_collapse_window(struct tcp_rack *rack, int line)
10420 {
10421         struct rack_sendmap *nrsm, *rsm, fe;
10422         int cnt = 0, split = 0;
10423 #ifdef INVARIANTS
10424         struct rack_sendmap *insret;
10425 #endif
10426
10427         memset(&fe, 0, sizeof(fe));
10428         rack->rc_has_collapsed = 0;
10429         fe.r_start = rack->r_ctl.last_collapse_point;
10430         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10431         if (rsm == NULL) {
10432                 /* Nothing to do maybe the peer ack'ed it all */
10433                 rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10434                 return;
10435         }
10436         /* Now do we need to split this one? */
10437         if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
10438                 rack_log_collapse(rack, rsm->r_start, rsm->r_end,
10439                                   rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
10440                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10441                 if (nrsm == NULL) {
10442                         /* We can't get a rsm, mark all? */
10443                         nrsm = rsm;
10444                         goto no_split;
10445                 }
10446                 /* Clone it */
10447                 split = 1;
10448                 rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
10449 #ifndef INVARIANTS
10450                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10451 #else
10452                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10453                 if (insret != NULL) {
10454                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10455                               nrsm, insret, rack, rsm);
10456                 }
10457 #endif
10458                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
10459                                  rack->r_ctl.last_collapse_point, __LINE__);
10460                 if (rsm->r_in_tmap) {
10461                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10462                         nrsm->r_in_tmap = 1;
10463                 }
10464                 /*
10465                  * Set in the new RSM as the
10466                  * collapsed starting point
10467                  */
10468                 rsm = nrsm;
10469         }
10470 no_split:
10471         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10472                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
10473                 rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
10474                 cnt++;
10475         }
10476         if (cnt) {
10477                 counter_u64_add(rack_collapsed_win, 1);
10478         }
10479         rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10480 }
10481
10482 static void
10483 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10484                         int32_t tlen, int32_t tfo_syn)
10485 {
10486         if (DELAY_ACK(tp, tlen) || tfo_syn) {
10487                 if (rack->rc_dack_mode &&
10488                     (tlen > 500) &&
10489                     (rack->rc_dack_toggle == 1)) {
10490                         goto no_delayed_ack;
10491                 }
10492                 rack_timer_cancel(tp, rack,
10493                                   rack->r_ctl.rc_rcvtime, __LINE__);
10494                 tp->t_flags |= TF_DELACK;
10495         } else {
10496 no_delayed_ack:
10497                 rack->r_wanted_output = 1;
10498                 tp->t_flags |= TF_ACKNOW;
10499                 if (rack->rc_dack_mode) {
10500                         if (tp->t_flags & TF_DELACK)
10501                                 rack->rc_dack_toggle = 1;
10502                         else
10503                                 rack->rc_dack_toggle = 0;
10504                 }
10505         }
10506 }
10507
10508 static void
10509 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10510 {
10511         /*
10512          * If fast output is in progress, lets validate that
10513          * the new window did not shrink on us and make it
10514          * so fast output should end.
10515          */
10516         if (rack->r_fast_output) {
10517                 uint32_t out;
10518
10519                 /*
10520                  * Calculate what we will send if left as is
10521                  * and compare that to our send window.
10522                  */
10523                 out = ctf_outstanding(tp);
10524                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10525                         /* ok we have an issue */
10526                         if (out >= tp->snd_wnd) {
10527                                 /* Turn off fast output the window is met or collapsed */
10528                                 rack->r_fast_output = 0;
10529                         } else {
10530                                 /* we have some room left */
10531                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10532                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10533                                         /* If not at least 1 full segment never mind */
10534                                         rack->r_fast_output = 0;
10535                                 }
10536                         }
10537                 }
10538         }
10539 }
10540
10541
10542 /*
10543  * Return value of 1, the TCB is unlocked and most
10544  * likely gone, return value of 0, the TCP is still
10545  * locked.
10546  */
10547 static int
10548 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10549     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10550     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10551 {
10552         /*
10553          * Update window information. Don't look at window if no ACK: TAC's
10554          * send garbage on first SYN.
10555          */
10556         int32_t nsegs;
10557         int32_t tfo_syn;
10558         struct tcp_rack *rack;
10559
10560         rack = (struct tcp_rack *)tp->t_fb_ptr;
10561         INP_WLOCK_ASSERT(tp->t_inpcb);
10562         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10563         if ((thflags & TH_ACK) &&
10564             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10565             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10566             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10567                 /* keep track of pure window updates */
10568                 if (tlen == 0 &&
10569                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10570                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10571                 tp->snd_wnd = tiwin;
10572                 rack_validate_fo_sendwin_up(tp, rack);
10573                 tp->snd_wl1 = th->th_seq;
10574                 tp->snd_wl2 = th->th_ack;
10575                 if (tp->snd_wnd > tp->max_sndwnd)
10576                         tp->max_sndwnd = tp->snd_wnd;
10577                 rack->r_wanted_output = 1;
10578         } else if (thflags & TH_ACK) {
10579                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10580                         tp->snd_wnd = tiwin;
10581                         rack_validate_fo_sendwin_up(tp, rack);
10582                         tp->snd_wl1 = th->th_seq;
10583                         tp->snd_wl2 = th->th_ack;
10584                 }
10585         }
10586         if (tp->snd_wnd < ctf_outstanding(tp))
10587                 /* The peer collapsed the window */
10588                 rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
10589         else if (rack->rc_has_collapsed)
10590                 rack_un_collapse_window(rack, __LINE__);
10591         if ((rack->r_collapse_point_valid) &&
10592             (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
10593                 rack->r_collapse_point_valid = 0;
10594         /* Was persist timer active and now we have window space? */
10595         if ((rack->rc_in_persist != 0) &&
10596             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10597                                 rack->r_ctl.rc_pace_min_segs))) {
10598                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10599                 tp->snd_nxt = tp->snd_max;
10600                 /* Make sure we output to start the timer */
10601                 rack->r_wanted_output = 1;
10602         }
10603         /* Do we enter persists? */
10604         if ((rack->rc_in_persist == 0) &&
10605             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10606             TCPS_HAVEESTABLISHED(tp->t_state) &&
10607             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10608             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10609             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10610                 /*
10611                  * Here the rwnd is less than
10612                  * the pacing size, we are established,
10613                  * nothing is outstanding, and there is
10614                  * data to send. Enter persists.
10615                  */
10616                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10617         }
10618         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10619                 m_freem(m);
10620                 return (0);
10621         }
10622         /*
10623          * don't process the URG bit, ignore them drag
10624          * along the up.
10625          */
10626         tp->rcv_up = tp->rcv_nxt;
10627         INP_WLOCK_ASSERT(tp->t_inpcb);
10628
10629         /*
10630          * Process the segment text, merging it into the TCP sequencing
10631          * queue, and arranging for acknowledgment of receipt if necessary.
10632          * This process logically involves adjusting tp->rcv_wnd as data is
10633          * presented to the user (this happens in tcp_usrreq.c, case
10634          * PRU_RCVD).  If a FIN has already been received on this connection
10635          * then we just ignore the text.
10636          */
10637         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10638                    IS_FASTOPEN(tp->t_flags));
10639         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10640             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10641                 tcp_seq save_start = th->th_seq;
10642                 tcp_seq save_rnxt  = tp->rcv_nxt;
10643                 int     save_tlen  = tlen;
10644
10645                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10646                 /*
10647                  * Insert segment which includes th into TCP reassembly
10648                  * queue with control block tp.  Set thflags to whether
10649                  * reassembly now includes a segment with FIN.  This handles
10650                  * the common case inline (segment is the next to be
10651                  * received on an established connection, and the queue is
10652                  * empty), avoiding linkage into and removal from the queue
10653                  * and repetition of various conversions. Set DELACK for
10654                  * segments received in order, but ack immediately when
10655                  * segments are out of order (so fast retransmit can work).
10656                  */
10657                 if (th->th_seq == tp->rcv_nxt &&
10658                     SEGQ_EMPTY(tp) &&
10659                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10660                     tfo_syn)) {
10661 #ifdef NETFLIX_SB_LIMITS
10662                         u_int mcnt, appended;
10663
10664                         if (so->so_rcv.sb_shlim) {
10665                                 mcnt = m_memcnt(m);
10666                                 appended = 0;
10667                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10668                                     CFO_NOSLEEP, NULL) == false) {
10669                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10670                                         m_freem(m);
10671                                         return (0);
10672                                 }
10673                         }
10674 #endif
10675                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10676                         tp->rcv_nxt += tlen;
10677                         if (tlen &&
10678                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10679                             (tp->t_fbyte_in == 0)) {
10680                                 tp->t_fbyte_in = ticks;
10681                                 if (tp->t_fbyte_in == 0)
10682                                         tp->t_fbyte_in = 1;
10683                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10684                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10685                         }
10686                         thflags = tcp_get_flags(th) & TH_FIN;
10687                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10688                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10689                         SOCKBUF_LOCK(&so->so_rcv);
10690                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10691                                 m_freem(m);
10692                         } else
10693 #ifdef NETFLIX_SB_LIMITS
10694                                 appended =
10695 #endif
10696                                         sbappendstream_locked(&so->so_rcv, m, 0);
10697
10698                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10699                         /* NB: sorwakeup_locked() does an implicit unlock. */
10700                         sorwakeup_locked(so);
10701 #ifdef NETFLIX_SB_LIMITS
10702                         if (so->so_rcv.sb_shlim && appended != mcnt)
10703                                 counter_fo_release(so->so_rcv.sb_shlim,
10704                                     mcnt - appended);
10705 #endif
10706                 } else {
10707                         /*
10708                          * XXX: Due to the header drop above "th" is
10709                          * theoretically invalid by now.  Fortunately
10710                          * m_adj() doesn't actually frees any mbufs when
10711                          * trimming from the head.
10712                          */
10713                         tcp_seq temp = save_start;
10714
10715                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10716                         tp->t_flags |= TF_ACKNOW;
10717                         if (tp->t_flags & TF_WAKESOR) {
10718                                 tp->t_flags &= ~TF_WAKESOR;
10719                                 /* NB: sorwakeup_locked() does an implicit unlock. */
10720                                 sorwakeup_locked(so);
10721                         }
10722                 }
10723                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10724                     (save_tlen > 0) &&
10725                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10726                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10727                                 /*
10728                                  * DSACK actually handled in the fastpath
10729                                  * above.
10730                                  */
10731                                 RACK_OPTS_INC(tcp_sack_path_1);
10732                                 tcp_update_sack_list(tp, save_start,
10733                                     save_start + save_tlen);
10734                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10735                                 if ((tp->rcv_numsacks >= 1) &&
10736                                     (tp->sackblks[0].end == save_start)) {
10737                                         /*
10738                                          * Partial overlap, recorded at todrop
10739                                          * above.
10740                                          */
10741                                         RACK_OPTS_INC(tcp_sack_path_2a);
10742                                         tcp_update_sack_list(tp,
10743                                             tp->sackblks[0].start,
10744                                             tp->sackblks[0].end);
10745                                 } else {
10746                                         RACK_OPTS_INC(tcp_sack_path_2b);
10747                                         tcp_update_dsack_list(tp, save_start,
10748                                             save_start + save_tlen);
10749                                 }
10750                         } else if (tlen >= save_tlen) {
10751                                 /* Update of sackblks. */
10752                                 RACK_OPTS_INC(tcp_sack_path_3);
10753                                 tcp_update_dsack_list(tp, save_start,
10754                                     save_start + save_tlen);
10755                         } else if (tlen > 0) {
10756                                 RACK_OPTS_INC(tcp_sack_path_4);
10757                                 tcp_update_dsack_list(tp, save_start,
10758                                     save_start + tlen);
10759                         }
10760                 }
10761         } else {
10762                 m_freem(m);
10763                 thflags &= ~TH_FIN;
10764         }
10765
10766         /*
10767          * If FIN is received ACK the FIN and let the user know that the
10768          * connection is closing.
10769          */
10770         if (thflags & TH_FIN) {
10771                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10772                         /* The socket upcall is handled by socantrcvmore. */
10773                         socantrcvmore(so);
10774                         /*
10775                          * If connection is half-synchronized (ie NEEDSYN
10776                          * flag on) then delay ACK, so it may be piggybacked
10777                          * when SYN is sent. Otherwise, since we received a
10778                          * FIN then no more input can be expected, send ACK
10779                          * now.
10780                          */
10781                         if (tp->t_flags & TF_NEEDSYN) {
10782                                 rack_timer_cancel(tp, rack,
10783                                     rack->r_ctl.rc_rcvtime, __LINE__);
10784                                 tp->t_flags |= TF_DELACK;
10785                         } else {
10786                                 tp->t_flags |= TF_ACKNOW;
10787                         }
10788                         tp->rcv_nxt++;
10789                 }
10790                 switch (tp->t_state) {
10791                         /*
10792                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10793                          * CLOSE_WAIT state.
10794                          */
10795                 case TCPS_SYN_RECEIVED:
10796                         tp->t_starttime = ticks;
10797                         /* FALLTHROUGH */
10798                 case TCPS_ESTABLISHED:
10799                         rack_timer_cancel(tp, rack,
10800                             rack->r_ctl.rc_rcvtime, __LINE__);
10801                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10802                         break;
10803
10804                         /*
10805                          * If still in FIN_WAIT_1 STATE FIN has not been
10806                          * acked so enter the CLOSING state.
10807                          */
10808                 case TCPS_FIN_WAIT_1:
10809                         rack_timer_cancel(tp, rack,
10810                             rack->r_ctl.rc_rcvtime, __LINE__);
10811                         tcp_state_change(tp, TCPS_CLOSING);
10812                         break;
10813
10814                         /*
10815                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10816                          * starting the time-wait timer, turning off the
10817                          * other standard timers.
10818                          */
10819                 case TCPS_FIN_WAIT_2:
10820                         rack_timer_cancel(tp, rack,
10821                             rack->r_ctl.rc_rcvtime, __LINE__);
10822                         tcp_twstart(tp);
10823                         return (1);
10824                 }
10825         }
10826         /*
10827          * Return any desired output.
10828          */
10829         if ((tp->t_flags & TF_ACKNOW) ||
10830             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10831                 rack->r_wanted_output = 1;
10832         }
10833         INP_WLOCK_ASSERT(tp->t_inpcb);
10834         return (0);
10835 }
10836
10837 /*
10838  * Here nothing is really faster, its just that we
10839  * have broken out the fast-data path also just like
10840  * the fast-ack.
10841  */
10842 static int
10843 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10844     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10845     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10846 {
10847         int32_t nsegs;
10848         int32_t newsize = 0;    /* automatic sockbuf scaling */
10849         struct tcp_rack *rack;
10850 #ifdef NETFLIX_SB_LIMITS
10851         u_int mcnt, appended;
10852 #endif
10853 #ifdef TCPDEBUG
10854         /*
10855          * The size of tcp_saveipgen must be the size of the max ip header,
10856          * now IPv6.
10857          */
10858         u_char tcp_saveipgen[IP6_HDR_LEN];
10859         struct tcphdr tcp_savetcp;
10860         short ostate = 0;
10861
10862 #endif
10863         /*
10864          * If last ACK falls within this segment's sequence numbers, record
10865          * the timestamp. NOTE that the test is modified according to the
10866          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10867          */
10868         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10869                 return (0);
10870         }
10871         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10872                 return (0);
10873         }
10874         if (tiwin && tiwin != tp->snd_wnd) {
10875                 return (0);
10876         }
10877         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10878                 return (0);
10879         }
10880         if (__predict_false((to->to_flags & TOF_TS) &&
10881             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10882                 return (0);
10883         }
10884         if (__predict_false((th->th_ack != tp->snd_una))) {
10885                 return (0);
10886         }
10887         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10888                 return (0);
10889         }
10890         if ((to->to_flags & TOF_TS) != 0 &&
10891             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10892                 tp->ts_recent_age = tcp_ts_getticks();
10893                 tp->ts_recent = to->to_tsval;
10894         }
10895         rack = (struct tcp_rack *)tp->t_fb_ptr;
10896         /*
10897          * This is a pure, in-sequence data packet with nothing on the
10898          * reassembly queue and we have enough buffer space to take it.
10899          */
10900         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10901
10902 #ifdef NETFLIX_SB_LIMITS
10903         if (so->so_rcv.sb_shlim) {
10904                 mcnt = m_memcnt(m);
10905                 appended = 0;
10906                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10907                     CFO_NOSLEEP, NULL) == false) {
10908                         counter_u64_add(tcp_sb_shlim_fails, 1);
10909                         m_freem(m);
10910                         return (1);
10911                 }
10912         }
10913 #endif
10914         /* Clean receiver SACK report if present */
10915         if (tp->rcv_numsacks)
10916                 tcp_clean_sackreport(tp);
10917         KMOD_TCPSTAT_INC(tcps_preddat);
10918         tp->rcv_nxt += tlen;
10919         if (tlen &&
10920             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10921             (tp->t_fbyte_in == 0)) {
10922                 tp->t_fbyte_in = ticks;
10923                 if (tp->t_fbyte_in == 0)
10924                         tp->t_fbyte_in = 1;
10925                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10926                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10927         }
10928         /*
10929          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10930          */
10931         tp->snd_wl1 = th->th_seq;
10932         /*
10933          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10934          */
10935         tp->rcv_up = tp->rcv_nxt;
10936         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10937         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10938 #ifdef TCPDEBUG
10939         if (so->so_options & SO_DEBUG)
10940                 tcp_trace(TA_INPUT, ostate, tp,
10941                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
10942 #endif
10943         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10944
10945         /* Add data to socket buffer. */
10946         SOCKBUF_LOCK(&so->so_rcv);
10947         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10948                 m_freem(m);
10949         } else {
10950                 /*
10951                  * Set new socket buffer size. Give up when limit is
10952                  * reached.
10953                  */
10954                 if (newsize)
10955                         if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
10956                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10957                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10958 #ifdef NETFLIX_SB_LIMITS
10959                 appended =
10960 #endif
10961                         sbappendstream_locked(&so->so_rcv, m, 0);
10962                 ctf_calc_rwin(so, tp);
10963         }
10964         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10965         /* NB: sorwakeup_locked() does an implicit unlock. */
10966         sorwakeup_locked(so);
10967 #ifdef NETFLIX_SB_LIMITS
10968         if (so->so_rcv.sb_shlim && mcnt != appended)
10969                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10970 #endif
10971         rack_handle_delayed_ack(tp, rack, tlen, 0);
10972         if (tp->snd_una == tp->snd_max)
10973                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10974         return (1);
10975 }
10976
10977 /*
10978  * This subfunction is used to try to highly optimize the
10979  * fast path. We again allow window updates that are
10980  * in sequence to remain in the fast-path. We also add
10981  * in the __predict's to attempt to help the compiler.
10982  * Note that if we return a 0, then we can *not* process
10983  * it and the caller should push the packet into the
10984  * slow-path.
10985  */
10986 static int
10987 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10988     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10989     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10990 {
10991         int32_t acked;
10992         int32_t nsegs;
10993 #ifdef TCPDEBUG
10994         /*
10995          * The size of tcp_saveipgen must be the size of the max ip header,
10996          * now IPv6.
10997          */
10998         u_char tcp_saveipgen[IP6_HDR_LEN];
10999         struct tcphdr tcp_savetcp;
11000         short ostate = 0;
11001 #endif
11002         int32_t under_pacing = 0;
11003         struct tcp_rack *rack;
11004
11005         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
11006                 /* Old ack, behind (or duplicate to) the last one rcv'd */
11007                 return (0);
11008         }
11009         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
11010                 /* Above what we have sent? */
11011                 return (0);
11012         }
11013         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
11014                 /* We are retransmitting */
11015                 return (0);
11016         }
11017         if (__predict_false(tiwin == 0)) {
11018                 /* zero window */
11019                 return (0);
11020         }
11021         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
11022                 /* We need a SYN or a FIN, unlikely.. */
11023                 return (0);
11024         }
11025         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
11026                 /* Timestamp is behind .. old ack with seq wrap? */
11027                 return (0);
11028         }
11029         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
11030                 /* Still recovering */
11031                 return (0);
11032         }
11033         rack = (struct tcp_rack *)tp->t_fb_ptr;
11034         if (rack->r_ctl.rc_sacked) {
11035                 /* We have sack holes on our scoreboard */
11036                 return (0);
11037         }
11038         /* Ok if we reach here, we can process a fast-ack */
11039         if (rack->gp_ready &&
11040             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11041                 under_pacing = 1;
11042         }
11043         nsegs = max(1, m->m_pkthdr.lro_nsegs);
11044         rack_log_ack(tp, to, th, 0, 0);
11045         /* Did the window get updated? */
11046         if (tiwin != tp->snd_wnd) {
11047                 tp->snd_wnd = tiwin;
11048                 rack_validate_fo_sendwin_up(tp, rack);
11049                 tp->snd_wl1 = th->th_seq;
11050                 if (tp->snd_wnd > tp->max_sndwnd)
11051                         tp->max_sndwnd = tp->snd_wnd;
11052         }
11053         /* Do we exit persists? */
11054         if ((rack->rc_in_persist != 0) &&
11055             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
11056                                rack->r_ctl.rc_pace_min_segs))) {
11057                 rack_exit_persist(tp, rack, cts);
11058         }
11059         /* Do we enter persists? */
11060         if ((rack->rc_in_persist == 0) &&
11061             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
11062             TCPS_HAVEESTABLISHED(tp->t_state) &&
11063             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
11064             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
11065             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
11066                 /*
11067                  * Here the rwnd is less than
11068                  * the pacing size, we are established,
11069                  * nothing is outstanding, and there is
11070                  * data to send. Enter persists.
11071                  */
11072                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
11073         }
11074         /*
11075          * If last ACK falls within this segment's sequence numbers, record
11076          * the timestamp. NOTE that the test is modified according to the
11077          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
11078          */
11079         if ((to->to_flags & TOF_TS) != 0 &&
11080             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
11081                 tp->ts_recent_age = tcp_ts_getticks();
11082                 tp->ts_recent = to->to_tsval;
11083         }
11084         /*
11085          * This is a pure ack for outstanding data.
11086          */
11087         KMOD_TCPSTAT_INC(tcps_predack);
11088
11089         /*
11090          * "bad retransmit" recovery.
11091          */
11092         if ((tp->t_flags & TF_PREVVALID) &&
11093             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11094                 tp->t_flags &= ~TF_PREVVALID;
11095                 if (tp->t_rxtshift == 1 &&
11096                     (int)(ticks - tp->t_badrxtwin) < 0)
11097                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
11098         }
11099         /*
11100          * Recalculate the transmit timer / rtt.
11101          *
11102          * Some boxes send broken timestamp replies during the SYN+ACK
11103          * phase, ignore timestamps of 0 or we could calculate a huge RTT
11104          * and blow up the retransmit timer.
11105          */
11106         acked = BYTES_THIS_ACK(tp, th);
11107
11108 #ifdef TCP_HHOOK
11109         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
11110         hhook_run_tcp_est_in(tp, th, to);
11111 #endif
11112         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11113         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11114         if (acked) {
11115                 struct mbuf *mfree;
11116
11117                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
11118                 SOCKBUF_LOCK(&so->so_snd);
11119                 mfree = sbcut_locked(&so->so_snd, acked);
11120                 tp->snd_una = th->th_ack;
11121                 /* Note we want to hold the sb lock through the sendmap adjust */
11122                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
11123                 /* Wake up the socket if we have room to write more */
11124                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11125                 sowwakeup_locked(so);
11126                 m_freem(mfree);
11127                 tp->t_rxtshift = 0;
11128                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11129                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11130                 rack->rc_tlp_in_progress = 0;
11131                 rack->r_ctl.rc_tlp_cnt_out = 0;
11132                 /*
11133                  * If it is the RXT timer we want to
11134                  * stop it, so we can restart a TLP.
11135                  */
11136                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11137                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11138 #ifdef NETFLIX_HTTP_LOGGING
11139                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
11140 #endif
11141         }
11142         /*
11143          * Let the congestion control algorithm update congestion control
11144          * related information. This typically means increasing the
11145          * congestion window.
11146          */
11147         if (tp->snd_wnd < ctf_outstanding(tp)) {
11148                 /* The peer collapsed the window */
11149                 rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
11150         } else if (rack->rc_has_collapsed)
11151                 rack_un_collapse_window(rack, __LINE__);
11152         if ((rack->r_collapse_point_valid) &&
11153             (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
11154                 rack->r_collapse_point_valid = 0;
11155         /*
11156          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
11157          */
11158         tp->snd_wl2 = th->th_ack;
11159         tp->t_dupacks = 0;
11160         m_freem(m);
11161         /* ND6_HINT(tp);         *//* Some progress has been made. */
11162
11163         /*
11164          * If all outstanding data are acked, stop retransmit timer,
11165          * otherwise restart timer using current (possibly backed-off)
11166          * value. If process is waiting for space, wakeup/selwakeup/signal.
11167          * If data are ready to send, let tcp_output decide between more
11168          * output or persist.
11169          */
11170 #ifdef TCPDEBUG
11171         if (so->so_options & SO_DEBUG)
11172                 tcp_trace(TA_INPUT, ostate, tp,
11173                     (void *)tcp_saveipgen,
11174                     &tcp_savetcp, 0);
11175 #endif
11176         if (under_pacing &&
11177             (rack->use_fixed_rate == 0) &&
11178             (rack->in_probe_rtt == 0) &&
11179             rack->rc_gp_dyn_mul &&
11180             rack->rc_always_pace) {
11181                 /* Check if we are dragging bottom */
11182                 rack_check_bottom_drag(tp, rack, so, acked);
11183         }
11184         if (tp->snd_una == tp->snd_max) {
11185                 tp->t_flags &= ~TF_PREVVALID;
11186                 rack->r_ctl.retran_during_recovery = 0;
11187                 rack->r_ctl.dsack_byte_cnt = 0;
11188                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11189                 if (rack->r_ctl.rc_went_idle_time == 0)
11190                         rack->r_ctl.rc_went_idle_time = 1;
11191                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11192                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
11193                         tp->t_acktime = 0;
11194                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11195         }
11196         if (acked && rack->r_fast_output)
11197                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11198         if (sbavail(&so->so_snd)) {
11199                 rack->r_wanted_output = 1;
11200         }
11201         return (1);
11202 }
11203
11204 /*
11205  * Return value of 1, the TCB is unlocked and most
11206  * likely gone, return value of 0, the TCP is still
11207  * locked.
11208  */
11209 static int
11210 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11211     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11212     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11213 {
11214         int32_t ret_val = 0;
11215         int32_t todrop;
11216         int32_t ourfinisacked = 0;
11217         struct tcp_rack *rack;
11218
11219         ctf_calc_rwin(so, tp);
11220         /*
11221          * If the state is SYN_SENT: if seg contains an ACK, but not for our
11222          * SYN, drop the input. if seg contains a RST, then drop the
11223          * connection. if seg does not contain SYN, then drop it. Otherwise
11224          * this is an acceptable SYN segment initialize tp->rcv_nxt and
11225          * tp->irs if seg contains ack then advance tp->snd_una if seg
11226          * contains an ECE and ECN support is enabled, the stream is ECN
11227          * capable. if SYN has been acked change to ESTABLISHED else
11228          * SYN_RCVD state arrange for segment to be acked (eventually)
11229          * continue processing rest of data/controls.
11230          */
11231         if ((thflags & TH_ACK) &&
11232             (SEQ_LEQ(th->th_ack, tp->iss) ||
11233             SEQ_GT(th->th_ack, tp->snd_max))) {
11234                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11235                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11236                 return (1);
11237         }
11238         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11239                 TCP_PROBE5(connect__refused, NULL, tp,
11240                     mtod(m, const char *), tp, th);
11241                 tp = tcp_drop(tp, ECONNREFUSED);
11242                 ctf_do_drop(m, tp);
11243                 return (1);
11244         }
11245         if (thflags & TH_RST) {
11246                 ctf_do_drop(m, tp);
11247                 return (1);
11248         }
11249         if (!(thflags & TH_SYN)) {
11250                 ctf_do_drop(m, tp);
11251                 return (1);
11252         }
11253         tp->irs = th->th_seq;
11254         tcp_rcvseqinit(tp);
11255         rack = (struct tcp_rack *)tp->t_fb_ptr;
11256         if (thflags & TH_ACK) {
11257                 int tfo_partial = 0;
11258
11259                 KMOD_TCPSTAT_INC(tcps_connects);
11260                 soisconnected(so);
11261 #ifdef MAC
11262                 mac_socketpeer_set_from_mbuf(m, so);
11263 #endif
11264                 /* Do window scaling on this connection? */
11265                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11266                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11267                         tp->rcv_scale = tp->request_r_scale;
11268                 }
11269                 tp->rcv_adv += min(tp->rcv_wnd,
11270                     TCP_MAXWIN << tp->rcv_scale);
11271                 /*
11272                  * If not all the data that was sent in the TFO SYN
11273                  * has been acked, resend the remainder right away.
11274                  */
11275                 if (IS_FASTOPEN(tp->t_flags) &&
11276                     (tp->snd_una != tp->snd_max)) {
11277                         tp->snd_nxt = th->th_ack;
11278                         tfo_partial = 1;
11279                 }
11280                 /*
11281                  * If there's data, delay ACK; if there's also a FIN ACKNOW
11282                  * will be turned on later.
11283                  */
11284                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11285                         rack_timer_cancel(tp, rack,
11286                                           rack->r_ctl.rc_rcvtime, __LINE__);
11287                         tp->t_flags |= TF_DELACK;
11288                 } else {
11289                         rack->r_wanted_output = 1;
11290                         tp->t_flags |= TF_ACKNOW;
11291                         rack->rc_dack_toggle = 0;
11292                 }
11293
11294                 tcp_ecn_input_syn_sent(tp, thflags, iptos);
11295
11296                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
11297                         /*
11298                          * We advance snd_una for the
11299                          * fast open case. If th_ack is
11300                          * acknowledging data beyond
11301                          * snd_una we can't just call
11302                          * ack-processing since the
11303                          * data stream in our send-map
11304                          * will start at snd_una + 1 (one
11305                          * beyond the SYN). If its just
11306                          * equal we don't need to do that
11307                          * and there is no send_map.
11308                          */
11309                         tp->snd_una++;
11310                 }
11311                 /*
11312                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11313                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11314                  */
11315                 tp->t_starttime = ticks;
11316                 if (tp->t_flags & TF_NEEDFIN) {
11317                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
11318                         tp->t_flags &= ~TF_NEEDFIN;
11319                         thflags &= ~TH_SYN;
11320                 } else {
11321                         tcp_state_change(tp, TCPS_ESTABLISHED);
11322                         TCP_PROBE5(connect__established, NULL, tp,
11323                             mtod(m, const char *), tp, th);
11324                         rack_cc_conn_init(tp);
11325                 }
11326         } else {
11327                 /*
11328                  * Received initial SYN in SYN-SENT[*] state => simultaneous
11329                  * open.  If segment contains CC option and there is a
11330                  * cached CC, apply TAO test. If it succeeds, connection is *
11331                  * half-synchronized. Otherwise, do 3-way handshake:
11332                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11333                  * there was no CC option, clear cached CC value.
11334                  */
11335                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
11336                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
11337         }
11338         INP_WLOCK_ASSERT(tp->t_inpcb);
11339         /*
11340          * Advance th->th_seq to correspond to first data byte. If data,
11341          * trim to stay within window, dropping FIN if necessary.
11342          */
11343         th->th_seq++;
11344         if (tlen > tp->rcv_wnd) {
11345                 todrop = tlen - tp->rcv_wnd;
11346                 m_adj(m, -todrop);
11347                 tlen = tp->rcv_wnd;
11348                 thflags &= ~TH_FIN;
11349                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11350                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11351         }
11352         tp->snd_wl1 = th->th_seq - 1;
11353         tp->rcv_up = th->th_seq;
11354         /*
11355          * Client side of transaction: already sent SYN and data. If the
11356          * remote host used T/TCP to validate the SYN, our data will be
11357          * ACK'd; if so, enter normal data segment processing in the middle
11358          * of step 5, ack processing. Otherwise, goto step 6.
11359          */
11360         if (thflags & TH_ACK) {
11361                 /* For syn-sent we need to possibly update the rtt */
11362                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11363                         uint32_t t, mcts;
11364
11365                         mcts = tcp_ts_getticks();
11366                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11367                         if (!tp->t_rttlow || tp->t_rttlow > t)
11368                                 tp->t_rttlow = t;
11369                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11370                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11371                         tcp_rack_xmit_timer_commit(rack, tp);
11372                 }
11373                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11374                         return (ret_val);
11375                 /* We may have changed to FIN_WAIT_1 above */
11376                 if (tp->t_state == TCPS_FIN_WAIT_1) {
11377                         /*
11378                          * In FIN_WAIT_1 STATE in addition to the processing
11379                          * for the ESTABLISHED state if our FIN is now
11380                          * acknowledged then enter FIN_WAIT_2.
11381                          */
11382                         if (ourfinisacked) {
11383                                 /*
11384                                  * If we can't receive any more data, then
11385                                  * closing user can proceed. Starting the
11386                                  * timer is contrary to the specification,
11387                                  * but if we don't get a FIN we'll hang
11388                                  * forever.
11389                                  *
11390                                  * XXXjl: we should release the tp also, and
11391                                  * use a compressed state.
11392                                  */
11393                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11394                                         soisdisconnected(so);
11395                                         tcp_timer_activate(tp, TT_2MSL,
11396                                             (tcp_fast_finwait2_recycle ?
11397                                             tcp_finwait2_timeout :
11398                                             TP_MAXIDLE(tp)));
11399                                 }
11400                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11401                         }
11402                 }
11403         }
11404         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11405            tiwin, thflags, nxt_pkt));
11406 }
11407
11408 /*
11409  * Return value of 1, the TCB is unlocked and most
11410  * likely gone, return value of 0, the TCP is still
11411  * locked.
11412  */
11413 static int
11414 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11415     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11416     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11417 {
11418         struct tcp_rack *rack;
11419         int32_t ret_val = 0;
11420         int32_t ourfinisacked = 0;
11421
11422         ctf_calc_rwin(so, tp);
11423         if ((thflags & TH_ACK) &&
11424             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11425             SEQ_GT(th->th_ack, tp->snd_max))) {
11426                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11427                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11428                 return (1);
11429         }
11430         rack = (struct tcp_rack *)tp->t_fb_ptr;
11431         if (IS_FASTOPEN(tp->t_flags)) {
11432                 /*
11433                  * When a TFO connection is in SYN_RECEIVED, the
11434                  * only valid packets are the initial SYN, a
11435                  * retransmit/copy of the initial SYN (possibly with
11436                  * a subset of the original data), a valid ACK, a
11437                  * FIN, or a RST.
11438                  */
11439                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11440                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11441                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11442                         return (1);
11443                 } else if (thflags & TH_SYN) {
11444                         /* non-initial SYN is ignored */
11445                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11446                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11447                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11448                                 ctf_do_drop(m, NULL);
11449                                 return (0);
11450                         }
11451                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11452                         ctf_do_drop(m, NULL);
11453                         return (0);
11454                 }
11455         }
11456
11457         if ((thflags & TH_RST) ||
11458             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11459                 return (__ctf_process_rst(m, th, so, tp,
11460                                           &rack->r_ctl.challenge_ack_ts,
11461                                           &rack->r_ctl.challenge_ack_cnt));
11462         /*
11463          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11464          * it's less than ts_recent, drop it.
11465          */
11466         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11467             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11468                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11469                         return (ret_val);
11470         }
11471         /*
11472          * In the SYN-RECEIVED state, validate that the packet belongs to
11473          * this connection before trimming the data to fit the receive
11474          * window.  Check the sequence number versus IRS since we know the
11475          * sequence numbers haven't wrapped.  This is a partial fix for the
11476          * "LAND" DoS attack.
11477          */
11478         if (SEQ_LT(th->th_seq, tp->irs)) {
11479                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11480                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11481                 return (1);
11482         }
11483         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11484                               &rack->r_ctl.challenge_ack_ts,
11485                               &rack->r_ctl.challenge_ack_cnt)) {
11486                 return (ret_val);
11487         }
11488         /*
11489          * If last ACK falls within this segment's sequence numbers, record
11490          * its timestamp. NOTE: 1) That the test incorporates suggestions
11491          * from the latest proposal of the tcplw@cray.com list (Braden
11492          * 1993/04/26). 2) That updating only on newer timestamps interferes
11493          * with our earlier PAWS tests, so this check should be solely
11494          * predicated on the sequence space of this segment. 3) That we
11495          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11496          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11497          * SEG.Len, This modified check allows us to overcome RFC1323's
11498          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11499          * p.869. In such cases, we can still calculate the RTT correctly
11500          * when RCV.NXT == Last.ACK.Sent.
11501          */
11502         if ((to->to_flags & TOF_TS) != 0 &&
11503             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11504             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11505             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11506                 tp->ts_recent_age = tcp_ts_getticks();
11507                 tp->ts_recent = to->to_tsval;
11508         }
11509         tp->snd_wnd = tiwin;
11510         rack_validate_fo_sendwin_up(tp, rack);
11511         /*
11512          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11513          * is on (half-synchronized state), then queue data for later
11514          * processing; else drop segment and return.
11515          */
11516         if ((thflags & TH_ACK) == 0) {
11517                 if (IS_FASTOPEN(tp->t_flags)) {
11518                         rack_cc_conn_init(tp);
11519                 }
11520                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11521                     tiwin, thflags, nxt_pkt));
11522         }
11523         KMOD_TCPSTAT_INC(tcps_connects);
11524         if (tp->t_flags & TF_SONOTCONN) {
11525                 tp->t_flags &= ~TF_SONOTCONN;
11526                 soisconnected(so);
11527         }
11528         /* Do window scaling? */
11529         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11530             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11531                 tp->rcv_scale = tp->request_r_scale;
11532         }
11533         /*
11534          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11535          * FIN-WAIT-1
11536          */
11537         tp->t_starttime = ticks;
11538         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11539                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11540                 tp->t_tfo_pending = NULL;
11541         }
11542         if (tp->t_flags & TF_NEEDFIN) {
11543                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11544                 tp->t_flags &= ~TF_NEEDFIN;
11545         } else {
11546                 tcp_state_change(tp, TCPS_ESTABLISHED);
11547                 TCP_PROBE5(accept__established, NULL, tp,
11548                     mtod(m, const char *), tp, th);
11549                 /*
11550                  * TFO connections call cc_conn_init() during SYN
11551                  * processing.  Calling it again here for such connections
11552                  * is not harmless as it would undo the snd_cwnd reduction
11553                  * that occurs when a TFO SYN|ACK is retransmitted.
11554                  */
11555                 if (!IS_FASTOPEN(tp->t_flags))
11556                         rack_cc_conn_init(tp);
11557         }
11558         /*
11559          * Account for the ACK of our SYN prior to
11560          * regular ACK processing below, except for
11561          * simultaneous SYN, which is handled later.
11562          */
11563         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11564                 tp->snd_una++;
11565         /*
11566          * If segment contains data or ACK, will call tcp_reass() later; if
11567          * not, do so now to pass queued data to user.
11568          */
11569         if (tlen == 0 && (thflags & TH_FIN) == 0) {
11570                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11571                     (struct mbuf *)0);
11572                 if (tp->t_flags & TF_WAKESOR) {
11573                         tp->t_flags &= ~TF_WAKESOR;
11574                         /* NB: sorwakeup_locked() does an implicit unlock. */
11575                         sorwakeup_locked(so);
11576                 }
11577         }
11578         tp->snd_wl1 = th->th_seq - 1;
11579         /* For syn-recv we need to possibly update the rtt */
11580         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11581                 uint32_t t, mcts;
11582
11583                 mcts = tcp_ts_getticks();
11584                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11585                 if (!tp->t_rttlow || tp->t_rttlow > t)
11586                         tp->t_rttlow = t;
11587                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11588                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11589                 tcp_rack_xmit_timer_commit(rack, tp);
11590         }
11591         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11592                 return (ret_val);
11593         }
11594         if (tp->t_state == TCPS_FIN_WAIT_1) {
11595                 /* We could have went to FIN_WAIT_1 (or EST) above */
11596                 /*
11597                  * In FIN_WAIT_1 STATE in addition to the processing for the
11598                  * ESTABLISHED state if our FIN is now acknowledged then
11599                  * enter FIN_WAIT_2.
11600                  */
11601                 if (ourfinisacked) {
11602                         /*
11603                          * If we can't receive any more data, then closing
11604                          * user can proceed. Starting the timer is contrary
11605                          * to the specification, but if we don't get a FIN
11606                          * we'll hang forever.
11607                          *
11608                          * XXXjl: we should release the tp also, and use a
11609                          * compressed state.
11610                          */
11611                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11612                                 soisdisconnected(so);
11613                                 tcp_timer_activate(tp, TT_2MSL,
11614                                     (tcp_fast_finwait2_recycle ?
11615                                     tcp_finwait2_timeout :
11616                                     TP_MAXIDLE(tp)));
11617                         }
11618                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11619                 }
11620         }
11621         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11622             tiwin, thflags, nxt_pkt));
11623 }
11624
11625 /*
11626  * Return value of 1, the TCB is unlocked and most
11627  * likely gone, return value of 0, the TCP is still
11628  * locked.
11629  */
11630 static int
11631 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11632     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11633     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11634 {
11635         int32_t ret_val = 0;
11636         struct tcp_rack *rack;
11637
11638         /*
11639          * Header prediction: check for the two common cases of a
11640          * uni-directional data xfer.  If the packet has no control flags,
11641          * is in-sequence, the window didn't change and we're not
11642          * retransmitting, it's a candidate.  If the length is zero and the
11643          * ack moved forward, we're the sender side of the xfer.  Just free
11644          * the data acked & wake any higher level process that was blocked
11645          * waiting for space.  If the length is non-zero and the ack didn't
11646          * move, we're the receiver side.  If we're getting packets in-order
11647          * (the reassembly queue is empty), add the data toc The socket
11648          * buffer and note that we need a delayed ack. Make sure that the
11649          * hidden state-flags are also off. Since we check for
11650          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11651          */
11652         rack = (struct tcp_rack *)tp->t_fb_ptr;
11653         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11654             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11655             __predict_true(SEGQ_EMPTY(tp)) &&
11656             __predict_true(th->th_seq == tp->rcv_nxt)) {
11657                 if (tlen == 0) {
11658                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11659                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11660                                 return (0);
11661                         }
11662                 } else {
11663                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11664                             tiwin, nxt_pkt, iptos)) {
11665                                 return (0);
11666                         }
11667                 }
11668         }
11669         ctf_calc_rwin(so, tp);
11670
11671         if ((thflags & TH_RST) ||
11672             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11673                 return (__ctf_process_rst(m, th, so, tp,
11674                                           &rack->r_ctl.challenge_ack_ts,
11675                                           &rack->r_ctl.challenge_ack_cnt));
11676
11677         /*
11678          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11679          * synchronized state.
11680          */
11681         if (thflags & TH_SYN) {
11682                 ctf_challenge_ack(m, th, tp, &ret_val);
11683                 return (ret_val);
11684         }
11685         /*
11686          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11687          * it's less than ts_recent, drop it.
11688          */
11689         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11690             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11691                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11692                         return (ret_val);
11693         }
11694         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11695                               &rack->r_ctl.challenge_ack_ts,
11696                               &rack->r_ctl.challenge_ack_cnt)) {
11697                 return (ret_val);
11698         }
11699         /*
11700          * If last ACK falls within this segment's sequence numbers, record
11701          * its timestamp. NOTE: 1) That the test incorporates suggestions
11702          * from the latest proposal of the tcplw@cray.com list (Braden
11703          * 1993/04/26). 2) That updating only on newer timestamps interferes
11704          * with our earlier PAWS tests, so this check should be solely
11705          * predicated on the sequence space of this segment. 3) That we
11706          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11707          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11708          * SEG.Len, This modified check allows us to overcome RFC1323's
11709          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11710          * p.869. In such cases, we can still calculate the RTT correctly
11711          * when RCV.NXT == Last.ACK.Sent.
11712          */
11713         if ((to->to_flags & TOF_TS) != 0 &&
11714             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11715             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11716             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11717                 tp->ts_recent_age = tcp_ts_getticks();
11718                 tp->ts_recent = to->to_tsval;
11719         }
11720         /*
11721          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11722          * is on (half-synchronized state), then queue data for later
11723          * processing; else drop segment and return.
11724          */
11725         if ((thflags & TH_ACK) == 0) {
11726                 if (tp->t_flags & TF_NEEDSYN) {
11727                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11728                             tiwin, thflags, nxt_pkt));
11729
11730                 } else if (tp->t_flags & TF_ACKNOW) {
11731                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11732                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11733                         return (ret_val);
11734                 } else {
11735                         ctf_do_drop(m, NULL);
11736                         return (0);
11737                 }
11738         }
11739         /*
11740          * Ack processing.
11741          */
11742         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11743                 return (ret_val);
11744         }
11745         if (sbavail(&so->so_snd)) {
11746                 if (ctf_progress_timeout_check(tp, true)) {
11747                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11748                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11749                         return (1);
11750                 }
11751         }
11752         /* State changes only happen in rack_process_data() */
11753         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11754             tiwin, thflags, nxt_pkt));
11755 }
11756
11757 /*
11758  * Return value of 1, the TCB is unlocked and most
11759  * likely gone, return value of 0, the TCP is still
11760  * locked.
11761  */
11762 static int
11763 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11764     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11765     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11766 {
11767         int32_t ret_val = 0;
11768         struct tcp_rack *rack;
11769
11770         rack = (struct tcp_rack *)tp->t_fb_ptr;
11771         ctf_calc_rwin(so, tp);
11772         if ((thflags & TH_RST) ||
11773             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11774                 return (__ctf_process_rst(m, th, so, tp,
11775                                           &rack->r_ctl.challenge_ack_ts,
11776                                           &rack->r_ctl.challenge_ack_cnt));
11777         /*
11778          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11779          * synchronized state.
11780          */
11781         if (thflags & TH_SYN) {
11782                 ctf_challenge_ack(m, th, tp, &ret_val);
11783                 return (ret_val);
11784         }
11785         /*
11786          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11787          * it's less than ts_recent, drop it.
11788          */
11789         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11790             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11791                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11792                         return (ret_val);
11793         }
11794         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11795                               &rack->r_ctl.challenge_ack_ts,
11796                               &rack->r_ctl.challenge_ack_cnt)) {
11797                 return (ret_val);
11798         }
11799         /*
11800          * If last ACK falls within this segment's sequence numbers, record
11801          * its timestamp. NOTE: 1) That the test incorporates suggestions
11802          * from the latest proposal of the tcplw@cray.com list (Braden
11803          * 1993/04/26). 2) That updating only on newer timestamps interferes
11804          * with our earlier PAWS tests, so this check should be solely
11805          * predicated on the sequence space of this segment. 3) That we
11806          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11807          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11808          * SEG.Len, This modified check allows us to overcome RFC1323's
11809          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11810          * p.869. In such cases, we can still calculate the RTT correctly
11811          * when RCV.NXT == Last.ACK.Sent.
11812          */
11813         if ((to->to_flags & TOF_TS) != 0 &&
11814             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11815             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11816             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11817                 tp->ts_recent_age = tcp_ts_getticks();
11818                 tp->ts_recent = to->to_tsval;
11819         }
11820         /*
11821          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11822          * is on (half-synchronized state), then queue data for later
11823          * processing; else drop segment and return.
11824          */
11825         if ((thflags & TH_ACK) == 0) {
11826                 if (tp->t_flags & TF_NEEDSYN) {
11827                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11828                             tiwin, thflags, nxt_pkt));
11829
11830                 } else if (tp->t_flags & TF_ACKNOW) {
11831                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11832                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11833                         return (ret_val);
11834                 } else {
11835                         ctf_do_drop(m, NULL);
11836                         return (0);
11837                 }
11838         }
11839         /*
11840          * Ack processing.
11841          */
11842         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11843                 return (ret_val);
11844         }
11845         if (sbavail(&so->so_snd)) {
11846                 if (ctf_progress_timeout_check(tp, true)) {
11847                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11848                                                 tp, tick, PROGRESS_DROP, __LINE__);
11849                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11850                         return (1);
11851                 }
11852         }
11853         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11854             tiwin, thflags, nxt_pkt));
11855 }
11856
11857 static int
11858 rack_check_data_after_close(struct mbuf *m,
11859     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11860 {
11861         struct tcp_rack *rack;
11862
11863         rack = (struct tcp_rack *)tp->t_fb_ptr;
11864         if (rack->rc_allow_data_af_clo == 0) {
11865         close_now:
11866                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11867                 /* tcp_close will kill the inp pre-log the Reset */
11868                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11869                 tp = tcp_close(tp);
11870                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11871                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11872                 return (1);
11873         }
11874         if (sbavail(&so->so_snd) == 0)
11875                 goto close_now;
11876         /* Ok we allow data that is ignored and a followup reset */
11877         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11878         tp->rcv_nxt = th->th_seq + *tlen;
11879         tp->t_flags2 |= TF2_DROP_AF_DATA;
11880         rack->r_wanted_output = 1;
11881         *tlen = 0;
11882         return (0);
11883 }
11884
11885 /*
11886  * Return value of 1, the TCB is unlocked and most
11887  * likely gone, return value of 0, the TCP is still
11888  * locked.
11889  */
11890 static int
11891 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11892     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11893     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11894 {
11895         int32_t ret_val = 0;
11896         int32_t ourfinisacked = 0;
11897         struct tcp_rack *rack;
11898
11899         rack = (struct tcp_rack *)tp->t_fb_ptr;
11900         ctf_calc_rwin(so, tp);
11901
11902         if ((thflags & TH_RST) ||
11903             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11904                 return (__ctf_process_rst(m, th, so, tp,
11905                                           &rack->r_ctl.challenge_ack_ts,
11906                                           &rack->r_ctl.challenge_ack_cnt));
11907         /*
11908          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11909          * synchronized state.
11910          */
11911         if (thflags & TH_SYN) {
11912                 ctf_challenge_ack(m, th, tp, &ret_val);
11913                 return (ret_val);
11914         }
11915         /*
11916          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11917          * it's less than ts_recent, drop it.
11918          */
11919         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11920             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11921                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11922                         return (ret_val);
11923         }
11924         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11925                               &rack->r_ctl.challenge_ack_ts,
11926                               &rack->r_ctl.challenge_ack_cnt)) {
11927                 return (ret_val);
11928         }
11929         /*
11930          * If new data are received on a connection after the user processes
11931          * are gone, then RST the other end.
11932          */
11933         if ((tp->t_flags & TF_CLOSED) && tlen &&
11934             rack_check_data_after_close(m, tp, &tlen, th, so))
11935                 return (1);
11936         /*
11937          * If last ACK falls within this segment's sequence numbers, record
11938          * its timestamp. NOTE: 1) That the test incorporates suggestions
11939          * from the latest proposal of the tcplw@cray.com list (Braden
11940          * 1993/04/26). 2) That updating only on newer timestamps interferes
11941          * with our earlier PAWS tests, so this check should be solely
11942          * predicated on the sequence space of this segment. 3) That we
11943          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11944          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11945          * SEG.Len, This modified check allows us to overcome RFC1323's
11946          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11947          * p.869. In such cases, we can still calculate the RTT correctly
11948          * when RCV.NXT == Last.ACK.Sent.
11949          */
11950         if ((to->to_flags & TOF_TS) != 0 &&
11951             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11952             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11953             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11954                 tp->ts_recent_age = tcp_ts_getticks();
11955                 tp->ts_recent = to->to_tsval;
11956         }
11957         /*
11958          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11959          * is on (half-synchronized state), then queue data for later
11960          * processing; else drop segment and return.
11961          */
11962         if ((thflags & TH_ACK) == 0) {
11963                 if (tp->t_flags & TF_NEEDSYN) {
11964                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11965                             tiwin, thflags, nxt_pkt));
11966                 } else if (tp->t_flags & TF_ACKNOW) {
11967                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11968                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11969                         return (ret_val);
11970                 } else {
11971                         ctf_do_drop(m, NULL);
11972                         return (0);
11973                 }
11974         }
11975         /*
11976          * Ack processing.
11977          */
11978         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11979                 return (ret_val);
11980         }
11981         if (ourfinisacked) {
11982                 /*
11983                  * If we can't receive any more data, then closing user can
11984                  * proceed. Starting the timer is contrary to the
11985                  * specification, but if we don't get a FIN we'll hang
11986                  * forever.
11987                  *
11988                  * XXXjl: we should release the tp also, and use a
11989                  * compressed state.
11990                  */
11991                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11992                         soisdisconnected(so);
11993                         tcp_timer_activate(tp, TT_2MSL,
11994                             (tcp_fast_finwait2_recycle ?
11995                             tcp_finwait2_timeout :
11996                             TP_MAXIDLE(tp)));
11997                 }
11998                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
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_closing(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, &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          * 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                 tcp_twstart(tp);
12110                 m_freem(m);
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_lastack(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         if ((thflags & TH_RST) ||
12143             (tp->t_fin_is_rst && (thflags & TH_FIN)))
12144                 return (__ctf_process_rst(m, th, so, tp,
12145                                           &rack->r_ctl.challenge_ack_ts,
12146                                           &rack->r_ctl.challenge_ack_cnt));
12147         /*
12148          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12149          * synchronized state.
12150          */
12151         if (thflags & TH_SYN) {
12152                 ctf_challenge_ack(m, th, tp, &ret_val);
12153                 return (ret_val);
12154         }
12155         /*
12156          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12157          * it's less than ts_recent, drop it.
12158          */
12159         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12160             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12161                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12162                         return (ret_val);
12163         }
12164         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12165                               &rack->r_ctl.challenge_ack_ts,
12166                               &rack->r_ctl.challenge_ack_cnt)) {
12167                 return (ret_val);
12168         }
12169         /*
12170          * If new data are received on a connection after the user processes
12171          * are gone, then RST the other end.
12172          */
12173         if ((tp->t_flags & TF_CLOSED) && tlen &&
12174             rack_check_data_after_close(m, tp, &tlen, th, so))
12175                 return (1);
12176         /*
12177          * If last ACK falls within this segment's sequence numbers, record
12178          * its timestamp. NOTE: 1) That the test incorporates suggestions
12179          * from the latest proposal of the tcplw@cray.com list (Braden
12180          * 1993/04/26). 2) That updating only on newer timestamps interferes
12181          * with our earlier PAWS tests, so this check should be solely
12182          * predicated on the sequence space of this segment. 3) That we
12183          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12184          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12185          * SEG.Len, This modified check allows us to overcome RFC1323's
12186          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12187          * p.869. In such cases, we can still calculate the RTT correctly
12188          * when RCV.NXT == Last.ACK.Sent.
12189          */
12190         if ((to->to_flags & TOF_TS) != 0 &&
12191             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12192             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12193             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12194                 tp->ts_recent_age = tcp_ts_getticks();
12195                 tp->ts_recent = to->to_tsval;
12196         }
12197         /*
12198          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12199          * is on (half-synchronized state), then queue data for later
12200          * processing; else drop segment and return.
12201          */
12202         if ((thflags & TH_ACK) == 0) {
12203                 if (tp->t_flags & TF_NEEDSYN) {
12204                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12205                             tiwin, thflags, nxt_pkt));
12206                 } else if (tp->t_flags & TF_ACKNOW) {
12207                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12208                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12209                         return (ret_val);
12210                 } else {
12211                         ctf_do_drop(m, NULL);
12212                         return (0);
12213                 }
12214         }
12215         /*
12216          * case TCPS_LAST_ACK: Ack processing.
12217          */
12218         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12219                 return (ret_val);
12220         }
12221         if (ourfinisacked) {
12222                 tp = tcp_close(tp);
12223                 ctf_do_drop(m, tp);
12224                 return (1);
12225         }
12226         if (sbavail(&so->so_snd)) {
12227                 if (ctf_progress_timeout_check(tp, true)) {
12228                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12229                                                 tp, tick, PROGRESS_DROP, __LINE__);
12230                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12231                         return (1);
12232                 }
12233         }
12234         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12235             tiwin, thflags, nxt_pkt));
12236 }
12237
12238 /*
12239  * Return value of 1, the TCB is unlocked and most
12240  * likely gone, return value of 0, the TCP is still
12241  * locked.
12242  */
12243 static int
12244 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12245     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12246     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12247 {
12248         int32_t ret_val = 0;
12249         int32_t ourfinisacked = 0;
12250         struct tcp_rack *rack;
12251
12252         rack = (struct tcp_rack *)tp->t_fb_ptr;
12253         ctf_calc_rwin(so, tp);
12254
12255         /* Reset receive buffer auto scaling when not in bulk receive mode. */
12256         if ((thflags & TH_RST) ||
12257             (tp->t_fin_is_rst && (thflags & TH_FIN)))
12258                 return (__ctf_process_rst(m, th, so, tp,
12259                                           &rack->r_ctl.challenge_ack_ts,
12260                                           &rack->r_ctl.challenge_ack_cnt));
12261         /*
12262          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12263          * synchronized state.
12264          */
12265         if (thflags & TH_SYN) {
12266                 ctf_challenge_ack(m, th, tp, &ret_val);
12267                 return (ret_val);
12268         }
12269         /*
12270          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12271          * it's less than ts_recent, drop it.
12272          */
12273         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12274             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12275                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12276                         return (ret_val);
12277         }
12278         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12279                               &rack->r_ctl.challenge_ack_ts,
12280                               &rack->r_ctl.challenge_ack_cnt)) {
12281                 return (ret_val);
12282         }
12283         /*
12284          * If new data are received on a connection after the user processes
12285          * are gone, then RST the other end.
12286          */
12287         if ((tp->t_flags & TF_CLOSED) && tlen &&
12288             rack_check_data_after_close(m, tp, &tlen, th, so))
12289                 return (1);
12290         /*
12291          * If last ACK falls within this segment's sequence numbers, record
12292          * its timestamp. NOTE: 1) That the test incorporates suggestions
12293          * from the latest proposal of the tcplw@cray.com list (Braden
12294          * 1993/04/26). 2) That updating only on newer timestamps interferes
12295          * with our earlier PAWS tests, so this check should be solely
12296          * predicated on the sequence space of this segment. 3) That we
12297          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12298          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12299          * SEG.Len, This modified check allows us to overcome RFC1323's
12300          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12301          * p.869. In such cases, we can still calculate the RTT correctly
12302          * when RCV.NXT == Last.ACK.Sent.
12303          */
12304         if ((to->to_flags & TOF_TS) != 0 &&
12305             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12306             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12307             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12308                 tp->ts_recent_age = tcp_ts_getticks();
12309                 tp->ts_recent = to->to_tsval;
12310         }
12311         /*
12312          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12313          * is on (half-synchronized state), then queue data for later
12314          * processing; else drop segment and return.
12315          */
12316         if ((thflags & TH_ACK) == 0) {
12317                 if (tp->t_flags & TF_NEEDSYN) {
12318                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12319                             tiwin, thflags, nxt_pkt));
12320                 } else if (tp->t_flags & TF_ACKNOW) {
12321                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12322                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12323                         return (ret_val);
12324                 } else {
12325                         ctf_do_drop(m, NULL);
12326                         return (0);
12327                 }
12328         }
12329         /*
12330          * Ack processing.
12331          */
12332         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12333                 return (ret_val);
12334         }
12335         if (sbavail(&so->so_snd)) {
12336                 if (ctf_progress_timeout_check(tp, true)) {
12337                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12338                                                 tp, tick, PROGRESS_DROP, __LINE__);
12339                         ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12340                         return (1);
12341                 }
12342         }
12343         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12344             tiwin, thflags, nxt_pkt));
12345 }
12346
12347 static void inline
12348 rack_clear_rate_sample(struct tcp_rack *rack)
12349 {
12350         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12351         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12352         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12353 }
12354
12355 static void
12356 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12357 {
12358         uint64_t bw_est, rate_wanted;
12359         int chged = 0;
12360         uint32_t user_max, orig_min, orig_max;
12361
12362         orig_min = rack->r_ctl.rc_pace_min_segs;
12363         orig_max = rack->r_ctl.rc_pace_max_segs;
12364         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12365         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12366                 chged = 1;
12367         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12368         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12369                 if (user_max != rack->r_ctl.rc_pace_max_segs)
12370                         chged = 1;
12371         }
12372         if (rack->rc_force_max_seg) {
12373                 rack->r_ctl.rc_pace_max_segs = user_max;
12374         } else if (rack->use_fixed_rate) {
12375                 bw_est = rack_get_bw(rack);
12376                 if ((rack->r_ctl.crte == NULL) ||
12377                     (bw_est != rack->r_ctl.crte->rate)) {
12378                         rack->r_ctl.rc_pace_max_segs = user_max;
12379                 } else {
12380                         /* We are pacing right at the hardware rate */
12381                         uint32_t segsiz;
12382
12383                         segsiz = min(ctf_fixed_maxseg(tp),
12384                                      rack->r_ctl.rc_pace_min_segs);
12385                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12386                                                            tp, bw_est, segsiz, 0,
12387                                                            rack->r_ctl.crte, NULL);
12388                 }
12389         } else if (rack->rc_always_pace) {
12390                 if (rack->r_ctl.gp_bw ||
12391 #ifdef NETFLIX_PEAKRATE
12392                     rack->rc_tp->t_maxpeakrate ||
12393 #endif
12394                     rack->r_ctl.init_rate) {
12395                         /* We have a rate of some sort set */
12396                         uint32_t  orig;
12397
12398                         bw_est = rack_get_bw(rack);
12399                         orig = rack->r_ctl.rc_pace_max_segs;
12400                         if (fill_override)
12401                                 rate_wanted = *fill_override;
12402                         else
12403                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12404                         if (rate_wanted) {
12405                                 /* We have something */
12406                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12407                                                                                    rate_wanted,
12408                                                                                    ctf_fixed_maxseg(rack->rc_tp));
12409                         } else
12410                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12411                         if (orig != rack->r_ctl.rc_pace_max_segs)
12412                                 chged = 1;
12413                 } else if ((rack->r_ctl.gp_bw == 0) &&
12414                            (rack->r_ctl.rc_pace_max_segs == 0)) {
12415                         /*
12416                          * If we have nothing limit us to bursting
12417                          * out IW sized pieces.
12418                          */
12419                         chged = 1;
12420                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12421                 }
12422         }
12423         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12424                 chged = 1;
12425                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12426         }
12427         if (chged)
12428                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12429 }
12430
12431
12432 static void
12433 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12434 {
12435 #ifdef INET6
12436         struct ip6_hdr *ip6 = NULL;
12437 #endif
12438 #ifdef INET
12439         struct ip *ip = NULL;
12440 #endif
12441         struct udphdr *udp = NULL;
12442
12443         /* Ok lets fill in the fast block, it can only be used with no IP options! */
12444 #ifdef INET6
12445         if (rack->r_is_v6) {
12446                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12447                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12448                 if (tp->t_port) {
12449                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12450                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12451                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12452                         udp->uh_dport = tp->t_port;
12453                         rack->r_ctl.fsb.udp = udp;
12454                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12455                 } else
12456                 {
12457                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12458                         rack->r_ctl.fsb.udp = NULL;
12459                 }
12460                 tcpip_fillheaders(rack->rc_inp,
12461                                   tp->t_port,
12462                                   ip6, rack->r_ctl.fsb.th);
12463         } else
12464 #endif                          /* INET6 */
12465         {
12466                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12467                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12468                 if (tp->t_port) {
12469                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12470                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12471                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12472                         udp->uh_dport = tp->t_port;
12473                         rack->r_ctl.fsb.udp = udp;
12474                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12475                 } else
12476                 {
12477                         rack->r_ctl.fsb.udp = NULL;
12478                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12479                 }
12480                 tcpip_fillheaders(rack->rc_inp,
12481                                   tp->t_port,
12482                                   ip, rack->r_ctl.fsb.th);
12483         }
12484         rack->r_fsb_inited = 1;
12485 }
12486
12487 static int
12488 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12489 {
12490         /*
12491          * Allocate the larger of spaces V6 if available else just
12492          * V4 and include udphdr (overbook)
12493          */
12494 #ifdef INET6
12495         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12496 #else
12497         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12498 #endif
12499         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12500                                             M_TCPFSB, M_NOWAIT|M_ZERO);
12501         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12502                 return (ENOMEM);
12503         }
12504         rack->r_fsb_inited = 0;
12505         return (0);
12506 }
12507
12508 static int
12509 rack_init(struct tcpcb *tp)
12510 {
12511         struct tcp_rack *rack = NULL;
12512 #ifdef INVARIANTS
12513         struct rack_sendmap *insret;
12514 #endif
12515         uint32_t iwin, snt, us_cts;
12516         int err;
12517
12518         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12519         if (tp->t_fb_ptr == NULL) {
12520                 /*
12521                  * We need to allocate memory but cant. The INP and INP_INFO
12522                  * locks and they are recursive (happens during setup. So a
12523                  * scheme to drop the locks fails :(
12524                  *
12525                  */
12526                 return (ENOMEM);
12527         }
12528         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12529
12530         rack = (struct tcp_rack *)tp->t_fb_ptr;
12531         RB_INIT(&rack->r_ctl.rc_mtree);
12532         TAILQ_INIT(&rack->r_ctl.rc_free);
12533         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12534         rack->rc_tp = tp;
12535         rack->rc_inp = tp->t_inpcb;
12536         /* Set the flag */
12537         rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12538         /* Probably not needed but lets be sure */
12539         rack_clear_rate_sample(rack);
12540         /*
12541          * Save off the default values, socket options will poke
12542          * at these if pacing is not on or we have not yet
12543          * reached where pacing is on (gp_ready/fixed enabled).
12544          * When they get set into the CC module (when gp_ready
12545          * is enabled or we enable fixed) then we will set these
12546          * values into the CC and place in here the old values
12547          * so we have a restoral. Then we will set the flag
12548          * rc_pacing_cc_set. That way whenever we turn off pacing
12549          * or switch off this stack, we will know to go restore
12550          * the saved values.
12551          */
12552         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12553         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12554         /* We want abe like behavior as well */
12555         rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12556         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12557         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12558         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12559         rack->r_ctl.roundends = tp->snd_max;
12560         if (use_rack_rr)
12561                 rack->use_rack_rr = 1;
12562         if (V_tcp_delack_enabled)
12563                 tp->t_delayed_ack = 1;
12564         else
12565                 tp->t_delayed_ack = 0;
12566 #ifdef TCP_ACCOUNTING
12567         if (rack_tcp_accounting) {
12568                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12569         }
12570 #endif
12571         if (rack_enable_shared_cwnd)
12572                 rack->rack_enable_scwnd = 1;
12573         rack->rc_user_set_max_segs = rack_hptsi_segments;
12574         rack->rc_force_max_seg = 0;
12575         if (rack_use_imac_dack)
12576                 rack->rc_dack_mode = 1;
12577         TAILQ_INIT(&rack->r_ctl.opt_list);
12578         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12579         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12580         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12581         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12582         rack->r_ctl.rc_highest_us_rtt = 0;
12583         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12584         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12585         if (rack_use_cmp_acks)
12586                 rack->r_use_cmp_ack = 1;
12587         if (rack_disable_prr)
12588                 rack->rack_no_prr = 1;
12589         if (rack_gp_no_rec_chg)
12590                 rack->rc_gp_no_rec_chg = 1;
12591         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12592                 rack->rc_always_pace = 1;
12593                 if (rack->use_fixed_rate || rack->gp_ready)
12594                         rack_set_cc_pacing(rack);
12595         } else
12596                 rack->rc_always_pace = 0;
12597         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12598                 rack->r_mbuf_queue = 1;
12599         else
12600                 rack->r_mbuf_queue = 0;
12601         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12602                 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12603         else
12604                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12605         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12606         if (rack_limits_scwnd)
12607                 rack->r_limit_scw = 1;
12608         else
12609                 rack->r_limit_scw = 0;
12610         rack->rc_labc = V_tcp_abc_l_var;
12611         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12612         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12613         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12614         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12615         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12616         rack->r_ctl.rc_min_to = rack_min_to;
12617         microuptime(&rack->r_ctl.act_rcv_time);
12618         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12619         rack->rc_init_win = rack_default_init_window;
12620         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12621         if (rack_hw_up_only)
12622                 rack->r_up_only = 1;
12623         if (rack_do_dyn_mul) {
12624                 /* When dynamic adjustment is on CA needs to start at 100% */
12625                 rack->rc_gp_dyn_mul = 1;
12626                 if (rack_do_dyn_mul >= 100)
12627                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12628         } else
12629                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12630         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12631         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12632         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12633         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12634                                 rack_probertt_filter_life);
12635         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12636         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12637         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12638         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12639         rack->r_ctl.rc_time_probertt_starts = 0;
12640         if (rack_dsack_std_based & 0x1) {
12641                 /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12642                 rack->rc_rack_tmr_std_based = 1;
12643         }
12644         if (rack_dsack_std_based & 0x2) {
12645                 /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12646                 rack->rc_rack_use_dsack = 1;
12647         }
12648         /* We require at least one measurement, even if the sysctl is 0 */
12649         if (rack_req_measurements)
12650                 rack->r_ctl.req_measurements = rack_req_measurements;
12651         else
12652                 rack->r_ctl.req_measurements = 1;
12653         if (rack_enable_hw_pacing)
12654                 rack->rack_hdw_pace_ena = 1;
12655         if (rack_hw_rate_caps)
12656                 rack->r_rack_hw_rate_caps = 1;
12657         /* Do we force on detection? */
12658 #ifdef NETFLIX_EXP_DETECTION
12659         if (tcp_force_detection)
12660                 rack->do_detection = 1;
12661         else
12662 #endif
12663                 rack->do_detection = 0;
12664         if (rack_non_rxt_use_cr)
12665                 rack->rack_rec_nonrxt_use_cr = 1;
12666         err = rack_init_fsb(tp, rack);
12667         if (err) {
12668                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12669                 tp->t_fb_ptr = NULL;
12670                 return (err);
12671         }
12672         if (tp->snd_una != tp->snd_max) {
12673                 /* Create a send map for the current outstanding data */
12674                 struct rack_sendmap *rsm;
12675
12676                 rsm = rack_alloc(rack);
12677                 if (rsm == NULL) {
12678                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12679                         tp->t_fb_ptr = NULL;
12680                         return (ENOMEM);
12681                 }
12682                 rsm->r_no_rtt_allowed = 1;
12683                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12684                 rsm->r_rtr_cnt = 1;
12685                 rsm->r_rtr_bytes = 0;
12686                 if (tp->t_flags & TF_SENTFIN)
12687                         rsm->r_flags |= RACK_HAS_FIN;
12688                 if ((tp->snd_una == tp->iss) &&
12689                     !TCPS_HAVEESTABLISHED(tp->t_state))
12690                         rsm->r_flags |= RACK_HAS_SYN;
12691                 rsm->r_start = tp->snd_una;
12692                 rsm->r_end = tp->snd_max;
12693                 rsm->r_dupack = 0;
12694                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12695                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12696                         if (rsm->m)
12697                                 rsm->orig_m_len = rsm->m->m_len;
12698                         else
12699                                 rsm->orig_m_len = 0;
12700                 } else {
12701                         /*
12702                          * This can happen if we have a stand-alone FIN or
12703                          *  SYN.
12704                          */
12705                         rsm->m = NULL;
12706                         rsm->orig_m_len = 0;
12707                         rsm->soff = 0;
12708                 }
12709 #ifndef INVARIANTS
12710                 (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12711 #else
12712                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12713                 if (insret != NULL) {
12714                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12715                               insret, rack, rsm);
12716                 }
12717 #endif
12718                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12719                 rsm->r_in_tmap = 1;
12720         }
12721         /*
12722          * Timers in Rack are kept in microseconds so lets
12723          * convert any initial incoming variables
12724          * from ticks into usecs. Note that we
12725          * also change the values of t_srtt and t_rttvar, if
12726          * they are non-zero. They are kept with a 5
12727          * bit decimal so we have to carefully convert
12728          * these to get the full precision.
12729          */
12730         rack_convert_rtts(tp);
12731         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12732         if (rack_do_hystart) {
12733                 tp->ccv->flags |= CCF_HYSTART_ALLOWED;
12734                 if (rack_do_hystart > 1)
12735                         tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND;
12736                 if (rack_do_hystart > 2)
12737                         tp->ccv->flags |= CCF_HYSTART_CONS_SSTH;
12738         }
12739         if (rack_def_profile)
12740                 rack_set_profile(rack, rack_def_profile);
12741         /* Cancel the GP measurement in progress */
12742         tp->t_flags &= ~TF_GPUTINPROG;
12743         if (SEQ_GT(tp->snd_max, tp->iss))
12744                 snt = tp->snd_max - tp->iss;
12745         else
12746                 snt = 0;
12747         iwin = rc_init_window(rack);
12748         if (snt < iwin) {
12749                 /* We are not past the initial window
12750                  * so we need to make sure cwnd is
12751                  * correct.
12752                  */
12753                 if (tp->snd_cwnd < iwin)
12754                         tp->snd_cwnd = iwin;
12755                 /*
12756                  * If we are within the initial window
12757                  * we want ssthresh to be unlimited. Setting
12758                  * it to the rwnd (which the default stack does
12759                  * and older racks) is not really a good idea
12760                  * since we want to be in SS and grow both the
12761                  * cwnd and the rwnd (via dynamic rwnd growth). If
12762                  * we set it to the rwnd then as the peer grows its
12763                  * rwnd we will be stuck in CA and never hit SS.
12764                  *
12765                  * Its far better to raise it up high (this takes the
12766                  * risk that there as been a loss already, probably
12767                  * we should have an indicator in all stacks of loss
12768                  * but we don't), but considering the normal use this
12769                  * is a risk worth taking. The consequences of not
12770                  * hitting SS are far worse than going one more time
12771                  * into it early on (before we have sent even a IW).
12772                  * It is highly unlikely that we will have had a loss
12773                  * before getting the IW out.
12774                  */
12775                 tp->snd_ssthresh = 0xffffffff;
12776         }
12777         rack_stop_all_timers(tp);
12778         /* Lets setup the fsb block */
12779         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12780         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12781                              __LINE__, RACK_RTTS_INIT);
12782         return (0);
12783 }
12784
12785 static int
12786 rack_handoff_ok(struct tcpcb *tp)
12787 {
12788         if ((tp->t_state == TCPS_CLOSED) ||
12789             (tp->t_state == TCPS_LISTEN)) {
12790                 /* Sure no problem though it may not stick */
12791                 return (0);
12792         }
12793         if ((tp->t_state == TCPS_SYN_SENT) ||
12794             (tp->t_state == TCPS_SYN_RECEIVED)) {
12795                 /*
12796                  * We really don't know if you support sack,
12797                  * you have to get to ESTAB or beyond to tell.
12798                  */
12799                 return (EAGAIN);
12800         }
12801         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12802                 /*
12803                  * Rack will only send a FIN after all data is acknowledged.
12804                  * So in this case we have more data outstanding. We can't
12805                  * switch stacks until either all data and only the FIN
12806                  * is left (in which case rack_init() now knows how
12807                  * to deal with that) <or> all is acknowledged and we
12808                  * are only left with incoming data, though why you
12809                  * would want to switch to rack after all data is acknowledged
12810                  * I have no idea (rrs)!
12811                  */
12812                 return (EAGAIN);
12813         }
12814         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12815                 return (0);
12816         }
12817         /*
12818          * If we reach here we don't do SACK on this connection so we can
12819          * never do rack.
12820          */
12821         return (EINVAL);
12822 }
12823
12824
12825 static void
12826 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12827 {
12828         if (tp->t_fb_ptr) {
12829                 struct tcp_rack *rack;
12830                 struct rack_sendmap *rsm, *nrsm;
12831 #ifdef INVARIANTS
12832                 struct rack_sendmap *rm;
12833 #endif
12834
12835                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12836                 if (tp->t_in_pkt) {
12837                         /*
12838                          * It is unsafe to process the packets since a
12839                          * reset may be lurking in them (its rare but it
12840                          * can occur). If we were to find a RST, then we
12841                          * would end up dropping the connection and the
12842                          * INP lock, so when we return the caller (tcp_usrreq)
12843                          * will blow up when it trys to unlock the inp.
12844                          */
12845                         struct mbuf *save, *m;
12846
12847                         m = tp->t_in_pkt;
12848                         tp->t_in_pkt = NULL;
12849                         tp->t_tail_pkt = NULL;
12850                         while (m) {
12851                                 save = m->m_nextpkt;
12852                                 m->m_nextpkt = NULL;
12853                                 m_freem(m);
12854                                 m = save;
12855                         }
12856                 }
12857                 tp->t_flags &= ~TF_FORCEDATA;
12858 #ifdef NETFLIX_SHARED_CWND
12859                 if (rack->r_ctl.rc_scw) {
12860                         uint32_t limit;
12861
12862                         if (rack->r_limit_scw)
12863                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12864                         else
12865                                 limit = 0;
12866                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12867                                                   rack->r_ctl.rc_scw_index,
12868                                                   limit);
12869                         rack->r_ctl.rc_scw = NULL;
12870                 }
12871 #endif
12872                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12873                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12874                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12875                         rack->r_ctl.fsb.th = NULL;
12876                 }
12877                 /* Convert back to ticks, with  */
12878                 if (tp->t_srtt > 1) {
12879                         uint32_t val, frac;
12880
12881                         val = USEC_2_TICKS(tp->t_srtt);
12882                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12883                         tp->t_srtt = val << TCP_RTT_SHIFT;
12884                         /*
12885                          * frac is the fractional part here is left
12886                          * over from converting to hz and shifting.
12887                          * We need to convert this to the 5 bit
12888                          * remainder.
12889                          */
12890                         if (frac) {
12891                                 if (hz == 1000) {
12892                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12893                                 } else {
12894                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12895                                 }
12896                                 tp->t_srtt += frac;
12897                         }
12898                 }
12899                 if (tp->t_rttvar) {
12900                         uint32_t val, frac;
12901
12902                         val = USEC_2_TICKS(tp->t_rttvar);
12903                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12904                         tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12905                         /*
12906                          * frac is the fractional part here is left
12907                          * over from converting to hz and shifting.
12908                          * We need to convert this to the 5 bit
12909                          * remainder.
12910                          */
12911                         if (frac) {
12912                                 if (hz == 1000) {
12913                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12914                                 } else {
12915                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12916                                 }
12917                                 tp->t_rttvar += frac;
12918                         }
12919                 }
12920                 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12921                 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12922                 if (rack->rc_always_pace) {
12923                         tcp_decrement_paced_conn();
12924                         rack_undo_cc_pacing(rack);
12925                         rack->rc_always_pace = 0;
12926                 }
12927                 /* Clean up any options if they were not applied */
12928                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12929                         struct deferred_opt_list *dol;
12930
12931                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12932                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12933                         free(dol, M_TCPDO);
12934                 }
12935                 /* rack does not use force data but other stacks may clear it */
12936                 if (rack->r_ctl.crte != NULL) {
12937                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12938                         rack->rack_hdrw_pacing = 0;
12939                         rack->r_ctl.crte = NULL;
12940                 }
12941 #ifdef TCP_BLACKBOX
12942                 tcp_log_flowend(tp);
12943 #endif
12944                 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12945 #ifndef INVARIANTS
12946                         (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12947 #else
12948                         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12949                         if (rm != rsm) {
12950                                 panic("At fini, rack:%p rsm:%p rm:%p",
12951                                       rack, rsm, rm);
12952                         }
12953 #endif
12954                         uma_zfree(rack_zone, rsm);
12955                 }
12956                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12957                 while (rsm) {
12958                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12959                         uma_zfree(rack_zone, rsm);
12960                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12961                 }
12962                 rack->rc_free_cnt = 0;
12963                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12964                 tp->t_fb_ptr = NULL;
12965         }
12966         if (tp->t_inpcb) {
12967                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12968                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12969                 tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12970                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
12971                 /* Cancel the GP measurement in progress */
12972                 tp->t_flags &= ~TF_GPUTINPROG;
12973                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
12974         }
12975         /* Make sure snd_nxt is correctly set */
12976         tp->snd_nxt = tp->snd_max;
12977 }
12978
12979 static void
12980 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12981 {
12982         if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12983                 rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12984         }
12985         switch (tp->t_state) {
12986         case TCPS_SYN_SENT:
12987                 rack->r_state = TCPS_SYN_SENT;
12988                 rack->r_substate = rack_do_syn_sent;
12989                 break;
12990         case TCPS_SYN_RECEIVED:
12991                 rack->r_state = TCPS_SYN_RECEIVED;
12992                 rack->r_substate = rack_do_syn_recv;
12993                 break;
12994         case TCPS_ESTABLISHED:
12995                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12996                 rack->r_state = TCPS_ESTABLISHED;
12997                 rack->r_substate = rack_do_established;
12998                 break;
12999         case TCPS_CLOSE_WAIT:
13000                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13001                 rack->r_state = TCPS_CLOSE_WAIT;
13002                 rack->r_substate = rack_do_close_wait;
13003                 break;
13004         case TCPS_FIN_WAIT_1:
13005                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13006                 rack->r_state = TCPS_FIN_WAIT_1;
13007                 rack->r_substate = rack_do_fin_wait_1;
13008                 break;
13009         case TCPS_CLOSING:
13010                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13011                 rack->r_state = TCPS_CLOSING;
13012                 rack->r_substate = rack_do_closing;
13013                 break;
13014         case TCPS_LAST_ACK:
13015                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13016                 rack->r_state = TCPS_LAST_ACK;
13017                 rack->r_substate = rack_do_lastack;
13018                 break;
13019         case TCPS_FIN_WAIT_2:
13020                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13021                 rack->r_state = TCPS_FIN_WAIT_2;
13022                 rack->r_substate = rack_do_fin_wait_2;
13023                 break;
13024         case TCPS_LISTEN:
13025         case TCPS_CLOSED:
13026         case TCPS_TIME_WAIT:
13027         default:
13028                 break;
13029         };
13030         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
13031                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
13032
13033 }
13034
13035 static void
13036 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
13037 {
13038         /*
13039          * We received an ack, and then did not
13040          * call send or were bounced out due to the
13041          * hpts was running. Now a timer is up as well, is
13042          * it the right timer?
13043          */
13044         struct rack_sendmap *rsm;
13045         int tmr_up;
13046
13047         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
13048         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
13049                 return;
13050         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
13051         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
13052             (tmr_up == PACE_TMR_RXT)) {
13053                 /* Should be an RXT */
13054                 return;
13055         }
13056         if (rsm == NULL) {
13057                 /* Nothing outstanding? */
13058                 if (tp->t_flags & TF_DELACK) {
13059                         if (tmr_up == PACE_TMR_DELACK)
13060                                 /* We are supposed to have delayed ack up and we do */
13061                                 return;
13062                 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
13063                         /*
13064                          * if we hit enobufs then we would expect the possibility
13065                          * of nothing outstanding and the RXT up (and the hptsi timer).
13066                          */
13067                         return;
13068                 } else if (((V_tcp_always_keepalive ||
13069                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13070                             (tp->t_state <= TCPS_CLOSING)) &&
13071                            (tmr_up == PACE_TMR_KEEP) &&
13072                            (tp->snd_max == tp->snd_una)) {
13073                         /* We should have keep alive up and we do */
13074                         return;
13075                 }
13076         }
13077         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
13078                    ((tmr_up == PACE_TMR_TLP) ||
13079                     (tmr_up == PACE_TMR_RACK) ||
13080                     (tmr_up == PACE_TMR_RXT))) {
13081                 /*
13082                  * Either a Rack, TLP or RXT is fine if  we
13083                  * have outstanding data.
13084                  */
13085                 return;
13086         } else if (tmr_up == PACE_TMR_DELACK) {
13087                 /*
13088                  * If the delayed ack was going to go off
13089                  * before the rtx/tlp/rack timer were going to
13090                  * expire, then that would be the timer in control.
13091                  * Note we don't check the time here trusting the
13092                  * code is correct.
13093                  */
13094                 return;
13095         }
13096         /*
13097          * Ok the timer originally started is not what we want now.
13098          * We will force the hpts to be stopped if any, and restart
13099          * with the slot set to what was in the saved slot.
13100          */
13101         if (tcp_in_hpts(rack->rc_inp)) {
13102                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13103                         uint32_t us_cts;
13104
13105                         us_cts = tcp_get_usecs(NULL);
13106                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13107                                 rack->r_early = 1;
13108                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13109                         }
13110                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13111                 }
13112                 tcp_hpts_remove(tp->t_inpcb);
13113         }
13114         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13115         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13116 }
13117
13118
13119 static void
13120 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)
13121 {
13122         if ((SEQ_LT(tp->snd_wl1, seq) ||
13123             (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
13124             (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
13125                 /* keep track of pure window updates */
13126                 if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
13127                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13128                 tp->snd_wnd = tiwin;
13129                 rack_validate_fo_sendwin_up(tp, rack);
13130                 tp->snd_wl1 = seq;
13131                 tp->snd_wl2 = ack;
13132                 if (tp->snd_wnd > tp->max_sndwnd)
13133                         tp->max_sndwnd = tp->snd_wnd;
13134             rack->r_wanted_output = 1;
13135         } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
13136                 tp->snd_wnd = tiwin;
13137                 rack_validate_fo_sendwin_up(tp, rack);
13138                 tp->snd_wl1 = seq;
13139                 tp->snd_wl2 = ack;
13140         } else {
13141                 /* Not a valid win update */
13142                 return;
13143         }
13144         /* Do we exit persists? */
13145         if ((rack->rc_in_persist != 0) &&
13146             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13147                                 rack->r_ctl.rc_pace_min_segs))) {
13148                 rack_exit_persist(tp, rack, cts);
13149         }
13150         /* Do we enter persists? */
13151         if ((rack->rc_in_persist == 0) &&
13152             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13153             TCPS_HAVEESTABLISHED(tp->t_state) &&
13154             ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13155             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
13156             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
13157                 /*
13158                  * Here the rwnd is less than
13159                  * the pacing size, we are established,
13160                  * nothing is outstanding, and there is
13161                  * data to send. Enter persists.
13162                  */
13163                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13164         }
13165 }
13166
13167 static void
13168 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13169 {
13170
13171         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13172                 union tcp_log_stackspecific log;
13173                 struct timeval ltv;
13174                 char tcp_hdr_buf[60];
13175                 struct tcphdr *th;
13176                 struct timespec ts;
13177                 uint32_t orig_snd_una;
13178                 uint8_t xx = 0;
13179
13180 #ifdef NETFLIX_HTTP_LOGGING
13181                 struct http_sendfile_track *http_req;
13182
13183                 if (SEQ_GT(ae->ack, tp->snd_una)) {
13184                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13185                 } else {
13186                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13187                 }
13188 #endif
13189                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13190                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13191                 if (rack->rack_no_prr == 0)
13192                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13193                 else
13194                         log.u_bbr.flex1 = 0;
13195                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13196                 log.u_bbr.use_lt_bw <<= 1;
13197                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13198                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13199                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13200                 log.u_bbr.pkts_out = tp->t_maxseg;
13201                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13202                 log.u_bbr.flex7 = 1;
13203                 log.u_bbr.lost = ae->flags;
13204                 log.u_bbr.cwnd_gain = ackval;
13205                 log.u_bbr.pacing_gain = 0x2;
13206                 if (ae->flags & TSTMP_HDWR) {
13207                         /* Record the hardware timestamp if present */
13208                         log.u_bbr.flex3 = M_TSTMP;
13209                         ts.tv_sec = ae->timestamp / 1000000000;
13210                         ts.tv_nsec = ae->timestamp % 1000000000;
13211                         ltv.tv_sec = ts.tv_sec;
13212                         ltv.tv_usec = ts.tv_nsec / 1000;
13213                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13214                 } else if (ae->flags & TSTMP_LRO) {
13215                         /* Record the LRO the arrival timestamp */
13216                         log.u_bbr.flex3 = M_TSTMP_LRO;
13217                         ts.tv_sec = ae->timestamp / 1000000000;
13218                         ts.tv_nsec = ae->timestamp % 1000000000;
13219                         ltv.tv_sec = ts.tv_sec;
13220                         ltv.tv_usec = ts.tv_nsec / 1000;
13221                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13222                 }
13223                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13224                 /* Log the rcv time */
13225                 log.u_bbr.delRate = ae->timestamp;
13226 #ifdef NETFLIX_HTTP_LOGGING
13227                 log.u_bbr.applimited = tp->t_http_closed;
13228                 log.u_bbr.applimited <<= 8;
13229                 log.u_bbr.applimited |= tp->t_http_open;
13230                 log.u_bbr.applimited <<= 8;
13231                 log.u_bbr.applimited |= tp->t_http_req;
13232                 if (http_req) {
13233                         /* Copy out any client req info */
13234                         /* seconds */
13235                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13236                         /* useconds */
13237                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13238                         log.u_bbr.rttProp = http_req->timestamp;
13239                         log.u_bbr.cur_del_rate = http_req->start;
13240                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13241                                 log.u_bbr.flex8 |= 1;
13242                         } else {
13243                                 log.u_bbr.flex8 |= 2;
13244                                 log.u_bbr.bw_inuse = http_req->end;
13245                         }
13246                         log.u_bbr.flex6 = http_req->start_seq;
13247                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13248                                 log.u_bbr.flex8 |= 4;
13249                                 log.u_bbr.epoch = http_req->end_seq;
13250                         }
13251                 }
13252 #endif
13253                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13254                 th = (struct tcphdr *)tcp_hdr_buf;
13255                 th->th_seq = ae->seq;
13256                 th->th_ack = ae->ack;
13257                 th->th_win = ae->win;
13258                 /* Now fill in the ports */
13259                 th->th_sport = tp->t_inpcb->inp_fport;
13260                 th->th_dport = tp->t_inpcb->inp_lport;
13261                 tcp_set_flags(th, ae->flags);
13262                 /* Now do we have a timestamp option? */
13263                 if (ae->flags & HAS_TSTMP) {
13264                         u_char *cp;
13265                         uint32_t val;
13266
13267                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13268                         cp = (u_char *)(th + 1);
13269                         *cp = TCPOPT_NOP;
13270                         cp++;
13271                         *cp = TCPOPT_NOP;
13272                         cp++;
13273                         *cp = TCPOPT_TIMESTAMP;
13274                         cp++;
13275                         *cp = TCPOLEN_TIMESTAMP;
13276                         cp++;
13277                         val = htonl(ae->ts_value);
13278                         bcopy((char *)&val,
13279                               (char *)cp, sizeof(uint32_t));
13280                         val = htonl(ae->ts_echo);
13281                         bcopy((char *)&val,
13282                               (char *)(cp + 4), sizeof(uint32_t));
13283                 } else
13284                         th->th_off = (sizeof(struct tcphdr) >> 2);
13285
13286                 /*
13287                  * For sane logging we need to play a little trick.
13288                  * If the ack were fully processed we would have moved
13289                  * snd_una to high_seq, but since compressed acks are
13290                  * processed in two phases, at this point (logging) snd_una
13291                  * won't be advanced. So we would see multiple acks showing
13292                  * the advancement. We can prevent that by "pretending" that
13293                  * snd_una was advanced and then un-advancing it so that the
13294                  * logging code has the right value for tlb_snd_una.
13295                  */
13296                 if (tp->snd_una != high_seq) {
13297                         orig_snd_una = tp->snd_una;
13298                         tp->snd_una = high_seq;
13299                         xx = 1;
13300                 } else
13301                         xx = 0;
13302                 TCP_LOG_EVENTP(tp, th,
13303                                &tp->t_inpcb->inp_socket->so_rcv,
13304                                &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
13305                                0, &log, true, &ltv);
13306                 if (xx) {
13307                         tp->snd_una = orig_snd_una;
13308                 }
13309         }
13310
13311 }
13312
13313 static void
13314 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13315 {
13316         uint32_t us_rtt;
13317         /*
13318          * A persist or keep-alive was forced out, update our
13319          * min rtt time. Note now worry about lost responses.
13320          * When a subsequent keep-alive or persist times out
13321          * and forced_ack is still on, then the last probe
13322          * was not responded to. In such cases we have a
13323          * sysctl that controls the behavior. Either we apply
13324          * the rtt but with reduced confidence (0). Or we just
13325          * plain don't apply the rtt estimate. Having data flow
13326          * will clear the probe_not_answered flag i.e. cum-ack
13327          * move forward <or> exiting and reentering persists.
13328          */
13329
13330         rack->forced_ack = 0;
13331         rack->rc_tp->t_rxtshift = 0;
13332         if ((rack->rc_in_persist &&
13333              (tiwin == rack->rc_tp->snd_wnd)) ||
13334             (rack->rc_in_persist == 0)) {
13335                 /*
13336                  * In persists only apply the RTT update if this is
13337                  * a response to our window probe. And that
13338                  * means the rwnd sent must match the current
13339                  * snd_wnd. If it does not, then we got a
13340                  * window update ack instead. For keepalive
13341                  * we allow the answer no matter what the window.
13342                  *
13343                  * Note that if the probe_not_answered is set then
13344                  * the forced_ack_ts is the oldest one i.e. the first
13345                  * probe sent that might have been lost. This assures
13346                  * us that if we do calculate an RTT it is longer not
13347                  * some short thing.
13348                  */
13349                 if (rack->rc_in_persist)
13350                         counter_u64_add(rack_persists_acks, 1);
13351                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13352                 if (us_rtt == 0)
13353                         us_rtt = 1;
13354                 if (rack->probe_not_answered == 0) {
13355                         rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13356                         tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13357                 } else {
13358                         /* We have a retransmitted probe here too */
13359                         if (rack_apply_rtt_with_reduced_conf) {
13360                                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13361                                 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13362                         }
13363                 }
13364         }
13365 }
13366
13367 static int
13368 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13369 {
13370         /*
13371          * Handle a "special" compressed ack mbuf. Each incoming
13372          * ack has only four possible dispositions:
13373          *
13374          * A) It moves the cum-ack forward
13375          * B) It is behind the cum-ack.
13376          * C) It is a window-update ack.
13377          * D) It is a dup-ack.
13378          *
13379          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13380          * in the incoming mbuf. We also need to still pay attention
13381          * to nxt_pkt since there may be another packet after this
13382          * one.
13383          */
13384 #ifdef TCP_ACCOUNTING
13385         uint64_t ts_val;
13386         uint64_t rdstc;
13387 #endif
13388         int segsiz;
13389         struct timespec ts;
13390         struct tcp_rack *rack;
13391         struct tcp_ackent *ae;
13392         uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13393         int cnt, i, did_out, ourfinisacked = 0;
13394         struct tcpopt to_holder, *to = NULL;
13395 #ifdef TCP_ACCOUNTING
13396         int win_up_req = 0;
13397 #endif
13398         int nsegs = 0;
13399         int under_pacing = 1;
13400         int recovery = 0;
13401 #ifdef TCP_ACCOUNTING
13402         sched_pin();
13403 #endif
13404         rack = (struct tcp_rack *)tp->t_fb_ptr;
13405         if (rack->gp_ready &&
13406             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13407                 under_pacing = 0;
13408         else
13409                 under_pacing = 1;
13410
13411         if (rack->r_state != tp->t_state)
13412                 rack_set_state(tp, rack);
13413         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13414             (tp->t_flags & TF_GPUTINPROG)) {
13415                 /*
13416                  * We have a goodput in progress
13417                  * and we have entered a late state.
13418                  * Do we have enough data in the sb
13419                  * to handle the GPUT request?
13420                  */
13421                 uint32_t bytes;
13422
13423                 bytes = tp->gput_ack - tp->gput_seq;
13424                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
13425                         bytes += tp->gput_seq - tp->snd_una;
13426                 if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
13427                         /*
13428                          * There are not enough bytes in the socket
13429                          * buffer that have been sent to cover this
13430                          * measurement. Cancel it.
13431                          */
13432                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13433                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
13434                                                    tp->gput_seq,
13435                                                    0, 0, 18, __LINE__, NULL, 0);
13436                         tp->t_flags &= ~TF_GPUTINPROG;
13437                 }
13438         }
13439         to = &to_holder;
13440         to->to_flags = 0;
13441         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13442                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13443         cnt = m->m_len / sizeof(struct tcp_ackent);
13444         counter_u64_add(rack_multi_single_eq, cnt);
13445         high_seq = tp->snd_una;
13446         the_win = tp->snd_wnd;
13447         win_seq = tp->snd_wl1;
13448         win_upd_ack = tp->snd_wl2;
13449         cts = tcp_tv_to_usectick(tv);
13450         ms_cts = tcp_tv_to_mssectick(tv);
13451         rack->r_ctl.rc_rcvtime = cts;
13452         segsiz = ctf_fixed_maxseg(tp);
13453         if ((rack->rc_gp_dyn_mul) &&
13454             (rack->use_fixed_rate == 0) &&
13455             (rack->rc_always_pace)) {
13456                 /* Check in on probertt */
13457                 rack_check_probe_rtt(rack, cts);
13458         }
13459         for (i = 0; i < cnt; i++) {
13460 #ifdef TCP_ACCOUNTING
13461                 ts_val = get_cyclecount();
13462 #endif
13463                 rack_clear_rate_sample(rack);
13464                 ae = ((mtod(m, struct tcp_ackent *)) + i);
13465                 /* Setup the window */
13466                 tiwin = ae->win << tp->snd_scale;
13467                 if (tiwin > rack->r_ctl.rc_high_rwnd)
13468                         rack->r_ctl.rc_high_rwnd = tiwin;
13469                 /* figure out the type of ack */
13470                 if (SEQ_LT(ae->ack, high_seq)) {
13471                         /* Case B*/
13472                         ae->ack_val_set = ACK_BEHIND;
13473                 } else if (SEQ_GT(ae->ack, high_seq)) {
13474                         /* Case A */
13475                         ae->ack_val_set = ACK_CUMACK;
13476                 } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13477                         /* Case D */
13478                         ae->ack_val_set = ACK_DUPACK;
13479                 } else {
13480                         /* Case C */
13481                         ae->ack_val_set = ACK_RWND;
13482                 }
13483                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13484                 /* Validate timestamp */
13485                 if (ae->flags & HAS_TSTMP) {
13486                         /* Setup for a timestamp */
13487                         to->to_flags = TOF_TS;
13488                         ae->ts_echo -= tp->ts_offset;
13489                         to->to_tsecr = ae->ts_echo;
13490                         to->to_tsval = ae->ts_value;
13491                         /*
13492                          * If echoed timestamp is later than the current time, fall back to
13493                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13494                          * were used when this connection was established.
13495                          */
13496                         if (TSTMP_GT(ae->ts_echo, ms_cts))
13497                                 to->to_tsecr = 0;
13498                         if (tp->ts_recent &&
13499                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13500                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13501 #ifdef TCP_ACCOUNTING
13502                                         rdstc = get_cyclecount();
13503                                         if (rdstc > ts_val) {
13504                                                 counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13505                                                                 (rdstc - ts_val));
13506                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13507                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13508                                                 }
13509                                         }
13510 #endif
13511                                         continue;
13512                                 }
13513                         }
13514                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13515                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13516                                 tp->ts_recent_age = tcp_ts_getticks();
13517                                 tp->ts_recent = ae->ts_value;
13518                         }
13519                 } else {
13520                         /* Setup for a no options */
13521                         to->to_flags = 0;
13522                 }
13523                 /* Update the rcv time and perform idle reduction possibly */
13524                 if  (tp->t_idle_reduce &&
13525                      (tp->snd_max == tp->snd_una) &&
13526                      (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13527                         counter_u64_add(rack_input_idle_reduces, 1);
13528                         rack_cc_after_idle(rack, tp);
13529                 }
13530                 tp->t_rcvtime = ticks;
13531                 /* Now what about ECN? */
13532                 if (tcp_ecn_input_segment(tp, ae->flags, ae->codepoint))
13533                         rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
13534 #ifdef TCP_ACCOUNTING
13535                 /* Count for the specific type of ack in */
13536                 counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13537                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13538                         tp->tcp_cnt_counters[ae->ack_val_set]++;
13539                 }
13540 #endif
13541                 /*
13542                  * Note how we could move up these in the determination
13543                  * above, but we don't so that way the timestamp checks (and ECN)
13544                  * is done first before we do any processing on the ACK.
13545                  * The non-compressed path through the code has this
13546                  * weakness (noted by @jtl) that it actually does some
13547                  * processing before verifying the timestamp information.
13548                  * We don't take that path here which is why we set
13549                  * the ack_val_set first, do the timestamp and ecn
13550                  * processing, and then look at what we have setup.
13551                  */
13552                 if (ae->ack_val_set == ACK_BEHIND) {
13553                         /*
13554                          * Case B flag reordering, if window is not closed
13555                          * or it could be a keep-alive or persists
13556                          */
13557                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13558                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13559                         }
13560                 } else if (ae->ack_val_set == ACK_DUPACK) {
13561                         /* Case D */
13562                         rack_strike_dupack(rack);
13563                 } else if (ae->ack_val_set == ACK_RWND) {
13564                         /* Case C */
13565                         if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13566                                 ts.tv_sec = ae->timestamp / 1000000000;
13567                                 ts.tv_nsec = ae->timestamp % 1000000000;
13568                                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13569                                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13570                         } else {
13571                                 rack->r_ctl.act_rcv_time = *tv;
13572                         }
13573                         if (rack->forced_ack) {
13574                                 rack_handle_probe_response(rack, tiwin,
13575                                                            tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13576                         }
13577 #ifdef TCP_ACCOUNTING
13578                         win_up_req = 1;
13579 #endif
13580                         win_upd_ack = ae->ack;
13581                         win_seq = ae->seq;
13582                         the_win = tiwin;
13583                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13584                 } else {
13585                         /* Case A */
13586                         if (SEQ_GT(ae->ack, tp->snd_max)) {
13587                                 /*
13588                                  * We just send an ack since the incoming
13589                                  * ack is beyond the largest seq we sent.
13590                                  */
13591                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
13592                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13593                                         if (tp->t_flags && TF_ACKNOW)
13594                                                 rack->r_wanted_output = 1;
13595                                 }
13596                         } else {
13597                                 nsegs++;
13598                                 /* If the window changed setup to update */
13599                                 if (tiwin != tp->snd_wnd) {
13600                                         win_upd_ack = ae->ack;
13601                                         win_seq = ae->seq;
13602                                         the_win = tiwin;
13603                                         rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13604                                 }
13605 #ifdef TCP_ACCOUNTING
13606                                 /* Account for the acks */
13607                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13608                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13609                                 }
13610                                 counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13611                                                 (((ae->ack - high_seq) + segsiz - 1) / segsiz));
13612 #endif
13613                                 high_seq = ae->ack;
13614                                 if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
13615                                         union tcp_log_stackspecific log;
13616                                         struct timeval tv;
13617
13618                                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13619                                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13620                                         log.u_bbr.flex1 = high_seq;
13621                                         log.u_bbr.flex2 = rack->r_ctl.roundends;
13622                                         log.u_bbr.flex3 = rack->r_ctl.current_round;
13623                                         log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
13624                                         log.u_bbr.flex8 = 8;
13625                                         tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
13626                                                        0, &log, false, NULL, NULL, 0, &tv);
13627                                 }
13628                                 /*
13629                                  * The draft (v3) calls for us to use SEQ_GEQ, but that
13630                                  * causes issues when we are just going app limited. Lets
13631                                  * instead use SEQ_GT <or> where its equal but more data
13632                                  * is outstanding.
13633                                  */
13634                                 if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) ||
13635                                     ((high_seq == rack->r_ctl.roundends) &&
13636                                      SEQ_GT(tp->snd_max, tp->snd_una))) {
13637                                         rack->r_ctl.current_round++;
13638                                         rack->r_ctl.roundends = tp->snd_max;
13639                                         if (CC_ALGO(tp)->newround != NULL) {
13640                                                 CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
13641                                         }
13642                                 }
13643                                 /* Setup our act_rcv_time */
13644                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13645                                         ts.tv_sec = ae->timestamp / 1000000000;
13646                                         ts.tv_nsec = ae->timestamp % 1000000000;
13647                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13648                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13649                                 } else {
13650                                         rack->r_ctl.act_rcv_time = *tv;
13651                                 }
13652                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13653                                 if (rack->rc_dsack_round_seen) {
13654                                         /* Is the dsack round over? */
13655                                         if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13656                                                 /* Yes it is */
13657                                                 rack->rc_dsack_round_seen = 0;
13658                                                 rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13659                                         }
13660                                 }
13661                         }
13662                 }
13663                 /* And lets be sure to commit the rtt measurements for this ack */
13664                 tcp_rack_xmit_timer_commit(rack, tp);
13665 #ifdef TCP_ACCOUNTING
13666                 rdstc = get_cyclecount();
13667                 if (rdstc > ts_val) {
13668                         counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13669                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13670                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13671                                 if (ae->ack_val_set == ACK_CUMACK)
13672                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13673                         }
13674                 }
13675 #endif
13676         }
13677 #ifdef TCP_ACCOUNTING
13678         ts_val = get_cyclecount();
13679 #endif
13680         /* Tend to any collapsed window */
13681         if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
13682                 /* The peer collapsed the window */
13683                 rack_collapsed_window(rack, (tp->snd_max - high_seq), __LINE__);
13684         } else if (rack->rc_has_collapsed)
13685                 rack_un_collapse_window(rack, __LINE__);
13686         if ((rack->r_collapse_point_valid) &&
13687             (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
13688                 rack->r_collapse_point_valid = 0;
13689         acked_amount = acked = (high_seq - tp->snd_una);
13690         if (acked) {
13691                 /*
13692                  * Clear the probe not answered flag
13693                  * since cum-ack moved forward.
13694                  */
13695                 rack->probe_not_answered = 0;
13696                 if (rack->sack_attack_disable == 0)
13697                         rack_do_decay(rack);
13698                 if (acked >= segsiz) {
13699                         /*
13700                          * You only get credit for
13701                          * MSS and greater (and you get extra
13702                          * credit for larger cum-ack moves).
13703                          */
13704                         int ac;
13705
13706                         ac = acked / segsiz;
13707                         rack->r_ctl.ack_count += ac;
13708                         counter_u64_add(rack_ack_total, ac);
13709                 }
13710                 if (rack->r_ctl.ack_count > 0xfff00000) {
13711                         /*
13712                          * reduce the number to keep us under
13713                          * a uint32_t.
13714                          */
13715                         rack->r_ctl.ack_count /= 2;
13716                         rack->r_ctl.sack_count /= 2;
13717                 }
13718                 if (tp->t_flags & TF_NEEDSYN) {
13719                         /*
13720                          * T/TCP: Connection was half-synchronized, and our SYN has
13721                          * been ACK'd (so connection is now fully synchronized).  Go
13722                          * to non-starred state, increment snd_una for ACK of SYN,
13723                          * and check if we can do window scaling.
13724                          */
13725                         tp->t_flags &= ~TF_NEEDSYN;
13726                         tp->snd_una++;
13727                         acked_amount = acked = (high_seq - tp->snd_una);
13728                 }
13729                 if (acked > sbavail(&so->so_snd))
13730                         acked_amount = sbavail(&so->so_snd);
13731 #ifdef NETFLIX_EXP_DETECTION
13732                 /*
13733                  * We only care on a cum-ack move if we are in a sack-disabled
13734                  * state. We have already added in to the ack_count, and we never
13735                  * would disable on a cum-ack move, so we only care to do the
13736                  * detection if it may "undo" it, i.e. we were in disabled already.
13737                  */
13738                 if (rack->sack_attack_disable)
13739                         rack_do_detection(tp, rack, acked_amount, segsiz);
13740 #endif
13741                 if (IN_FASTRECOVERY(tp->t_flags) &&
13742                     (rack->rack_no_prr == 0))
13743                         rack_update_prr(tp, rack, acked_amount, high_seq);
13744                 if (IN_RECOVERY(tp->t_flags)) {
13745                         if (SEQ_LT(high_seq, tp->snd_recover) &&
13746                             (SEQ_LT(high_seq, tp->snd_max))) {
13747                                 tcp_rack_partialack(tp);
13748                         } else {
13749                                 rack_post_recovery(tp, high_seq);
13750                                 recovery = 1;
13751                         }
13752                 }
13753                 /* Handle the rack-log-ack part (sendmap) */
13754                 if ((sbused(&so->so_snd) == 0) &&
13755                     (acked > acked_amount) &&
13756                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
13757                     (tp->t_flags & TF_SENTFIN)) {
13758                         /*
13759                          * We must be sure our fin
13760                          * was sent and acked (we can be
13761                          * in FIN_WAIT_1 without having
13762                          * sent the fin).
13763                          */
13764                         ourfinisacked = 1;
13765                         /*
13766                          * Lets make sure snd_una is updated
13767                          * since most likely acked_amount = 0 (it
13768                          * should be).
13769                          */
13770                         tp->snd_una = high_seq;
13771                 }
13772                 /* Did we make a RTO error? */
13773                 if ((tp->t_flags & TF_PREVVALID) &&
13774                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13775                         tp->t_flags &= ~TF_PREVVALID;
13776                         if (tp->t_rxtshift == 1 &&
13777                             (int)(ticks - tp->t_badrxtwin) < 0)
13778                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
13779                 }
13780                 /* Handle the data in the socket buffer */
13781                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13782                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13783                 if (acked_amount > 0) {
13784                         struct mbuf *mfree;
13785
13786                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13787                         SOCKBUF_LOCK(&so->so_snd);
13788                         mfree = sbcut_locked(&so->so_snd, acked_amount);
13789                         tp->snd_una = high_seq;
13790                         /* Note we want to hold the sb lock through the sendmap adjust */
13791                         rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13792                         /* Wake up the socket if we have room to write more */
13793                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13794                         sowwakeup_locked(so);
13795                         m_freem(mfree);
13796                 }
13797                 /* update progress */
13798                 tp->t_acktime = ticks;
13799                 rack_log_progress_event(rack, tp, tp->t_acktime,
13800                                         PROGRESS_UPDATE, __LINE__);
13801                 /* Clear out shifts and such */
13802                 tp->t_rxtshift = 0;
13803                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13804                                    rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13805                 rack->rc_tlp_in_progress = 0;
13806                 rack->r_ctl.rc_tlp_cnt_out = 0;
13807                 /* Send recover and snd_nxt must be dragged along */
13808                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13809                         tp->snd_recover = tp->snd_una;
13810                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13811                         tp->snd_nxt = tp->snd_una;
13812                 /*
13813                  * If the RXT timer is running we want to
13814                  * stop it, so we can restart a TLP (or new RXT).
13815                  */
13816                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13817                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13818 #ifdef NETFLIX_HTTP_LOGGING
13819                 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13820 #endif
13821                 tp->snd_wl2 = high_seq;
13822                 tp->t_dupacks = 0;
13823                 if (under_pacing &&
13824                     (rack->use_fixed_rate == 0) &&
13825                     (rack->in_probe_rtt == 0) &&
13826                     rack->rc_gp_dyn_mul &&
13827                     rack->rc_always_pace) {
13828                         /* Check if we are dragging bottom */
13829                         rack_check_bottom_drag(tp, rack, so, acked);
13830                 }
13831                 if (tp->snd_una == tp->snd_max) {
13832                         tp->t_flags &= ~TF_PREVVALID;
13833                         rack->r_ctl.retran_during_recovery = 0;
13834                         rack->r_ctl.dsack_byte_cnt = 0;
13835                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13836                         if (rack->r_ctl.rc_went_idle_time == 0)
13837                                 rack->r_ctl.rc_went_idle_time = 1;
13838                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13839                         if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13840                                 tp->t_acktime = 0;
13841                         /* Set so we might enter persists... */
13842                         rack->r_wanted_output = 1;
13843                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13844                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13845                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13846                             (sbavail(&so->so_snd) == 0) &&
13847                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13848                                 /*
13849                                  * The socket was gone and the
13850                                  * peer sent data (not now in the past), time to
13851                                  * reset him.
13852                                  */
13853                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13854                                 /* tcp_close will kill the inp pre-log the Reset */
13855                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13856 #ifdef TCP_ACCOUNTING
13857                                 rdstc = get_cyclecount();
13858                                 if (rdstc > ts_val) {
13859                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13860                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13861                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13862                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13863                                         }
13864                                 }
13865 #endif
13866                                 m_freem(m);
13867                                 tp = tcp_close(tp);
13868                                 if (tp == NULL) {
13869 #ifdef TCP_ACCOUNTING
13870                                         sched_unpin();
13871 #endif
13872                                         return (1);
13873                                 }
13874                                 /*
13875                                  * We would normally do drop-with-reset which would
13876                                  * send back a reset. We can't since we don't have
13877                                  * all the needed bits. Instead lets arrange for
13878                                  * a call to tcp_output(). That way since we
13879                                  * are in the closed state we will generate a reset.
13880                                  *
13881                                  * Note if tcp_accounting is on we don't unpin since
13882                                  * we do that after the goto label.
13883                                  */
13884                                 goto send_out_a_rst;
13885                         }
13886                         if ((sbused(&so->so_snd) == 0) &&
13887                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
13888                             (tp->t_flags & TF_SENTFIN)) {
13889                                 /*
13890                                  * If we can't receive any more data, then closing user can
13891                                  * proceed. Starting the timer is contrary to the
13892                                  * specification, but if we don't get a FIN we'll hang
13893                                  * forever.
13894                                  *
13895                                  */
13896                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13897                                         soisdisconnected(so);
13898                                         tcp_timer_activate(tp, TT_2MSL,
13899                                                            (tcp_fast_finwait2_recycle ?
13900                                                             tcp_finwait2_timeout :
13901                                                             TP_MAXIDLE(tp)));
13902                                 }
13903                                 if (ourfinisacked == 0) {
13904                                         /*
13905                                          * We don't change to fin-wait-2 if we have our fin acked
13906                                          * which means we are probably in TCPS_CLOSING.
13907                                          */
13908                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13909                                 }
13910                         }
13911                 }
13912                 /* Wake up the socket if we have room to write more */
13913                 if (sbavail(&so->so_snd)) {
13914                         rack->r_wanted_output = 1;
13915                         if (ctf_progress_timeout_check(tp, true)) {
13916                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13917                                                         tp, tick, PROGRESS_DROP, __LINE__);
13918                                 /*
13919                                  * We cheat here and don't send a RST, we should send one
13920                                  * when the pacer drops the connection.
13921                                  */
13922 #ifdef TCP_ACCOUNTING
13923                                 rdstc = get_cyclecount();
13924                                 if (rdstc > ts_val) {
13925                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13926                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13927                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13928                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13929                                         }
13930                                 }
13931                                 sched_unpin();
13932 #endif
13933                                 (void)tcp_drop(tp, ETIMEDOUT);
13934                                 m_freem(m);
13935                                 return (1);
13936                         }
13937                 }
13938                 if (ourfinisacked) {
13939                         switch(tp->t_state) {
13940                         case TCPS_CLOSING:
13941 #ifdef TCP_ACCOUNTING
13942                                 rdstc = get_cyclecount();
13943                                 if (rdstc > ts_val) {
13944                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13945                                                         (rdstc - ts_val));
13946                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13947                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13948                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13949                                         }
13950                                 }
13951                                 sched_unpin();
13952 #endif
13953                                 tcp_twstart(tp);
13954                                 m_freem(m);
13955                                 return (1);
13956                                 break;
13957                         case TCPS_LAST_ACK:
13958 #ifdef TCP_ACCOUNTING
13959                                 rdstc = get_cyclecount();
13960                                 if (rdstc > ts_val) {
13961                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13962                                                         (rdstc - ts_val));
13963                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13964                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13965                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13966                                         }
13967                                 }
13968                                 sched_unpin();
13969 #endif
13970                                 tp = tcp_close(tp);
13971                                 ctf_do_drop(m, tp);
13972                                 return (1);
13973                                 break;
13974                         case TCPS_FIN_WAIT_1:
13975 #ifdef TCP_ACCOUNTING
13976                                 rdstc = get_cyclecount();
13977                                 if (rdstc > ts_val) {
13978                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13979                                                         (rdstc - ts_val));
13980                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13981                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13982                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13983                                         }
13984                                 }
13985 #endif
13986                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13987                                         soisdisconnected(so);
13988                                         tcp_timer_activate(tp, TT_2MSL,
13989                                                            (tcp_fast_finwait2_recycle ?
13990                                                             tcp_finwait2_timeout :
13991                                                             TP_MAXIDLE(tp)));
13992                                 }
13993                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13994                                 break;
13995                         default:
13996                                 break;
13997                         }
13998                 }
13999                 if (rack->r_fast_output) {
14000                         /*
14001                          * We re doing fast output.. can we expand that?
14002                          */
14003                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
14004                 }
14005 #ifdef TCP_ACCOUNTING
14006                 rdstc = get_cyclecount();
14007                 if (rdstc > ts_val) {
14008                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
14009                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14010                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
14011                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
14012                         }
14013                 }
14014
14015         } else if (win_up_req) {
14016                 rdstc = get_cyclecount();
14017                 if (rdstc > ts_val) {
14018                         counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
14019                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14020                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
14021                         }
14022                 }
14023 #endif
14024         }
14025         /* Now is there a next packet, if so we are done */
14026         m_freem(m);
14027         did_out = 0;
14028         if (nxt_pkt) {
14029 #ifdef TCP_ACCOUNTING
14030                 sched_unpin();
14031 #endif
14032                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
14033                 return (0);
14034         }
14035         rack_handle_might_revert(tp, rack);
14036         ctf_calc_rwin(so, tp);
14037         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14038         send_out_a_rst:
14039                 if (tcp_output(tp) < 0) {
14040 #ifdef TCP_ACCOUNTING
14041                         sched_unpin();
14042 #endif
14043                         return (1);
14044                 }
14045                 did_out = 1;
14046         }
14047         rack_free_trim(rack);
14048 #ifdef TCP_ACCOUNTING
14049         sched_unpin();
14050 #endif
14051         rack_timer_audit(tp, rack, &so->so_snd);
14052         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
14053         return (0);
14054 }
14055
14056
14057 static int
14058 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
14059     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
14060     int32_t nxt_pkt, struct timeval *tv)
14061 {
14062 #ifdef TCP_ACCOUNTING
14063         uint64_t ts_val;
14064 #endif
14065         int32_t thflags, retval, did_out = 0;
14066         int32_t way_out = 0;
14067         /*
14068          * cts - is the current time from tv (caller gets ts) in microseconds.
14069          * ms_cts - is the current time from tv in milliseconds.
14070          * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
14071          */
14072         uint32_t cts, us_cts, ms_cts;
14073         uint32_t tiwin, high_seq;
14074         struct timespec ts;
14075         struct tcpopt to;
14076         struct tcp_rack *rack;
14077         struct rack_sendmap *rsm;
14078         int32_t prev_state = 0;
14079 #ifdef TCP_ACCOUNTING
14080         int ack_val_set = 0xf;
14081 #endif
14082         int nsegs;
14083         /*
14084          * tv passed from common code is from either M_TSTMP_LRO or
14085          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
14086          */
14087         rack = (struct tcp_rack *)tp->t_fb_ptr;
14088         if (m->m_flags & M_ACKCMP) {
14089                 /*
14090                  * All compressed ack's are ack's by definition so
14091                  * remove any ack required flag and then do the processing.
14092                  */
14093                 rack->rc_ack_required = 0;
14094                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
14095         }
14096         if (m->m_flags & M_ACKCMP) {
14097                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
14098         }
14099         cts = tcp_tv_to_usectick(tv);
14100         ms_cts =  tcp_tv_to_mssectick(tv);
14101         nsegs = m->m_pkthdr.lro_nsegs;
14102         counter_u64_add(rack_proc_non_comp_ack, 1);
14103         thflags = tcp_get_flags(th);
14104 #ifdef TCP_ACCOUNTING
14105         sched_pin();
14106         if (thflags & TH_ACK)
14107                 ts_val = get_cyclecount();
14108 #endif
14109         if ((m->m_flags & M_TSTMP) ||
14110             (m->m_flags & M_TSTMP_LRO)) {
14111                 mbuf_tstmp2timespec(m, &ts);
14112                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
14113                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
14114         } else
14115                 rack->r_ctl.act_rcv_time = *tv;
14116         kern_prefetch(rack, &prev_state);
14117         prev_state = 0;
14118         /*
14119          * Unscale the window into a 32-bit value. For the SYN_SENT state
14120          * the scale is zero.
14121          */
14122         tiwin = th->th_win << tp->snd_scale;
14123 #ifdef TCP_ACCOUNTING
14124         if (thflags & TH_ACK) {
14125                 /*
14126                  * We have a tradeoff here. We can either do what we are
14127                  * doing i.e. pinning to this CPU and then doing the accounting
14128                  * <or> we could do a critical enter, setup the rdtsc and cpu
14129                  * as in below, and then validate we are on the same CPU on
14130                  * exit. I have choosen to not do the critical enter since
14131                  * that often will gain you a context switch, and instead lock
14132                  * us (line above this if) to the same CPU with sched_pin(). This
14133                  * means we may be context switched out for a higher priority
14134                  * interupt but we won't be moved to another CPU.
14135                  *
14136                  * If this occurs (which it won't very often since we most likely
14137                  * are running this code in interupt context and only a higher
14138                  * priority will bump us ... clock?) we will falsely add in
14139                  * to the time the interupt processing time plus the ack processing
14140                  * time. This is ok since its a rare event.
14141                  */
14142                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
14143                                                     ctf_fixed_maxseg(tp));
14144         }
14145 #endif
14146         /*
14147          * Parse options on any incoming segment.
14148          */
14149         memset(&to, 0, sizeof(to));
14150         tcp_dooptions(&to, (u_char *)(th + 1),
14151             (th->th_off << 2) - sizeof(struct tcphdr),
14152             (thflags & TH_SYN) ? TO_SYN : 0);
14153         NET_EPOCH_ASSERT();
14154         INP_WLOCK_ASSERT(tp->t_inpcb);
14155         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
14156             __func__));
14157         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
14158             __func__));
14159         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
14160             (tp->t_flags & TF_GPUTINPROG)) {
14161                 /*
14162                  * We have a goodput in progress
14163                  * and we have entered a late state.
14164                  * Do we have enough data in the sb
14165                  * to handle the GPUT request?
14166                  */
14167                 uint32_t bytes;
14168
14169                 bytes = tp->gput_ack - tp->gput_seq;
14170                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
14171                         bytes += tp->gput_seq - tp->snd_una;
14172                 if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
14173                         /*
14174                          * There are not enough bytes in the socket
14175                          * buffer that have been sent to cover this
14176                          * measurement. Cancel it.
14177                          */
14178                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14179                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
14180                                                    tp->gput_seq,
14181                                                    0, 0, 18, __LINE__, NULL, 0);
14182                         tp->t_flags &= ~TF_GPUTINPROG;
14183                 }
14184         }
14185         high_seq = th->th_ack;
14186         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14187                 union tcp_log_stackspecific log;
14188                 struct timeval ltv;
14189 #ifdef NETFLIX_HTTP_LOGGING
14190                 struct http_sendfile_track *http_req;
14191
14192                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
14193                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14194                 } else {
14195                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14196                 }
14197 #endif
14198                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14199                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14200                 if (rack->rack_no_prr == 0)
14201                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14202                 else
14203                         log.u_bbr.flex1 = 0;
14204                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14205                 log.u_bbr.use_lt_bw <<= 1;
14206                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
14207                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14208                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14209                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14210                 log.u_bbr.flex3 = m->m_flags;
14211                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14212                 log.u_bbr.lost = thflags;
14213                 log.u_bbr.pacing_gain = 0x1;
14214 #ifdef TCP_ACCOUNTING
14215                 log.u_bbr.cwnd_gain = ack_val_set;
14216 #endif
14217                 log.u_bbr.flex7 = 2;
14218                 if (m->m_flags & M_TSTMP) {
14219                         /* Record the hardware timestamp if present */
14220                         mbuf_tstmp2timespec(m, &ts);
14221                         ltv.tv_sec = ts.tv_sec;
14222                         ltv.tv_usec = ts.tv_nsec / 1000;
14223                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14224                 } else if (m->m_flags & M_TSTMP_LRO) {
14225                         /* Record the LRO the arrival timestamp */
14226                         mbuf_tstmp2timespec(m, &ts);
14227                         ltv.tv_sec = ts.tv_sec;
14228                         ltv.tv_usec = ts.tv_nsec / 1000;
14229                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14230                 }
14231                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14232                 /* Log the rcv time */
14233                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14234 #ifdef NETFLIX_HTTP_LOGGING
14235                 log.u_bbr.applimited = tp->t_http_closed;
14236                 log.u_bbr.applimited <<= 8;
14237                 log.u_bbr.applimited |= tp->t_http_open;
14238                 log.u_bbr.applimited <<= 8;
14239                 log.u_bbr.applimited |= tp->t_http_req;
14240                 if (http_req) {
14241                         /* Copy out any client req info */
14242                         /* seconds */
14243                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14244                         /* useconds */
14245                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14246                         log.u_bbr.rttProp = http_req->timestamp;
14247                         log.u_bbr.cur_del_rate = http_req->start;
14248                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14249                                 log.u_bbr.flex8 |= 1;
14250                         } else {
14251                                 log.u_bbr.flex8 |= 2;
14252                                 log.u_bbr.bw_inuse = http_req->end;
14253                         }
14254                         log.u_bbr.flex6 = http_req->start_seq;
14255                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14256                                 log.u_bbr.flex8 |= 4;
14257                                 log.u_bbr.epoch = http_req->end_seq;
14258                         }
14259                 }
14260 #endif
14261                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14262                     tlen, &log, true, &ltv);
14263         }
14264         /* Remove ack required flag if set, we have one  */
14265         if (thflags & TH_ACK)
14266                 rack->rc_ack_required = 0;
14267         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14268                 way_out = 4;
14269                 retval = 0;
14270                 m_freem(m);
14271                 goto done_with_input;
14272         }
14273         /*
14274          * If a segment with the ACK-bit set arrives in the SYN-SENT state
14275          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14276          */
14277         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14278             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14279                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14280                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14281 #ifdef TCP_ACCOUNTING
14282                 sched_unpin();
14283 #endif
14284                 return (1);
14285         }
14286         /*
14287          * If timestamps were negotiated during SYN/ACK and a
14288          * segment without a timestamp is received, silently drop
14289          * the segment, unless it is a RST segment or missing timestamps are
14290          * tolerated.
14291          * See section 3.2 of RFC 7323.
14292          */
14293         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14294             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14295                 way_out = 5;
14296                 retval = 0;
14297                 m_freem(m);
14298                 goto done_with_input;
14299         }
14300
14301         /*
14302          * Segment received on connection. Reset idle time and keep-alive
14303          * timer. XXX: This should be done after segment validation to
14304          * ignore broken/spoofed segs.
14305          */
14306         if  (tp->t_idle_reduce &&
14307              (tp->snd_max == tp->snd_una) &&
14308              (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14309                 counter_u64_add(rack_input_idle_reduces, 1);
14310                 rack_cc_after_idle(rack, tp);
14311         }
14312         tp->t_rcvtime = ticks;
14313 #ifdef STATS
14314         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14315 #endif
14316         if (tiwin > rack->r_ctl.rc_high_rwnd)
14317                 rack->r_ctl.rc_high_rwnd = tiwin;
14318         /*
14319          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14320          * this to occur after we've validated the segment.
14321          */
14322         if (tcp_ecn_input_segment(tp, thflags, iptos))
14323                 rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
14324
14325         /*
14326          * If echoed timestamp is later than the current time, fall back to
14327          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14328          * were used when this connection was established.
14329          */
14330         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14331                 to.to_tsecr -= tp->ts_offset;
14332                 if (TSTMP_GT(to.to_tsecr, ms_cts))
14333                         to.to_tsecr = 0;
14334         }
14335
14336         /*
14337          * If its the first time in we need to take care of options and
14338          * verify we can do SACK for rack!
14339          */
14340         if (rack->r_state == 0) {
14341                 /* Should be init'd by rack_init() */
14342                 KASSERT(rack->rc_inp != NULL,
14343                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
14344                 if (rack->rc_inp == NULL) {
14345                         rack->rc_inp = tp->t_inpcb;
14346                 }
14347
14348                 /*
14349                  * Process options only when we get SYN/ACK back. The SYN
14350                  * case for incoming connections is handled in tcp_syncache.
14351                  * According to RFC1323 the window field in a SYN (i.e., a
14352                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14353                  * this is traditional behavior, may need to be cleaned up.
14354                  */
14355                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14356                         /* Handle parallel SYN for ECN */
14357                         tcp_ecn_input_parallel_syn(tp, thflags, iptos);
14358                         if ((to.to_flags & TOF_SCALE) &&
14359                             (tp->t_flags & TF_REQ_SCALE)) {
14360                                 tp->t_flags |= TF_RCVD_SCALE;
14361                                 tp->snd_scale = to.to_wscale;
14362                         } else
14363                                 tp->t_flags &= ~TF_REQ_SCALE;
14364                         /*
14365                          * Initial send window.  It will be updated with the
14366                          * next incoming segment to the scaled value.
14367                          */
14368                         tp->snd_wnd = th->th_win;
14369                         rack_validate_fo_sendwin_up(tp, rack);
14370                         if ((to.to_flags & TOF_TS) &&
14371                             (tp->t_flags & TF_REQ_TSTMP)) {
14372                                 tp->t_flags |= TF_RCVD_TSTMP;
14373                                 tp->ts_recent = to.to_tsval;
14374                                 tp->ts_recent_age = cts;
14375                         } else
14376                                 tp->t_flags &= ~TF_REQ_TSTMP;
14377                         if (to.to_flags & TOF_MSS) {
14378                                 tcp_mss(tp, to.to_mss);
14379                         }
14380                         if ((tp->t_flags & TF_SACK_PERMIT) &&
14381                             (to.to_flags & TOF_SACKPERM) == 0)
14382                                 tp->t_flags &= ~TF_SACK_PERMIT;
14383                         if (IS_FASTOPEN(tp->t_flags)) {
14384                                 if (to.to_flags & TOF_FASTOPEN) {
14385                                         uint16_t mss;
14386
14387                                         if (to.to_flags & TOF_MSS)
14388                                                 mss = to.to_mss;
14389                                         else
14390                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
14391                                                         mss = TCP6_MSS;
14392                                                 else
14393                                                         mss = TCP_MSS;
14394                                         tcp_fastopen_update_cache(tp, mss,
14395                                             to.to_tfo_len, to.to_tfo_cookie);
14396                                 } else
14397                                         tcp_fastopen_disable_path(tp);
14398                         }
14399                 }
14400                 /*
14401                  * At this point we are at the initial call. Here we decide
14402                  * if we are doing RACK or not. We do this by seeing if
14403                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
14404                  * The code now does do dup-ack counting so if you don't
14405                  * switch back you won't get rack & TLP, but you will still
14406                  * get this stack.
14407                  */
14408
14409                 if ((rack_sack_not_required == 0) &&
14410                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14411                         tcp_switch_back_to_default(tp);
14412                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14413                             tlen, iptos);
14414 #ifdef TCP_ACCOUNTING
14415                         sched_unpin();
14416 #endif
14417                         return (1);
14418                 }
14419                 tcp_set_hpts(tp->t_inpcb);
14420                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14421         }
14422         if (thflags & TH_FIN)
14423                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14424         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14425         if ((rack->rc_gp_dyn_mul) &&
14426             (rack->use_fixed_rate == 0) &&
14427             (rack->rc_always_pace)) {
14428                 /* Check in on probertt */
14429                 rack_check_probe_rtt(rack, us_cts);
14430         }
14431         rack_clear_rate_sample(rack);
14432         if ((rack->forced_ack) &&
14433             ((tcp_get_flags(th) & TH_RST) == 0)) {
14434                 rack_handle_probe_response(rack, tiwin, us_cts);
14435         }
14436         /*
14437          * This is the one exception case where we set the rack state
14438          * always. All other times (timers etc) we must have a rack-state
14439          * set (so we assure we have done the checks above for SACK).
14440          */
14441         rack->r_ctl.rc_rcvtime = cts;
14442         if (rack->r_state != tp->t_state)
14443                 rack_set_state(tp, rack);
14444         if (SEQ_GT(th->th_ack, tp->snd_una) &&
14445             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14446                 kern_prefetch(rsm, &prev_state);
14447         prev_state = rack->r_state;
14448         retval = (*rack->r_substate) (m, th, so,
14449             tp, &to, drop_hdrlen,
14450             tlen, tiwin, thflags, nxt_pkt, iptos);
14451 #ifdef INVARIANTS
14452         if ((retval == 0) &&
14453             (tp->t_inpcb == NULL)) {
14454                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
14455                     retval, tp, prev_state);
14456         }
14457 #endif
14458         if (retval == 0) {
14459                 /*
14460                  * If retval is 1 the tcb is unlocked and most likely the tp
14461                  * is gone.
14462                  */
14463                 INP_WLOCK_ASSERT(tp->t_inpcb);
14464                 if ((rack->rc_gp_dyn_mul) &&
14465                     (rack->rc_always_pace) &&
14466                     (rack->use_fixed_rate == 0) &&
14467                     rack->in_probe_rtt &&
14468                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
14469                         /*
14470                          * If we are going for target, lets recheck before
14471                          * we output.
14472                          */
14473                         rack_check_probe_rtt(rack, us_cts);
14474                 }
14475                 if (rack->set_pacing_done_a_iw == 0) {
14476                         /* How much has been acked? */
14477                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14478                                 /* We have enough to set in the pacing segment size */
14479                                 rack->set_pacing_done_a_iw = 1;
14480                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
14481                         }
14482                 }
14483                 tcp_rack_xmit_timer_commit(rack, tp);
14484 #ifdef TCP_ACCOUNTING
14485                 /*
14486                  * If we set the ack_val_se to what ack processing we are doing
14487                  * we also want to track how many cycles we burned. Note
14488                  * the bits after tcp_output we let be "free". This is because
14489                  * we are also tracking the tcp_output times as well. Note the
14490                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
14491                  * 0xf cannot be returned and is what we initialize it too to
14492                  * indicate we are not doing the tabulations.
14493                  */
14494                 if (ack_val_set != 0xf) {
14495                         uint64_t crtsc;
14496
14497                         crtsc = get_cyclecount();
14498                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14499                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14500                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14501                         }
14502                 }
14503 #endif
14504                 if (nxt_pkt == 0) {
14505                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14506 do_output_now:
14507                                 if (tcp_output(tp) < 0)
14508                                         return (1);
14509                                 did_out = 1;
14510                         }
14511                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14512                         rack_free_trim(rack);
14513                 }
14514                 /* Update any rounds needed */
14515                 if (rack_verbose_logging &&  (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
14516                         union tcp_log_stackspecific log;
14517                         struct timeval tv;
14518
14519                         memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14520                         log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14521                         log.u_bbr.flex1 = high_seq;
14522                         log.u_bbr.flex2 = rack->r_ctl.roundends;
14523                         log.u_bbr.flex3 = rack->r_ctl.current_round;
14524                         log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
14525                         log.u_bbr.flex8 = 9;
14526                         tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
14527                                        0, &log, false, NULL, NULL, 0, &tv);
14528                 }
14529                 /*
14530                  * The draft (v3) calls for us to use SEQ_GEQ, but that
14531                  * causes issues when we are just going app limited. Lets
14532                  * instead use SEQ_GT <or> where its equal but more data
14533                  * is outstanding.
14534                  */
14535                 if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) ||
14536                     ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) {
14537                         rack->r_ctl.current_round++;
14538                         rack->r_ctl.roundends = tp->snd_max;
14539                         if (CC_ALGO(tp)->newround != NULL) {
14540                                 CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
14541                         }
14542                 }
14543                 if ((nxt_pkt == 0) &&
14544                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14545                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
14546                      (tp->t_flags & TF_DELACK) ||
14547                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14548                       (tp->t_state <= TCPS_CLOSING)))) {
14549                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
14550                         if ((tp->snd_max == tp->snd_una) &&
14551                             ((tp->t_flags & TF_DELACK) == 0) &&
14552                             (tcp_in_hpts(rack->rc_inp)) &&
14553                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14554                                 /* keep alive not needed if we are hptsi output yet */
14555                                 ;
14556                         } else {
14557                                 int late = 0;
14558                                 if (tcp_in_hpts(rack->rc_inp)) {
14559                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14560                                                 us_cts = tcp_get_usecs(NULL);
14561                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14562                                                         rack->r_early = 1;
14563                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14564                                                 } else
14565                                                         late = 1;
14566                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14567                                         }
14568                                         tcp_hpts_remove(tp->t_inpcb);
14569                                 }
14570                                 if (late && (did_out == 0)) {
14571                                         /*
14572                                          * We are late in the sending
14573                                          * and we did not call the output
14574                                          * (this probably should not happen).
14575                                          */
14576                                         goto do_output_now;
14577                                 }
14578                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14579                         }
14580                         way_out = 1;
14581                 } else if (nxt_pkt == 0) {
14582                         /* Do we have the correct timer running? */
14583                         rack_timer_audit(tp, rack, &so->so_snd);
14584                         way_out = 2;
14585                 }
14586         done_with_input:
14587                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14588                 if (did_out)
14589                         rack->r_wanted_output = 0;
14590 #ifdef INVARIANTS
14591                 if (tp->t_inpcb == NULL) {
14592                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
14593                               did_out,
14594                               retval, tp, prev_state);
14595                 }
14596 #endif
14597 #ifdef TCP_ACCOUNTING
14598         } else {
14599                 /*
14600                  * Track the time (see above).
14601                  */
14602                 if (ack_val_set != 0xf) {
14603                         uint64_t crtsc;
14604
14605                         crtsc = get_cyclecount();
14606                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14607                         /*
14608                          * Note we *DO NOT* increment the per-tcb counters since
14609                          * in the else the TP may be gone!!
14610                          */
14611                 }
14612 #endif
14613         }
14614 #ifdef TCP_ACCOUNTING
14615         sched_unpin();
14616 #endif
14617         return (retval);
14618 }
14619
14620 void
14621 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14622     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14623 {
14624         struct timeval tv;
14625
14626         /* First lets see if we have old packets */
14627         if (tp->t_in_pkt) {
14628                 if (ctf_do_queued_segments(so, tp, 1)) {
14629                         m_freem(m);
14630                         return;
14631                 }
14632         }
14633         if (m->m_flags & M_TSTMP_LRO) {
14634                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
14635                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
14636         } else {
14637                 /* Should not be should we kassert instead? */
14638                 tcp_get_usecs(&tv);
14639         }
14640         if (rack_do_segment_nounlock(m, th, so, tp,
14641                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14642                 INP_WUNLOCK(tp->t_inpcb);
14643         }
14644 }
14645
14646 struct rack_sendmap *
14647 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14648 {
14649         struct rack_sendmap *rsm = NULL;
14650         int32_t idx;
14651         uint32_t srtt = 0, thresh = 0, ts_low = 0;
14652
14653         /* Return the next guy to be re-transmitted */
14654         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14655                 return (NULL);
14656         }
14657         if (tp->t_flags & TF_SENTFIN) {
14658                 /* retran the end FIN? */
14659                 return (NULL);
14660         }
14661         /* ok lets look at this one */
14662         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14663         if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
14664                 return (rsm);
14665         }
14666         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14667                 goto check_it;
14668         }
14669         rsm = rack_find_lowest_rsm(rack);
14670         if (rsm == NULL) {
14671                 return (NULL);
14672         }
14673 check_it:
14674         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14675             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14676                 /*
14677                  * No sack so we automatically do the 3 strikes and
14678                  * retransmit (no rack timer would be started).
14679                  */
14680
14681                 return (rsm);
14682         }
14683         if (rsm->r_flags & RACK_ACKED) {
14684                 return (NULL);
14685         }
14686         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14687             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14688                 /* Its not yet ready */
14689                 return (NULL);
14690         }
14691         srtt = rack_grab_rtt(tp, rack);
14692         idx = rsm->r_rtr_cnt - 1;
14693         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14694         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14695         if ((tsused == ts_low) ||
14696             (TSTMP_LT(tsused, ts_low))) {
14697                 /* No time since sending */
14698                 return (NULL);
14699         }
14700         if ((tsused - ts_low) < thresh) {
14701                 /* It has not been long enough yet */
14702                 return (NULL);
14703         }
14704         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14705             ((rsm->r_flags & RACK_SACK_PASSED) &&
14706              (rack->sack_attack_disable == 0))) {
14707                 /*
14708                  * We have passed the dup-ack threshold <or>
14709                  * a SACK has indicated this is missing.
14710                  * Note that if you are a declared attacker
14711                  * it is only the dup-ack threshold that
14712                  * will cause retransmits.
14713                  */
14714                 /* log retransmit reason */
14715                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14716                 rack->r_fast_output = 0;
14717                 return (rsm);
14718         }
14719         return (NULL);
14720 }
14721
14722 static void
14723 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14724                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14725                            int line, struct rack_sendmap *rsm, uint8_t quality)
14726 {
14727         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14728                 union tcp_log_stackspecific log;
14729                 struct timeval tv;
14730
14731                 memset(&log, 0, sizeof(log));
14732                 log.u_bbr.flex1 = slot;
14733                 log.u_bbr.flex2 = len;
14734                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14735                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14736                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14737                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14738                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14739                 log.u_bbr.use_lt_bw <<= 1;
14740                 log.u_bbr.use_lt_bw |= rack->r_late;
14741                 log.u_bbr.use_lt_bw <<= 1;
14742                 log.u_bbr.use_lt_bw |= rack->r_early;
14743                 log.u_bbr.use_lt_bw <<= 1;
14744                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14745                 log.u_bbr.use_lt_bw <<= 1;
14746                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14747                 log.u_bbr.use_lt_bw <<= 1;
14748                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14749                 log.u_bbr.use_lt_bw <<= 1;
14750                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14751                 log.u_bbr.use_lt_bw <<= 1;
14752                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14753                 log.u_bbr.pkt_epoch = line;
14754                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14755                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14756                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14757                 log.u_bbr.bw_inuse = bw_est;
14758                 log.u_bbr.delRate = bw;
14759                 if (rack->r_ctl.gp_bw == 0)
14760                         log.u_bbr.cur_del_rate = 0;
14761                 else
14762                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14763                 log.u_bbr.rttProp = len_time;
14764                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14765                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14766                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14767                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14768                         /* We are in slow start */
14769                         log.u_bbr.flex7 = 1;
14770                 } else {
14771                         /* we are on congestion avoidance */
14772                         log.u_bbr.flex7 = 0;
14773                 }
14774                 log.u_bbr.flex8 = method;
14775                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14776                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14777                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14778                 log.u_bbr.cwnd_gain <<= 1;
14779                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14780                 log.u_bbr.cwnd_gain <<= 1;
14781                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14782                 log.u_bbr.bbr_substate = quality;
14783                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14784                     &rack->rc_inp->inp_socket->so_rcv,
14785                     &rack->rc_inp->inp_socket->so_snd,
14786                     BBR_LOG_HPTSI_CALC, 0,
14787                     0, &log, false, &tv);
14788         }
14789 }
14790
14791 static uint32_t
14792 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14793 {
14794         uint32_t new_tso, user_max;
14795
14796         user_max = rack->rc_user_set_max_segs * mss;
14797         if (rack->rc_force_max_seg) {
14798                 return (user_max);
14799         }
14800         if (rack->use_fixed_rate &&
14801             ((rack->r_ctl.crte == NULL) ||
14802              (bw != rack->r_ctl.crte->rate))) {
14803                 /* Use the user mss since we are not exactly matched */
14804                 return (user_max);
14805         }
14806         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14807         if (new_tso > user_max)
14808                 new_tso = user_max;
14809         return (new_tso);
14810 }
14811
14812 static int32_t
14813 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)
14814 {
14815         uint64_t lentim, fill_bw;
14816
14817         /* Lets first see if we are full, if so continue with normal rate */
14818         rack->r_via_fill_cw = 0;
14819         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14820                 return (slot);
14821         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14822                 return (slot);
14823         if (rack->r_ctl.rc_last_us_rtt == 0)
14824                 return (slot);
14825         if (rack->rc_pace_fill_if_rttin_range &&
14826             (rack->r_ctl.rc_last_us_rtt >=
14827              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14828                 /* The rtt is huge, N * smallest, lets not fill */
14829                 return (slot);
14830         }
14831         /*
14832          * first lets calculate the b/w based on the last us-rtt
14833          * and the sndwnd.
14834          */
14835         fill_bw = rack->r_ctl.cwnd_to_use;
14836         /* Take the rwnd if its smaller */
14837         if (fill_bw > rack->rc_tp->snd_wnd)
14838                 fill_bw = rack->rc_tp->snd_wnd;
14839         if (rack->r_fill_less_agg) {
14840                 /*
14841                  * Now take away the inflight (this will reduce our
14842                  * aggressiveness and yeah, if we get that much out in 1RTT
14843                  * we will have had acks come back and still be behind).
14844                  */
14845                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14846         }
14847         /* Now lets make it into a b/w */
14848         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14849         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14850         /* We are below the min b/w */
14851         if (non_paced)
14852                 *rate_wanted = fill_bw;
14853         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14854                 return (slot);
14855         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14856                 fill_bw = rack->r_ctl.bw_rate_cap;
14857         rack->r_via_fill_cw = 1;
14858         if (rack->r_rack_hw_rate_caps &&
14859             (rack->r_ctl.crte != NULL)) {
14860                 uint64_t high_rate;
14861
14862                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14863                 if (fill_bw > high_rate) {
14864                         /* We are capping bw at the highest rate table entry */
14865                         if (*rate_wanted > high_rate) {
14866                                 /* The original rate was also capped */
14867                                 rack->r_via_fill_cw = 0;
14868                         }
14869                         rack_log_hdwr_pacing(rack,
14870                                              fill_bw, high_rate, __LINE__,
14871                                              0, 3);
14872                         fill_bw = high_rate;
14873                         if (capped)
14874                                 *capped = 1;
14875                 }
14876         } else if ((rack->r_ctl.crte == NULL) &&
14877                    (rack->rack_hdrw_pacing == 0) &&
14878                    (rack->rack_hdw_pace_ena) &&
14879                    rack->r_rack_hw_rate_caps &&
14880                    (rack->rack_attempt_hdwr_pace == 0) &&
14881                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14882                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14883                 /*
14884                  * Ok we may have a first attempt that is greater than our top rate
14885                  * lets check.
14886                  */
14887                 uint64_t high_rate;
14888
14889                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14890                 if (high_rate) {
14891                         if (fill_bw > high_rate) {
14892                                 fill_bw = high_rate;
14893                                 if (capped)
14894                                         *capped = 1;
14895                         }
14896                 }
14897         }
14898         /*
14899          * Ok fill_bw holds our mythical b/w to fill the cwnd
14900          * in a rtt, what does that time wise equate too?
14901          */
14902         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14903         lentim /= fill_bw;
14904         *rate_wanted = fill_bw;
14905         if (non_paced || (lentim < slot)) {
14906                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14907                                            0, lentim, 12, __LINE__, NULL, 0);
14908                 return ((int32_t)lentim);
14909         } else
14910                 return (slot);
14911 }
14912
14913 static int32_t
14914 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14915 {
14916         uint64_t srtt;
14917         int32_t slot = 0;
14918         int can_start_hw_pacing = 1;
14919         int err;
14920
14921         if (rack->rc_always_pace == 0) {
14922                 /*
14923                  * We use the most optimistic possible cwnd/srtt for
14924                  * sending calculations. This will make our
14925                  * calculation anticipate getting more through
14926                  * quicker then possible. But thats ok we don't want
14927                  * the peer to have a gap in data sending.
14928                  */
14929                 uint64_t cwnd, tr_perms = 0;
14930                 int32_t reduce = 0;
14931
14932         old_method:
14933                 /*
14934                  * We keep no precise pacing with the old method
14935                  * instead we use the pacer to mitigate bursts.
14936                  */
14937                 if (rack->r_ctl.rc_rack_min_rtt)
14938                         srtt = rack->r_ctl.rc_rack_min_rtt;
14939                 else
14940                         srtt = max(tp->t_srtt, 1);
14941                 if (rack->r_ctl.rc_rack_largest_cwnd)
14942                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14943                 else
14944                         cwnd = rack->r_ctl.cwnd_to_use;
14945                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14946                 tr_perms = (cwnd * 1000) / srtt;
14947                 if (tr_perms == 0) {
14948                         tr_perms = ctf_fixed_maxseg(tp);
14949                 }
14950                 /*
14951                  * Calculate how long this will take to drain, if
14952                  * the calculation comes out to zero, thats ok we
14953                  * will use send_a_lot to possibly spin around for
14954                  * more increasing tot_len_this_send to the point
14955                  * that its going to require a pace, or we hit the
14956                  * cwnd. Which in that case we are just waiting for
14957                  * a ACK.
14958                  */
14959                 slot = len / tr_perms;
14960                 /* Now do we reduce the time so we don't run dry? */
14961                 if (slot && rack_slot_reduction) {
14962                         reduce = (slot / rack_slot_reduction);
14963                         if (reduce < slot) {
14964                                 slot -= reduce;
14965                         } else
14966                                 slot = 0;
14967                 }
14968                 slot *= HPTS_USEC_IN_MSEC;
14969                 if (rack->rc_pace_to_cwnd) {
14970                         uint64_t rate_wanted = 0;
14971
14972                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14973                         rack->rc_ack_can_sendout_data = 1;
14974                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14975                 } else
14976                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14977         } else {
14978                 uint64_t bw_est, res, lentim, rate_wanted;
14979                 uint32_t orig_val, segs, oh;
14980                 int capped = 0;
14981                 int prev_fill;
14982
14983                 if ((rack->r_rr_config == 1) && rsm) {
14984                         return (rack->r_ctl.rc_min_to);
14985                 }
14986                 if (rack->use_fixed_rate) {
14987                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14988                 } else if ((rack->r_ctl.init_rate == 0) &&
14989 #ifdef NETFLIX_PEAKRATE
14990                            (rack->rc_tp->t_maxpeakrate == 0) &&
14991 #endif
14992                            (rack->r_ctl.gp_bw == 0)) {
14993                         /* no way to yet do an estimate */
14994                         bw_est = rate_wanted = 0;
14995                 } else {
14996                         bw_est = rack_get_bw(rack);
14997                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14998                 }
14999                 if ((bw_est == 0) || (rate_wanted == 0) ||
15000                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
15001                         /*
15002                          * No way yet to make a b/w estimate or
15003                          * our raise is set incorrectly.
15004                          */
15005                         goto old_method;
15006                 }
15007                 /* We need to account for all the overheads */
15008                 segs = (len + segsiz - 1) / segsiz;
15009                 /*
15010                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
15011                  * and how much data we put in each packet. Yes this
15012                  * means we may be off if we are larger than 1500 bytes
15013                  * or smaller. But this just makes us more conservative.
15014                  */
15015                 if (rack_hw_rate_min &&
15016                     (bw_est < rack_hw_rate_min))
15017                         can_start_hw_pacing = 0;
15018                 if (ETHERNET_SEGMENT_SIZE > segsiz)
15019                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
15020                 else
15021                         oh = 0;
15022                 segs *= oh;
15023                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
15024                 res = lentim / rate_wanted;
15025                 slot = (uint32_t)res;
15026                 orig_val = rack->r_ctl.rc_pace_max_segs;
15027                 if (rack->r_ctl.crte == NULL) {
15028                         /*
15029                          * Only do this if we are not hardware pacing
15030                          * since if we are doing hw-pacing below we will
15031                          * set make a call after setting up or changing
15032                          * the rate.
15033                          */
15034                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
15035                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
15036                         /*
15037                          * We lost our rate somehow, this can happen
15038                          * if the interface changed underneath us.
15039                          */
15040                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15041                         rack->r_ctl.crte = NULL;
15042                         /* Lets re-allow attempting to setup pacing */
15043                         rack->rack_hdrw_pacing = 0;
15044                         rack->rack_attempt_hdwr_pace = 0;
15045                         rack_log_hdwr_pacing(rack,
15046                                              rate_wanted, bw_est, __LINE__,
15047                                              0, 6);
15048                 }
15049                 /* Did we change the TSO size, if so log it */
15050                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
15051                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
15052                 prev_fill = rack->r_via_fill_cw;
15053                 if ((rack->rc_pace_to_cwnd) &&
15054                     (capped == 0) &&
15055                     (rack->use_fixed_rate == 0) &&
15056                     (rack->in_probe_rtt == 0) &&
15057                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
15058                         /*
15059                          * We want to pace at our rate *or* faster to
15060                          * fill the cwnd to the max if its not full.
15061                          */
15062                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
15063                 }
15064                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
15065                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
15066                         if ((rack->rack_hdw_pace_ena) &&
15067                             (can_start_hw_pacing > 0) &&
15068                             (rack->rack_hdrw_pacing == 0) &&
15069                             (rack->rack_attempt_hdwr_pace == 0)) {
15070                                 /*
15071                                  * Lets attempt to turn on hardware pacing
15072                                  * if we can.
15073                                  */
15074                                 rack->rack_attempt_hdwr_pace = 1;
15075                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
15076                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
15077                                                                        rate_wanted,
15078                                                                        RS_PACING_GEQ,
15079                                                                        &err, &rack->r_ctl.crte_prev_rate);
15080                                 if (rack->r_ctl.crte) {
15081                                         rack->rack_hdrw_pacing = 1;
15082                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
15083                                                                                                  0, rack->r_ctl.crte,
15084                                                                                                  NULL);
15085                                         rack_log_hdwr_pacing(rack,
15086                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15087                                                              err, 0);
15088                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
15089                                 } else {
15090                                         counter_u64_add(rack_hw_pace_init_fail, 1);
15091                                 }
15092                         } else if (rack->rack_hdrw_pacing &&
15093                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
15094                                 /* Do we need to adjust our rate? */
15095                                 const struct tcp_hwrate_limit_table *nrte;
15096
15097                                 if (rack->r_up_only &&
15098                                     (rate_wanted < rack->r_ctl.crte->rate)) {
15099                                         /**
15100                                          * We have four possible states here
15101                                          * having to do with the previous time
15102                                          * and this time.
15103                                          *   previous  |  this-time
15104                                          * A)     0      |     0   -- fill_cw not in the picture
15105                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
15106                                          * C)     1      |     1   -- all rates from fill_cw
15107                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
15108                                          *
15109                                          * For case A, C and D we don't allow a drop. But for
15110                                          * case B where we now our on our steady rate we do
15111                                          * allow a drop.
15112                                          *
15113                                          */
15114                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
15115                                                 goto done_w_hdwr;
15116                                 }
15117                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
15118                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
15119                                         if (rack_hw_rate_to_low &&
15120                                             (bw_est < rack_hw_rate_to_low)) {
15121                                                 /*
15122                                                  * The pacing rate is too low for hardware, but
15123                                                  * do allow hardware pacing to be restarted.
15124                                                  */
15125                                                 rack_log_hdwr_pacing(rack,
15126                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
15127                                                              0, 5);
15128                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15129                                                 rack->r_ctl.crte = NULL;
15130                                                 rack->rack_attempt_hdwr_pace = 0;
15131                                                 rack->rack_hdrw_pacing = 0;
15132                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15133                                                 goto done_w_hdwr;
15134                                         }
15135                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
15136                                                                    rack->rc_tp,
15137                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
15138                                                                    rate_wanted,
15139                                                                    RS_PACING_GEQ,
15140                                                                    &err, &rack->r_ctl.crte_prev_rate);
15141                                         if (nrte == NULL) {
15142                                                 /* Lost the rate */
15143                                                 rack->rack_hdrw_pacing = 0;
15144                                                 rack->r_ctl.crte = NULL;
15145                                                 rack_log_hdwr_pacing(rack,
15146                                                                      rate_wanted, 0, __LINE__,
15147                                                                      err, 1);
15148                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15149                                                 counter_u64_add(rack_hw_pace_lost, 1);
15150                                         } else if (nrte != rack->r_ctl.crte) {
15151                                                 rack->r_ctl.crte = nrte;
15152                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
15153                                                                                                          segsiz, 0,
15154                                                                                                          rack->r_ctl.crte,
15155                                                                                                          NULL);
15156                                                 rack_log_hdwr_pacing(rack,
15157                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15158                                                                      err, 2);
15159                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
15160                                         }
15161                                 } else {
15162                                         /* We just need to adjust the segment size */
15163                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15164                                         rack_log_hdwr_pacing(rack,
15165                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15166                                                              0, 4);
15167                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
15168                                 }
15169                         }
15170                 }
15171                 if ((rack->r_ctl.crte != NULL) &&
15172                     (rack->r_ctl.crte->rate == rate_wanted)) {
15173                         /*
15174                          * We need to add a extra if the rates
15175                          * are exactly matched. The idea is
15176                          * we want the software to make sure the
15177                          * queue is empty before adding more, this
15178                          * gives us N MSS extra pace times where
15179                          * N is our sysctl
15180                          */
15181                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15182                 }
15183 done_w_hdwr:
15184                 if (rack_limit_time_with_srtt &&
15185                     (rack->use_fixed_rate == 0) &&
15186 #ifdef NETFLIX_PEAKRATE
15187                     (rack->rc_tp->t_maxpeakrate == 0) &&
15188 #endif
15189                     (rack->rack_hdrw_pacing == 0)) {
15190                         /*
15191                          * Sanity check, we do not allow the pacing delay
15192                          * to be longer than the SRTT of the path. If it is
15193                          * a slow path, then adding a packet should increase
15194                          * the RTT and compensate for this i.e. the srtt will
15195                          * be greater so the allowed pacing time will be greater.
15196                          *
15197                          * Note this restriction is not for where a peak rate
15198                          * is set, we are doing fixed pacing or hardware pacing.
15199                          */
15200                         if (rack->rc_tp->t_srtt)
15201                                 srtt = rack->rc_tp->t_srtt;
15202                         else
15203                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
15204                         if (srtt < (uint64_t)slot) {
15205                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15206                                 slot = srtt;
15207                         }
15208                 }
15209                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15210         }
15211         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15212                 /*
15213                  * If this rate is seeing enobufs when it
15214                  * goes to send then either the nic is out
15215                  * of gas or we are mis-estimating the time
15216                  * somehow and not letting the queue empty
15217                  * completely. Lets add to the pacing time.
15218                  */
15219                 int hw_boost_delay;
15220
15221                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15222                 if (hw_boost_delay > rack_enobuf_hw_max)
15223                         hw_boost_delay = rack_enobuf_hw_max;
15224                 else if (hw_boost_delay < rack_enobuf_hw_min)
15225                         hw_boost_delay = rack_enobuf_hw_min;
15226                 slot += hw_boost_delay;
15227         }
15228         return (slot);
15229 }
15230
15231 static void
15232 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15233     tcp_seq startseq, uint32_t sb_offset)
15234 {
15235         struct rack_sendmap *my_rsm = NULL;
15236         struct rack_sendmap fe;
15237
15238         if (tp->t_state < TCPS_ESTABLISHED) {
15239                 /*
15240                  * We don't start any measurements if we are
15241                  * not at least established.
15242                  */
15243                 return;
15244         }
15245         if (tp->t_state >= TCPS_FIN_WAIT_1) {
15246                 /*
15247                  * We will get no more data into the SB
15248                  * this means we need to have the data available
15249                  * before we start a measurement.
15250                  */
15251
15252                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) <
15253                     max(rc_init_window(rack),
15254                         (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15255                         /* Nope not enough data */
15256                         return;
15257                 }
15258         }
15259         tp->t_flags |= TF_GPUTINPROG;
15260         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15261         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15262         tp->gput_seq = startseq;
15263         rack->app_limited_needs_set = 0;
15264         if (rack->in_probe_rtt)
15265                 rack->measure_saw_probe_rtt = 1;
15266         else if ((rack->measure_saw_probe_rtt) &&
15267                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15268                 rack->measure_saw_probe_rtt = 0;
15269         if (rack->rc_gp_filled)
15270                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15271         else {
15272                 /* Special case initial measurement */
15273                 struct timeval tv;
15274
15275                 tp->gput_ts = tcp_get_usecs(&tv);
15276                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15277         }
15278         /*
15279          * We take a guess out into the future,
15280          * if we have no measurement and no
15281          * initial rate, we measure the first
15282          * initial-windows worth of data to
15283          * speed up getting some GP measurement and
15284          * thus start pacing.
15285          */
15286         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15287                 rack->app_limited_needs_set = 1;
15288                 tp->gput_ack = startseq + max(rc_init_window(rack),
15289                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15290                 rack_log_pacing_delay_calc(rack,
15291                                            tp->gput_seq,
15292                                            tp->gput_ack,
15293                                            0,
15294                                            tp->gput_ts,
15295                                            rack->r_ctl.rc_app_limited_cnt,
15296                                            9,
15297                                            __LINE__, NULL, 0);
15298                 return;
15299         }
15300         if (sb_offset) {
15301                 /*
15302                  * We are out somewhere in the sb
15303                  * can we use the already outstanding data?
15304                  */
15305                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
15306                         /*
15307                          * Yes first one is good and in this case
15308                          * the tp->gput_ts is correctly set based on
15309                          * the last ack that arrived (no need to
15310                          * set things up when an ack comes in).
15311                          */
15312                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15313                         if ((my_rsm == NULL) ||
15314                             (my_rsm->r_rtr_cnt != 1)) {
15315                                 /* retransmission? */
15316                                 goto use_latest;
15317                         }
15318                 } else {
15319                         if (rack->r_ctl.rc_first_appl == NULL) {
15320                                 /*
15321                                  * If rc_first_appl is NULL
15322                                  * then the cnt should be 0.
15323                                  * This is probably an error, maybe
15324                                  * a KASSERT would be approprate.
15325                                  */
15326                                 goto use_latest;
15327                         }
15328                         /*
15329                          * If we have a marker pointer to the last one that is
15330                          * app limited we can use that, but we need to set
15331                          * things up so that when it gets ack'ed we record
15332                          * the ack time (if its not already acked).
15333                          */
15334                         rack->app_limited_needs_set = 1;
15335                         /*
15336                          * We want to get to the rsm that is either
15337                          * next with space i.e. over 1 MSS or the one
15338                          * after that (after the app-limited).
15339                          */
15340                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15341                                          rack->r_ctl.rc_first_appl);
15342                         if (my_rsm) {
15343                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15344                                         /* Have to use the next one */
15345                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15346                                                          my_rsm);
15347                                 else {
15348                                         /* Use after the first MSS of it is acked */
15349                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15350                                         goto start_set;
15351                                 }
15352                         }
15353                         if ((my_rsm == NULL) ||
15354                             (my_rsm->r_rtr_cnt != 1)) {
15355                                 /*
15356                                  * Either its a retransmit or
15357                                  * the last is the app-limited one.
15358                                  */
15359                                 goto use_latest;
15360                         }
15361                 }
15362                 tp->gput_seq = my_rsm->r_start;
15363 start_set:
15364                 if (my_rsm->r_flags & RACK_ACKED) {
15365                         /*
15366                          * This one has been acked use the arrival ack time
15367                          */
15368                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15369                         rack->app_limited_needs_set = 0;
15370                 }
15371                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15372                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15373                 rack_log_pacing_delay_calc(rack,
15374                                            tp->gput_seq,
15375                                            tp->gput_ack,
15376                                            (uint64_t)my_rsm,
15377                                            tp->gput_ts,
15378                                            rack->r_ctl.rc_app_limited_cnt,
15379                                            9,
15380                                            __LINE__, NULL, 0);
15381                 return;
15382         }
15383
15384 use_latest:
15385         /*
15386          * We don't know how long we may have been
15387          * idle or if this is the first-send. Lets
15388          * setup the flag so we will trim off
15389          * the first ack'd data so we get a true
15390          * measurement.
15391          */
15392         rack->app_limited_needs_set = 1;
15393         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15394         /* Find this guy so we can pull the send time */
15395         fe.r_start = startseq;
15396         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15397         if (my_rsm) {
15398                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15399                 if (my_rsm->r_flags & RACK_ACKED) {
15400                         /*
15401                          * Unlikely since its probably what was
15402                          * just transmitted (but I am paranoid).
15403                          */
15404                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15405                         rack->app_limited_needs_set = 0;
15406                 }
15407                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15408                         /* This also is unlikely */
15409                         tp->gput_seq = my_rsm->r_start;
15410                 }
15411         } else {
15412                 /*
15413                  * TSNH unless we have some send-map limit,
15414                  * and even at that it should not be hitting
15415                  * that limit (we should have stopped sending).
15416                  */
15417                 struct timeval tv;
15418
15419                 microuptime(&tv);
15420                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15421         }
15422         rack_log_pacing_delay_calc(rack,
15423                                    tp->gput_seq,
15424                                    tp->gput_ack,
15425                                    (uint64_t)my_rsm,
15426                                    tp->gput_ts,
15427                                    rack->r_ctl.rc_app_limited_cnt,
15428                                    9, __LINE__, NULL, 0);
15429 }
15430
15431 static inline uint32_t
15432 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15433     uint32_t avail, int32_t sb_offset)
15434 {
15435         uint32_t len;
15436         uint32_t sendwin;
15437
15438         if (tp->snd_wnd > cwnd_to_use)
15439                 sendwin = cwnd_to_use;
15440         else
15441                 sendwin = tp->snd_wnd;
15442         if (ctf_outstanding(tp) >= tp->snd_wnd) {
15443                 /* We never want to go over our peers rcv-window */
15444                 len = 0;
15445         } else {
15446                 uint32_t flight;
15447
15448                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15449                 if (flight >= sendwin) {
15450                         /*
15451                          * We have in flight what we are allowed by cwnd (if
15452                          * it was rwnd blocking it would have hit above out
15453                          * >= tp->snd_wnd).
15454                          */
15455                         return (0);
15456                 }
15457                 len = sendwin - flight;
15458                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15459                         /* We would send too much (beyond the rwnd) */
15460                         len = tp->snd_wnd - ctf_outstanding(tp);
15461                 }
15462                 if ((len + sb_offset) > avail) {
15463                         /*
15464                          * We don't have that much in the SB, how much is
15465                          * there?
15466                          */
15467                         len = avail - sb_offset;
15468                 }
15469         }
15470         return (len);
15471 }
15472
15473 static void
15474 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15475              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15476              int rsm_is_null, int optlen, int line, uint16_t mode)
15477 {
15478         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15479                 union tcp_log_stackspecific log;
15480                 struct timeval tv;
15481
15482                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15483                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15484                 log.u_bbr.flex1 = error;
15485                 log.u_bbr.flex2 = flags;
15486                 log.u_bbr.flex3 = rsm_is_null;
15487                 log.u_bbr.flex4 = ipoptlen;
15488                 log.u_bbr.flex5 = tp->rcv_numsacks;
15489                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15490                 log.u_bbr.flex7 = optlen;
15491                 log.u_bbr.flex8 = rack->r_fsb_inited;
15492                 log.u_bbr.applimited = rack->r_fast_output;
15493                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15494                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15495                 log.u_bbr.cwnd_gain = mode;
15496                 log.u_bbr.pkts_out = orig_len;
15497                 log.u_bbr.lt_epoch = len;
15498                 log.u_bbr.delivered = line;
15499                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15500                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15501                 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15502                                len, &log, false, NULL, NULL, 0, &tv);
15503         }
15504 }
15505
15506
15507 static struct mbuf *
15508 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15509                    struct rack_fast_send_blk *fsb,
15510                    int32_t seglimit, int32_t segsize, int hw_tls)
15511 {
15512 #ifdef KERN_TLS
15513         struct ktls_session *tls, *ntls;
15514 #ifdef INVARIANTS
15515         struct mbuf *start;
15516 #endif
15517 #endif
15518         struct mbuf *m, *n, **np, *smb;
15519         struct mbuf *top;
15520         int32_t off, soff;
15521         int32_t len = *plen;
15522         int32_t fragsize;
15523         int32_t len_cp = 0;
15524         uint32_t mlen, frags;
15525
15526         soff = off = the_off;
15527         smb = m = the_m;
15528         np = &top;
15529         top = NULL;
15530 #ifdef KERN_TLS
15531         if (hw_tls && (m->m_flags & M_EXTPG))
15532                 tls = m->m_epg_tls;
15533         else
15534                 tls = NULL;
15535 #ifdef INVARIANTS
15536         start = m;
15537 #endif
15538 #endif
15539         while (len > 0) {
15540                 if (m == NULL) {
15541                         *plen = len_cp;
15542                         break;
15543                 }
15544 #ifdef KERN_TLS
15545                 if (hw_tls) {
15546                         if (m->m_flags & M_EXTPG)
15547                                 ntls = m->m_epg_tls;
15548                         else
15549                                 ntls = NULL;
15550
15551                         /*
15552                          * Avoid mixing TLS records with handshake
15553                          * data or TLS records from different
15554                          * sessions.
15555                          */
15556                         if (tls != ntls) {
15557                                 MPASS(m != start);
15558                                 *plen = len_cp;
15559                                 break;
15560                         }
15561                 }
15562 #endif
15563                 mlen = min(len, m->m_len - off);
15564                 if (seglimit) {
15565                         /*
15566                          * For M_EXTPG mbufs, add 3 segments
15567                          * + 1 in case we are crossing page boundaries
15568                          * + 2 in case the TLS hdr/trailer are used
15569                          * It is cheaper to just add the segments
15570                          * than it is to take the cache miss to look
15571                          * at the mbuf ext_pgs state in detail.
15572                          */
15573                         if (m->m_flags & M_EXTPG) {
15574                                 fragsize = min(segsize, PAGE_SIZE);
15575                                 frags = 3;
15576                         } else {
15577                                 fragsize = segsize;
15578                                 frags = 0;
15579                         }
15580
15581                         /* Break if we really can't fit anymore. */
15582                         if ((frags + 1) >= seglimit) {
15583                                 *plen = len_cp;
15584                                 break;
15585                         }
15586
15587                         /*
15588                          * Reduce size if you can't copy the whole
15589                          * mbuf. If we can't copy the whole mbuf, also
15590                          * adjust len so the loop will end after this
15591                          * mbuf.
15592                          */
15593                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15594                                 mlen = (seglimit - frags - 1) * fragsize;
15595                                 len = mlen;
15596                                 *plen = len_cp + len;
15597                         }
15598                         frags += howmany(mlen, fragsize);
15599                         if (frags == 0)
15600                                 frags++;
15601                         seglimit -= frags;
15602                         KASSERT(seglimit > 0,
15603                             ("%s: seglimit went too low", __func__));
15604                 }
15605                 n = m_get(M_NOWAIT, m->m_type);
15606                 *np = n;
15607                 if (n == NULL)
15608                         goto nospace;
15609                 n->m_len = mlen;
15610                 soff += mlen;
15611                 len_cp += n->m_len;
15612                 if (m->m_flags & (M_EXT|M_EXTPG)) {
15613                         n->m_data = m->m_data + off;
15614                         mb_dupcl(n, m);
15615                 } else {
15616                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15617                             (u_int)n->m_len);
15618                 }
15619                 len -= n->m_len;
15620                 off = 0;
15621                 m = m->m_next;
15622                 np = &n->m_next;
15623                 if (len || (soff == smb->m_len)) {
15624                         /*
15625                          * We have more so we move forward  or
15626                          * we have consumed the entire mbuf and
15627                          * len has fell to 0.
15628                          */
15629                         soff = 0;
15630                         smb = m;
15631                 }
15632
15633         }
15634         if (fsb != NULL) {
15635                 fsb->m = smb;
15636                 fsb->off = soff;
15637                 if (smb) {
15638                         /*
15639                          * Save off the size of the mbuf. We do
15640                          * this so that we can recognize when it
15641                          * has been trimmed by sbcut() as acks
15642                          * come in.
15643                          */
15644                         fsb->o_m_len = smb->m_len;
15645                 } else {
15646                         /*
15647                          * This is the case where the next mbuf went to NULL. This
15648                          * means with this copy we have sent everything in the sb.
15649                          * In theory we could clear the fast_output flag, but lets
15650                          * not since its possible that we could get more added
15651                          * and acks that call the extend function which would let
15652                          * us send more.
15653                          */
15654                         fsb->o_m_len = 0;
15655                 }
15656         }
15657         return (top);
15658 nospace:
15659         if (top)
15660                 m_freem(top);
15661         return (NULL);
15662
15663 }
15664
15665 /*
15666  * This is a copy of m_copym(), taking the TSO segment size/limit
15667  * constraints into account, and advancing the sndptr as it goes.
15668  */
15669 static struct mbuf *
15670 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15671                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15672 {
15673         struct mbuf *m, *n;
15674         int32_t soff;
15675
15676         soff = rack->r_ctl.fsb.off;
15677         m = rack->r_ctl.fsb.m;
15678         if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15679                 /*
15680                  * The mbuf had the front of it chopped off by an ack
15681                  * we need to adjust the soff/off by that difference.
15682                  */
15683                 uint32_t delta;
15684
15685                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15686                 soff -= delta;
15687         } else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15688                 /*
15689                  * The mbuf was expanded probably by
15690                  * a m_compress. Just update o_m_len.
15691                  */
15692                 rack->r_ctl.fsb.o_m_len = m->m_len;
15693         }
15694         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15695         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15696         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15697                                  __FUNCTION__,
15698                                  rack, *plen, m, m->m_len));
15699         /* Save off the right location before we copy and advance */
15700         *s_soff = soff;
15701         *s_mb = rack->r_ctl.fsb.m;
15702         n = rack_fo_base_copym(m, soff, plen,
15703                                &rack->r_ctl.fsb,
15704                                seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15705         return (n);
15706 }
15707
15708 static int
15709 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15710                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15711 {
15712         /*
15713          * Enter the fast retransmit path. We are given that a sched_pin is
15714          * in place (if accounting is compliled in) and the cycle count taken
15715          * at the entry is in the ts_val. The concept her is that the rsm
15716          * now holds the mbuf offsets and such so we can directly transmit
15717          * without a lot of overhead, the len field is already set for
15718          * us to prohibit us from sending too much (usually its 1MSS).
15719          */
15720         struct ip *ip = NULL;
15721         struct udphdr *udp = NULL;
15722         struct tcphdr *th = NULL;
15723         struct mbuf *m = NULL;
15724         struct inpcb *inp;
15725         uint8_t *cpto;
15726         struct tcp_log_buffer *lgb;
15727 #ifdef TCP_ACCOUNTING
15728         uint64_t crtsc;
15729         int cnt_thru = 1;
15730 #endif
15731         struct tcpopt to;
15732         u_char opt[TCP_MAXOLEN];
15733         uint32_t hdrlen, optlen;
15734         int32_t slot, segsiz, max_val, tso = 0, error, ulen = 0;
15735         uint16_t flags;
15736         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15737         uint32_t if_hw_tsomaxsegsize;
15738
15739 #ifdef INET6
15740         struct ip6_hdr *ip6 = NULL;
15741
15742         if (rack->r_is_v6) {
15743                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15744                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15745         } else
15746 #endif                          /* INET6 */
15747         {
15748                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15749                 hdrlen = sizeof(struct tcpiphdr);
15750         }
15751         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15752                 goto failed;
15753         }
15754         if (doing_tlp) {
15755                 /* Its a TLP add the flag, it may already be there but be sure */
15756                 rsm->r_flags |= RACK_TLP;
15757         } else {
15758                 /* If it was a TLP it is not not on this retransmit */
15759                 rsm->r_flags &= ~RACK_TLP;
15760         }
15761         startseq = rsm->r_start;
15762         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15763         inp = rack->rc_inp;
15764         to.to_flags = 0;
15765         flags = tcp_outflags[tp->t_state];
15766         if (flags & (TH_SYN|TH_RST)) {
15767                 goto failed;
15768         }
15769         if (rsm->r_flags & RACK_HAS_FIN) {
15770                 /* We can't send a FIN here */
15771                 goto failed;
15772         }
15773         if (flags & TH_FIN) {
15774                 /* We never send a FIN */
15775                 flags &= ~TH_FIN;
15776         }
15777         if (tp->t_flags & TF_RCVD_TSTMP) {
15778                 to.to_tsval = ms_cts + tp->ts_offset;
15779                 to.to_tsecr = tp->ts_recent;
15780                 to.to_flags = TOF_TS;
15781         }
15782         optlen = tcp_addoptions(&to, opt);
15783         hdrlen += optlen;
15784         udp = rack->r_ctl.fsb.udp;
15785         if (udp)
15786                 hdrlen += sizeof(struct udphdr);
15787         if (rack->r_ctl.rc_pace_max_segs)
15788                 max_val = rack->r_ctl.rc_pace_max_segs;
15789         else if (rack->rc_user_set_max_segs)
15790                 max_val = rack->rc_user_set_max_segs * segsiz;
15791         else
15792                 max_val = len;
15793         if ((tp->t_flags & TF_TSO) &&
15794             V_tcp_do_tso &&
15795             (len > segsiz) &&
15796             (tp->t_port == 0))
15797                 tso = 1;
15798 #ifdef INET6
15799         if (MHLEN < hdrlen + max_linkhdr)
15800                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15801         else
15802 #endif
15803                 m = m_gethdr(M_NOWAIT, MT_DATA);
15804         if (m == NULL)
15805                 goto failed;
15806         m->m_data += max_linkhdr;
15807         m->m_len = hdrlen;
15808         th = rack->r_ctl.fsb.th;
15809         /* Establish the len to send */
15810         if (len > max_val)
15811                 len = max_val;
15812         if ((tso) && (len + optlen > tp->t_maxseg)) {
15813                 uint32_t if_hw_tsomax;
15814                 int32_t max_len;
15815
15816                 /* extract TSO information */
15817                 if_hw_tsomax = tp->t_tsomax;
15818                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15819                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15820                 /*
15821                  * Check if we should limit by maximum payload
15822                  * length:
15823                  */
15824                 if (if_hw_tsomax != 0) {
15825                         /* compute maximum TSO length */
15826                         max_len = (if_hw_tsomax - hdrlen -
15827                                    max_linkhdr);
15828                         if (max_len <= 0) {
15829                                 goto failed;
15830                         } else if (len > max_len) {
15831                                 len = max_len;
15832                         }
15833                 }
15834                 if (len <= segsiz) {
15835                         /*
15836                          * In case there are too many small fragments don't
15837                          * use TSO:
15838                          */
15839                         tso = 0;
15840                 }
15841         } else {
15842                 tso = 0;
15843         }
15844         if ((tso == 0) && (len > segsiz))
15845                 len = segsiz;
15846         if ((len == 0) ||
15847             (len <= MHLEN - hdrlen - max_linkhdr)) {
15848                 goto failed;
15849         }
15850         th->th_seq = htonl(rsm->r_start);
15851         th->th_ack = htonl(tp->rcv_nxt);
15852         /*
15853          * The PUSH bit should only be applied
15854          * if the full retransmission is made. If
15855          * we are sending less than this is the
15856          * left hand edge and should not have
15857          * the PUSH bit.
15858          */
15859         if ((rsm->r_flags & RACK_HAD_PUSH) &&
15860             (len == (rsm->r_end - rsm->r_start)))
15861                 flags |= TH_PUSH;
15862         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15863         if (th->th_win == 0) {
15864                 tp->t_sndzerowin++;
15865                 tp->t_flags |= TF_RXWIN0SENT;
15866         } else
15867                 tp->t_flags &= ~TF_RXWIN0SENT;
15868         if (rsm->r_flags & RACK_TLP) {
15869                 /*
15870                  * TLP should not count in retran count, but
15871                  * in its own bin
15872                  */
15873                 counter_u64_add(rack_tlp_retran, 1);
15874                 counter_u64_add(rack_tlp_retran_bytes, len);
15875         } else {
15876                 tp->t_sndrexmitpack++;
15877                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15878                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15879         }
15880 #ifdef STATS
15881         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15882                                  len);
15883 #endif
15884         if (rsm->m == NULL)
15885                 goto failed;
15886         if (rsm->orig_m_len != rsm->m->m_len) {
15887                 /* Fix up the orig_m_len and possibly the mbuf offset */
15888                 rack_adjust_orig_mlen(rsm);
15889         }
15890         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15891         if (len <= segsiz) {
15892                 /*
15893                  * Must have ran out of mbufs for the copy
15894                  * shorten it to no longer need tso. Lets
15895                  * not put on sendalot since we are low on
15896                  * mbufs.
15897                  */
15898                 tso = 0;
15899         }
15900         if ((m->m_next == NULL) || (len <= 0)){
15901                 goto failed;
15902         }
15903         if (udp) {
15904                 if (rack->r_is_v6)
15905                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15906                 else
15907                         ulen = hdrlen + len - sizeof(struct ip);
15908                 udp->uh_ulen = htons(ulen);
15909         }
15910         m->m_pkthdr.rcvif = (struct ifnet *)0;
15911         if (TCPS_HAVERCVDSYN(tp->t_state) &&
15912             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
15913                 int ect = tcp_ecn_output_established(tp, &flags, len, true);
15914                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
15915                     (tp->t_flags2 & TF2_ECN_SND_ECE))
15916                     tp->t_flags2 &= ~TF2_ECN_SND_ECE;
15917 #ifdef INET6
15918                 if (rack->r_is_v6) {
15919                     ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
15920                     ip6->ip6_flow |= htonl(ect << 20);
15921                 }
15922                 else
15923 #endif
15924                 {
15925                     ip->ip_tos &= ~IPTOS_ECN_MASK;
15926                     ip->ip_tos |= ect;
15927                 }
15928         }
15929         tcp_set_flags(th, flags);
15930         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15931 #ifdef INET6
15932         if (rack->r_is_v6) {
15933                 if (tp->t_port) {
15934                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15935                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15936                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15937                         th->th_sum = htons(0);
15938                         UDPSTAT_INC(udps_opackets);
15939                 } else {
15940                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15941                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15942                         th->th_sum = in6_cksum_pseudo(ip6,
15943                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15944                                                       0);
15945                 }
15946         }
15947 #endif
15948 #if defined(INET6) && defined(INET)
15949         else
15950 #endif
15951 #ifdef INET
15952         {
15953                 if (tp->t_port) {
15954                         m->m_pkthdr.csum_flags = CSUM_UDP;
15955                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15956                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15957                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15958                         th->th_sum = htons(0);
15959                         UDPSTAT_INC(udps_opackets);
15960                 } else {
15961                         m->m_pkthdr.csum_flags = CSUM_TCP;
15962                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15963                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15964                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15965                                                                         IPPROTO_TCP + len + optlen));
15966                 }
15967                 /* IP version must be set here for ipv4/ipv6 checking later */
15968                 KASSERT(ip->ip_v == IPVERSION,
15969                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15970         }
15971 #endif
15972         if (tso) {
15973                 KASSERT(len > tp->t_maxseg - optlen,
15974                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15975                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15976                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15977         }
15978 #ifdef INET6
15979         if (rack->r_is_v6) {
15980                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15981                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15982                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15983                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15984                 else
15985                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15986         }
15987 #endif
15988 #if defined(INET) && defined(INET6)
15989         else
15990 #endif
15991 #ifdef INET
15992         {
15993                 ip->ip_len = htons(m->m_pkthdr.len);
15994                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15995                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15996                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15997                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15998                                 ip->ip_off |= htons(IP_DF);
15999                         }
16000                 } else {
16001                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16002                 }
16003         }
16004 #endif
16005         /* Time to copy in our header */
16006         cpto = mtod(m, uint8_t *);
16007         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16008         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16009         if (optlen) {
16010                 bcopy(opt, th + 1, optlen);
16011                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16012         } else {
16013                 th->th_off = sizeof(struct tcphdr) >> 2;
16014         }
16015         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16016                 union tcp_log_stackspecific log;
16017
16018                 if (rsm->r_flags & RACK_RWND_COLLAPSED) {
16019                         rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
16020                         counter_u64_add(rack_collapsed_win_rxt, 1);
16021                         counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
16022                 }
16023                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16024                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16025                 if (rack->rack_no_prr)
16026                         log.u_bbr.flex1 = 0;
16027                 else
16028                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16029                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16030                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16031                 log.u_bbr.flex4 = max_val;
16032                 log.u_bbr.flex5 = 0;
16033                 /* Save off the early/late values */
16034                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16035                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16036                 log.u_bbr.bw_inuse = rack_get_bw(rack);
16037                 if (doing_tlp == 0)
16038                         log.u_bbr.flex8 = 1;
16039                 else
16040                         log.u_bbr.flex8 = 2;
16041                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16042                 log.u_bbr.flex7 = 55;
16043                 log.u_bbr.pkts_out = tp->t_maxseg;
16044                 log.u_bbr.timeStamp = cts;
16045                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16046                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16047                 log.u_bbr.delivered = 0;
16048                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16049                                      len, &log, false, NULL, NULL, 0, tv);
16050         } else
16051                 lgb = NULL;
16052 #ifdef INET6
16053         if (rack->r_is_v6) {
16054                 error = ip6_output(m, NULL,
16055                                    &inp->inp_route6,
16056                                    0, NULL, NULL, inp);
16057         }
16058 #endif
16059 #if defined(INET) && defined(INET6)
16060         else
16061 #endif
16062 #ifdef INET
16063         {
16064                 error = ip_output(m, NULL,
16065                                   &inp->inp_route,
16066                                   0, 0, inp);
16067         }
16068 #endif
16069         m = NULL;
16070         if (lgb) {
16071                 lgb->tlb_errno = error;
16072                 lgb = NULL;
16073         }
16074         if (error) {
16075                 goto failed;
16076         }
16077         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
16078                         rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
16079         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
16080                 rack->rc_tlp_in_progress = 1;
16081                 rack->r_ctl.rc_tlp_cnt_out++;
16082         }
16083         if (error == 0) {
16084                 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
16085                 if (doing_tlp) {
16086                         rack->rc_last_sent_tlp_past_cumack = 0;
16087                         rack->rc_last_sent_tlp_seq_valid = 1;
16088                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
16089                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
16090                 }
16091         }
16092         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16093         rack->forced_ack = 0;   /* If we send something zap the FA flag */
16094         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
16095                 rack->r_ctl.retran_during_recovery += len;
16096         {
16097                 int idx;
16098
16099                 idx = (len / segsiz) + 3;
16100                 if (idx >= TCP_MSS_ACCT_ATIMER)
16101                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16102                 else
16103                         counter_u64_add(rack_out_size[idx], 1);
16104         }
16105         if (tp->t_rtttime == 0) {
16106                 tp->t_rtttime = ticks;
16107                 tp->t_rtseq = startseq;
16108                 KMOD_TCPSTAT_INC(tcps_segstimed);
16109         }
16110         counter_u64_add(rack_fto_rsm_send, 1);
16111         if (error && (error == ENOBUFS)) {
16112                 if (rack->r_ctl.crte != NULL) {
16113                         rack_trace_point(rack, RACK_TP_HWENOBUF);
16114                 } else
16115                         rack_trace_point(rack, RACK_TP_ENOBUF);
16116                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
16117                 if (rack->rc_enobuf < 0x7f)
16118                         rack->rc_enobuf++;
16119                 if (slot < (10 * HPTS_USEC_IN_MSEC))
16120                         slot = 10 * HPTS_USEC_IN_MSEC;
16121         } else
16122                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
16123         if ((slot == 0) ||
16124             (rack->rc_always_pace == 0) ||
16125             (rack->r_rr_config == 1)) {
16126                 /*
16127                  * We have no pacing set or we
16128                  * are using old-style rack or
16129                  * we are overridden to use the old 1ms pacing.
16130                  */
16131                 slot = rack->r_ctl.rc_min_to;
16132         }
16133         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
16134 #ifdef TCP_ACCOUNTING
16135         crtsc = get_cyclecount();
16136         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16137                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16138         }
16139         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16140         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16141                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16142         }
16143         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16144         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16145                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
16146         }
16147         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
16148         sched_unpin();
16149 #endif
16150         return (0);
16151 failed:
16152         if (m)
16153                 m_free(m);
16154         return (-1);
16155 }
16156
16157 static void
16158 rack_sndbuf_autoscale(struct tcp_rack *rack)
16159 {
16160         /*
16161          * Automatic sizing of send socket buffer.  Often the send buffer
16162          * size is not optimally adjusted to the actual network conditions
16163          * at hand (delay bandwidth product).  Setting the buffer size too
16164          * small limits throughput on links with high bandwidth and high
16165          * delay (eg. trans-continental/oceanic links).  Setting the
16166          * buffer size too big consumes too much real kernel memory,
16167          * especially with many connections on busy servers.
16168          *
16169          * The criteria to step up the send buffer one notch are:
16170          *  1. receive window of remote host is larger than send buffer
16171          *     (with a fudge factor of 5/4th);
16172          *  2. send buffer is filled to 7/8th with data (so we actually
16173          *     have data to make use of it);
16174          *  3. send buffer fill has not hit maximal automatic size;
16175          *  4. our send window (slow start and cogestion controlled) is
16176          *     larger than sent but unacknowledged data in send buffer.
16177          *
16178          * Note that the rack version moves things much faster since
16179          * we want to avoid hitting cache lines in the rack_fast_output()
16180          * path so this is called much less often and thus moves
16181          * the SB forward by a percentage.
16182          */
16183         struct socket *so;
16184         struct tcpcb *tp;
16185         uint32_t sendwin, scaleup;
16186
16187         tp = rack->rc_tp;
16188         so = rack->rc_inp->inp_socket;
16189         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16190         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16191                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16192                     sbused(&so->so_snd) >=
16193                     (so->so_snd.sb_hiwat / 8 * 7) &&
16194                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16195                     sendwin >= (sbused(&so->so_snd) -
16196                     (tp->snd_nxt - tp->snd_una))) {
16197                         if (rack_autosndbuf_inc)
16198                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16199                         else
16200                                 scaleup = V_tcp_autosndbuf_inc;
16201                         if (scaleup < V_tcp_autosndbuf_inc)
16202                                 scaleup = V_tcp_autosndbuf_inc;
16203                         scaleup += so->so_snd.sb_hiwat;
16204                         if (scaleup > V_tcp_autosndbuf_max)
16205                                 scaleup = V_tcp_autosndbuf_max;
16206                         if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
16207                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16208                 }
16209         }
16210 }
16211
16212 static int
16213 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16214                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16215 {
16216         /*
16217          * Enter to do fast output. We are given that the sched_pin is
16218          * in place (if accounting is compiled in) and the cycle count taken
16219          * at entry is in place in ts_val. The idea here is that
16220          * we know how many more bytes needs to be sent (presumably either
16221          * during pacing or to fill the cwnd and that was greater than
16222          * the max-burst). We have how much to send and all the info we
16223          * need to just send.
16224          */
16225         struct ip *ip = NULL;
16226         struct udphdr *udp = NULL;
16227         struct tcphdr *th = NULL;
16228         struct mbuf *m, *s_mb;
16229         struct inpcb *inp;
16230         uint8_t *cpto;
16231         struct tcp_log_buffer *lgb;
16232 #ifdef TCP_ACCOUNTING
16233         uint64_t crtsc;
16234 #endif
16235         struct tcpopt to;
16236         u_char opt[TCP_MAXOLEN];
16237         uint32_t hdrlen, optlen;
16238 #ifdef TCP_ACCOUNTING
16239         int cnt_thru = 1;
16240 #endif
16241         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
16242         uint16_t flags;
16243         uint32_t s_soff;
16244         uint32_t if_hw_tsomaxsegcount = 0, startseq;
16245         uint32_t if_hw_tsomaxsegsize;
16246         uint16_t add_flag = RACK_SENT_FP;
16247 #ifdef INET6
16248         struct ip6_hdr *ip6 = NULL;
16249
16250         if (rack->r_is_v6) {
16251                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16252                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16253         } else
16254 #endif                          /* INET6 */
16255         {
16256                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16257                 hdrlen = sizeof(struct tcpiphdr);
16258         }
16259         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16260                 m = NULL;
16261                 goto failed;
16262         }
16263         startseq = tp->snd_max;
16264         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16265         inp = rack->rc_inp;
16266         len = rack->r_ctl.fsb.left_to_send;
16267         to.to_flags = 0;
16268         flags = rack->r_ctl.fsb.tcp_flags;
16269         if (tp->t_flags & TF_RCVD_TSTMP) {
16270                 to.to_tsval = ms_cts + tp->ts_offset;
16271                 to.to_tsecr = tp->ts_recent;
16272                 to.to_flags = TOF_TS;
16273         }
16274         optlen = tcp_addoptions(&to, opt);
16275         hdrlen += optlen;
16276         udp = rack->r_ctl.fsb.udp;
16277         if (udp)
16278                 hdrlen += sizeof(struct udphdr);
16279         if (rack->r_ctl.rc_pace_max_segs)
16280                 max_val = rack->r_ctl.rc_pace_max_segs;
16281         else if (rack->rc_user_set_max_segs)
16282                 max_val = rack->rc_user_set_max_segs * segsiz;
16283         else
16284                 max_val = len;
16285         if ((tp->t_flags & TF_TSO) &&
16286             V_tcp_do_tso &&
16287             (len > segsiz) &&
16288             (tp->t_port == 0))
16289                 tso = 1;
16290 again:
16291 #ifdef INET6
16292         if (MHLEN < hdrlen + max_linkhdr)
16293                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16294         else
16295 #endif
16296                 m = m_gethdr(M_NOWAIT, MT_DATA);
16297         if (m == NULL)
16298                 goto failed;
16299         m->m_data += max_linkhdr;
16300         m->m_len = hdrlen;
16301         th = rack->r_ctl.fsb.th;
16302         /* Establish the len to send */
16303         if (len > max_val)
16304                 len = max_val;
16305         if ((tso) && (len + optlen > tp->t_maxseg)) {
16306                 uint32_t if_hw_tsomax;
16307                 int32_t max_len;
16308
16309                 /* extract TSO information */
16310                 if_hw_tsomax = tp->t_tsomax;
16311                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16312                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16313                 /*
16314                  * Check if we should limit by maximum payload
16315                  * length:
16316                  */
16317                 if (if_hw_tsomax != 0) {
16318                         /* compute maximum TSO length */
16319                         max_len = (if_hw_tsomax - hdrlen -
16320                                    max_linkhdr);
16321                         if (max_len <= 0) {
16322                                 goto failed;
16323                         } else if (len > max_len) {
16324                                 len = max_len;
16325                         }
16326                 }
16327                 if (len <= segsiz) {
16328                         /*
16329                          * In case there are too many small fragments don't
16330                          * use TSO:
16331                          */
16332                         tso = 0;
16333                 }
16334         } else {
16335                 tso = 0;
16336         }
16337         if ((tso == 0) && (len > segsiz))
16338                 len = segsiz;
16339         if ((len == 0) ||
16340             (len <= MHLEN - hdrlen - max_linkhdr)) {
16341                 goto failed;
16342         }
16343         sb_offset = tp->snd_max - tp->snd_una;
16344         th->th_seq = htonl(tp->snd_max);
16345         th->th_ack = htonl(tp->rcv_nxt);
16346         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16347         if (th->th_win == 0) {
16348                 tp->t_sndzerowin++;
16349                 tp->t_flags |= TF_RXWIN0SENT;
16350         } else
16351                 tp->t_flags &= ~TF_RXWIN0SENT;
16352         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
16353         KMOD_TCPSTAT_INC(tcps_sndpack);
16354         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16355 #ifdef STATS
16356         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16357                                  len);
16358 #endif
16359         if (rack->r_ctl.fsb.m == NULL)
16360                 goto failed;
16361
16362         /* s_mb and s_soff are saved for rack_log_output */
16363         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16364                                     &s_mb, &s_soff);
16365         if (len <= segsiz) {
16366                 /*
16367                  * Must have ran out of mbufs for the copy
16368                  * shorten it to no longer need tso. Lets
16369                  * not put on sendalot since we are low on
16370                  * mbufs.
16371                  */
16372                 tso = 0;
16373         }
16374         if (rack->r_ctl.fsb.rfo_apply_push &&
16375             (len == rack->r_ctl.fsb.left_to_send)) {
16376                 flags |= TH_PUSH;
16377                 add_flag |= RACK_HAD_PUSH;
16378         }
16379         if ((m->m_next == NULL) || (len <= 0)){
16380                 goto failed;
16381         }
16382         if (udp) {
16383                 if (rack->r_is_v6)
16384                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
16385                 else
16386                         ulen = hdrlen + len - sizeof(struct ip);
16387                 udp->uh_ulen = htons(ulen);
16388         }
16389         m->m_pkthdr.rcvif = (struct ifnet *)0;
16390         if (TCPS_HAVERCVDSYN(tp->t_state) &&
16391             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
16392                 int ect = tcp_ecn_output_established(tp, &flags, len, false);
16393                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
16394                     (tp->t_flags2 & TF2_ECN_SND_ECE))
16395                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
16396 #ifdef INET6
16397                 if (rack->r_is_v6) {
16398                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
16399                         ip6->ip6_flow |= htonl(ect << 20);
16400                 }
16401                 else
16402 #endif
16403                 {
16404                         ip->ip_tos &= ~IPTOS_ECN_MASK;
16405                         ip->ip_tos |= ect;
16406                 }
16407         }
16408         tcp_set_flags(th, flags);
16409         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
16410 #ifdef INET6
16411         if (rack->r_is_v6) {
16412                 if (tp->t_port) {
16413                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16414                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16415                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16416                         th->th_sum = htons(0);
16417                         UDPSTAT_INC(udps_opackets);
16418                 } else {
16419                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16420                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16421                         th->th_sum = in6_cksum_pseudo(ip6,
16422                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16423                                                       0);
16424                 }
16425         }
16426 #endif
16427 #if defined(INET6) && defined(INET)
16428         else
16429 #endif
16430 #ifdef INET
16431         {
16432                 if (tp->t_port) {
16433                         m->m_pkthdr.csum_flags = CSUM_UDP;
16434                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16435                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16436                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16437                         th->th_sum = htons(0);
16438                         UDPSTAT_INC(udps_opackets);
16439                 } else {
16440                         m->m_pkthdr.csum_flags = CSUM_TCP;
16441                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16442                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
16443                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16444                                                                         IPPROTO_TCP + len + optlen));
16445                 }
16446                 /* IP version must be set here for ipv4/ipv6 checking later */
16447                 KASSERT(ip->ip_v == IPVERSION,
16448                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
16449         }
16450 #endif
16451         if (tso) {
16452                 KASSERT(len > tp->t_maxseg - optlen,
16453                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
16454                 m->m_pkthdr.csum_flags |= CSUM_TSO;
16455                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16456         }
16457 #ifdef INET6
16458         if (rack->r_is_v6) {
16459                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16460                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16461                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16462                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16463                 else
16464                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16465         }
16466 #endif
16467 #if defined(INET) && defined(INET6)
16468         else
16469 #endif
16470 #ifdef INET
16471         {
16472                 ip->ip_len = htons(m->m_pkthdr.len);
16473                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16474                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16475                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16476                         if (tp->t_port == 0 || len < V_tcp_minmss) {
16477                                 ip->ip_off |= htons(IP_DF);
16478                         }
16479                 } else {
16480                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16481                 }
16482         }
16483 #endif
16484         /* Time to copy in our header */
16485         cpto = mtod(m, uint8_t *);
16486         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16487         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16488         if (optlen) {
16489                 bcopy(opt, th + 1, optlen);
16490                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16491         } else {
16492                 th->th_off = sizeof(struct tcphdr) >> 2;
16493         }
16494         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16495                 union tcp_log_stackspecific log;
16496
16497                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16498                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16499                 if (rack->rack_no_prr)
16500                         log.u_bbr.flex1 = 0;
16501                 else
16502                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16503                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16504                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16505                 log.u_bbr.flex4 = max_val;
16506                 log.u_bbr.flex5 = 0;
16507                 /* Save off the early/late values */
16508                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16509                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16510                 log.u_bbr.bw_inuse = rack_get_bw(rack);
16511                 log.u_bbr.flex8 = 0;
16512                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16513                 log.u_bbr.flex7 = 44;
16514                 log.u_bbr.pkts_out = tp->t_maxseg;
16515                 log.u_bbr.timeStamp = cts;
16516                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16517                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16518                 log.u_bbr.delivered = 0;
16519                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16520                                      len, &log, false, NULL, NULL, 0, tv);
16521         } else
16522                 lgb = NULL;
16523 #ifdef INET6
16524         if (rack->r_is_v6) {
16525                 error = ip6_output(m, NULL,
16526                                    &inp->inp_route6,
16527                                    0, NULL, NULL, inp);
16528         }
16529 #endif
16530 #if defined(INET) && defined(INET6)
16531         else
16532 #endif
16533 #ifdef INET
16534         {
16535                 error = ip_output(m, NULL,
16536                                   &inp->inp_route,
16537                                   0, 0, inp);
16538         }
16539 #endif
16540         if (lgb) {
16541                 lgb->tlb_errno = error;
16542                 lgb = NULL;
16543         }
16544         if (error) {
16545                 *send_err = error;
16546                 m = NULL;
16547                 goto failed;
16548         }
16549         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16550                         NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16551         m = NULL;
16552         if (tp->snd_una == tp->snd_max) {
16553                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
16554                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16555                 tp->t_acktime = ticks;
16556         }
16557         if (error == 0)
16558                 tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16559
16560         rack->forced_ack = 0;   /* If we send something zap the FA flag */
16561         tot_len += len;
16562         if ((tp->t_flags & TF_GPUTINPROG) == 0)
16563                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16564         tp->snd_max += len;
16565         tp->snd_nxt = tp->snd_max;
16566         {
16567                 int idx;
16568
16569                 idx = (len / segsiz) + 3;
16570                 if (idx >= TCP_MSS_ACCT_ATIMER)
16571                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16572                 else
16573                         counter_u64_add(rack_out_size[idx], 1);
16574         }
16575         if (len <= rack->r_ctl.fsb.left_to_send)
16576                 rack->r_ctl.fsb.left_to_send -= len;
16577         else
16578                 rack->r_ctl.fsb.left_to_send = 0;
16579         if (rack->r_ctl.fsb.left_to_send < segsiz) {
16580                 rack->r_fast_output = 0;
16581                 rack->r_ctl.fsb.left_to_send = 0;
16582                 /* At the end of fast_output scale up the sb */
16583                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16584                 rack_sndbuf_autoscale(rack);
16585                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16586         }
16587         if (tp->t_rtttime == 0) {
16588                 tp->t_rtttime = ticks;
16589                 tp->t_rtseq = startseq;
16590                 KMOD_TCPSTAT_INC(tcps_segstimed);
16591         }
16592         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16593             (max_val > len) &&
16594             (tso == 0)) {
16595                 max_val -= len;
16596                 len = segsiz;
16597                 th = rack->r_ctl.fsb.th;
16598 #ifdef TCP_ACCOUNTING
16599                 cnt_thru++;
16600 #endif
16601                 goto again;
16602         }
16603         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16604         counter_u64_add(rack_fto_send, 1);
16605         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16606         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16607 #ifdef TCP_ACCOUNTING
16608         crtsc = get_cyclecount();
16609         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16610                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16611         }
16612         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16613         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16614                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16615         }
16616         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16617         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16618                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16619         }
16620         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16621         sched_unpin();
16622 #endif
16623         return (0);
16624 failed:
16625         if (m)
16626                 m_free(m);
16627         rack->r_fast_output = 0;
16628         return (-1);
16629 }
16630
16631 static struct rack_sendmap *
16632 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
16633 {
16634         struct rack_sendmap *rsm = NULL;
16635         struct rack_sendmap fe;
16636         int thresh;
16637
16638 restart:
16639         fe.r_start = rack->r_ctl.last_collapse_point;
16640         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
16641         if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
16642                 /* Nothing, strange turn off validity  */
16643                 rack->r_collapse_point_valid = 0;
16644                 return (NULL);
16645         }
16646         /* Can we send it yet? */
16647         if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
16648                 /*
16649                  * Receiver window has not grown enough for
16650                  * the segment to be put on the wire.
16651                  */
16652                 return (NULL);
16653         }
16654         if (rsm->r_flags & RACK_ACKED) {
16655                 /*
16656                  * It has been sacked, lets move to the
16657                  * next one if possible.
16658                  */
16659                 rack->r_ctl.last_collapse_point = rsm->r_end;
16660                 /* Are we done? */
16661                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16662                             rack->r_ctl.high_collapse_point)) {
16663                         rack->r_collapse_point_valid = 0;
16664                         return (NULL);
16665                 }
16666                 goto restart;
16667         }
16668         /* Now has it been long enough ? */
16669         thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
16670         if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
16671                 rack_log_collapse(rack, rsm->r_start,
16672                                   (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16673                                   thresh, __LINE__, 6, rsm->r_flags, rsm);
16674                 return (rsm);
16675         }
16676         /* Not enough time */
16677         rack_log_collapse(rack, rsm->r_start,
16678                           (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16679                           thresh, __LINE__, 7, rsm->r_flags, rsm);
16680         return (NULL);
16681 }
16682
16683 static int
16684 rack_output(struct tcpcb *tp)
16685 {
16686         struct socket *so;
16687         uint32_t recwin;
16688         uint32_t sb_offset, s_moff = 0;
16689         int32_t len, error = 0;
16690         uint16_t flags;
16691         struct mbuf *m, *s_mb = NULL;
16692         struct mbuf *mb;
16693         uint32_t if_hw_tsomaxsegcount = 0;
16694         uint32_t if_hw_tsomaxsegsize;
16695         int32_t segsiz, minseg;
16696         long tot_len_this_send = 0;
16697 #ifdef INET
16698         struct ip *ip = NULL;
16699 #endif
16700         struct udphdr *udp = NULL;
16701         struct tcp_rack *rack;
16702         struct tcphdr *th;
16703         uint8_t pass = 0;
16704         uint8_t mark = 0;
16705         uint8_t wanted_cookie = 0;
16706         u_char opt[TCP_MAXOLEN];
16707         unsigned ipoptlen, optlen, hdrlen, ulen=0;
16708         uint32_t rack_seq;
16709
16710 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16711         unsigned ipsec_optlen = 0;
16712
16713 #endif
16714         int32_t idle, sendalot;
16715         int32_t sub_from_prr = 0;
16716         volatile int32_t sack_rxmit;
16717         struct rack_sendmap *rsm = NULL;
16718         int32_t tso, mtu;
16719         struct tcpopt to;
16720         int32_t slot = 0;
16721         int32_t sup_rack = 0;
16722         uint32_t cts, ms_cts, delayed, early;
16723         uint16_t add_flag = RACK_SENT_SP;
16724         /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16725         uint8_t hpts_calling,  doing_tlp = 0;
16726         uint32_t cwnd_to_use, pace_max_seg;
16727         int32_t do_a_prefetch = 0;
16728         int32_t prefetch_rsm = 0;
16729         int32_t orig_len = 0;
16730         struct timeval tv;
16731         int32_t prefetch_so_done = 0;
16732         struct tcp_log_buffer *lgb;
16733         struct inpcb *inp;
16734         struct sockbuf *sb;
16735         uint64_t ts_val = 0;
16736 #ifdef TCP_ACCOUNTING
16737         uint64_t crtsc;
16738 #endif
16739 #ifdef INET6
16740         struct ip6_hdr *ip6 = NULL;
16741         int32_t isipv6;
16742 #endif
16743         bool hw_tls = false;
16744
16745         /* setup and take the cache hits here */
16746         rack = (struct tcp_rack *)tp->t_fb_ptr;
16747 #ifdef TCP_ACCOUNTING
16748         sched_pin();
16749         ts_val = get_cyclecount();
16750 #endif
16751         hpts_calling = rack->rc_inp->inp_hpts_calls;
16752         NET_EPOCH_ASSERT();
16753         INP_WLOCK_ASSERT(rack->rc_inp);
16754 #ifdef TCP_OFFLOAD
16755         if (tp->t_flags & TF_TOE) {
16756 #ifdef TCP_ACCOUNTING
16757                 sched_unpin();
16758 #endif
16759                 return (tcp_offload_output(tp));
16760         }
16761 #endif
16762         /*
16763          * For TFO connections in SYN_RECEIVED, only allow the initial
16764          * SYN|ACK and those sent by the retransmit timer.
16765          */
16766         if (IS_FASTOPEN(tp->t_flags) &&
16767             (tp->t_state == TCPS_SYN_RECEIVED) &&
16768             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16769             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16770 #ifdef TCP_ACCOUNTING
16771                 sched_unpin();
16772 #endif
16773                 return (0);
16774         }
16775 #ifdef INET6
16776         if (rack->r_state) {
16777                 /* Use the cache line loaded if possible */
16778                 isipv6 = rack->r_is_v6;
16779         } else {
16780                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16781         }
16782 #endif
16783         early = 0;
16784         cts = tcp_get_usecs(&tv);
16785         ms_cts = tcp_tv_to_mssectick(&tv);
16786         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16787             tcp_in_hpts(rack->rc_inp)) {
16788                 /*
16789                  * We are on the hpts for some timer but not hptsi output.
16790                  * Remove from the hpts unconditionally.
16791                  */
16792                 rack_timer_cancel(tp, rack, cts, __LINE__);
16793         }
16794         /* Are we pacing and late? */
16795         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16796             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16797                 /* We are delayed */
16798                 delayed = cts - rack->r_ctl.rc_last_output_to;
16799         } else {
16800                 delayed = 0;
16801         }
16802         /* Do the timers, which may override the pacer */
16803         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16804                 int retval;
16805
16806                 retval = rack_process_timers(tp, rack, cts, hpts_calling,
16807                     &doing_tlp);
16808                 if (retval != 0) {
16809                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16810 #ifdef TCP_ACCOUNTING
16811                         sched_unpin();
16812 #endif
16813                         /*
16814                          * If timers want tcp_drop(), then pass error out,
16815                          * otherwise suppress it.
16816                          */
16817                         return (retval < 0 ? retval : 0);
16818                 }
16819         }
16820         if (rack->rc_in_persist) {
16821                 if (tcp_in_hpts(rack->rc_inp) == 0) {
16822                         /* Timer is not running */
16823                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16824                 }
16825 #ifdef TCP_ACCOUNTING
16826                 sched_unpin();
16827 #endif
16828                 return (0);
16829         }
16830         if ((rack->rc_ack_required == 1) &&
16831             (rack->r_timer_override == 0)){
16832                 /* A timeout occurred and no ack has arrived */
16833                 if (tcp_in_hpts(rack->rc_inp) == 0) {
16834                         /* Timer is not running */
16835                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16836                 }
16837 #ifdef TCP_ACCOUNTING
16838                 sched_unpin();
16839 #endif
16840                 return (0);
16841         }
16842         if ((rack->r_timer_override) ||
16843             (rack->rc_ack_can_sendout_data) ||
16844             (delayed) ||
16845             (tp->t_state < TCPS_ESTABLISHED)) {
16846                 rack->rc_ack_can_sendout_data = 0;
16847                 if (tcp_in_hpts(rack->rc_inp))
16848                         tcp_hpts_remove(rack->rc_inp);
16849         } else if (tcp_in_hpts(rack->rc_inp)) {
16850                 /*
16851                  * On the hpts you can't pass even if ACKNOW is on, we will
16852                  * when the hpts fires.
16853                  */
16854 #ifdef TCP_ACCOUNTING
16855                 crtsc = get_cyclecount();
16856                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16857                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16858                 }
16859                 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16860                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16861                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16862                 }
16863                 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16864                 sched_unpin();
16865 #endif
16866                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16867                 return (0);
16868         }
16869         rack->rc_inp->inp_hpts_calls = 0;
16870         /* Finish out both pacing early and late accounting */
16871         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16872             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16873                 early = rack->r_ctl.rc_last_output_to - cts;
16874         } else
16875                 early = 0;
16876         if (delayed) {
16877                 rack->r_ctl.rc_agg_delayed += delayed;
16878                 rack->r_late = 1;
16879         } else if (early) {
16880                 rack->r_ctl.rc_agg_early += early;
16881                 rack->r_early = 1;
16882         }
16883         /* Now that early/late accounting is done turn off the flag */
16884         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16885         rack->r_wanted_output = 0;
16886         rack->r_timer_override = 0;
16887         if ((tp->t_state != rack->r_state) &&
16888             TCPS_HAVEESTABLISHED(tp->t_state)) {
16889                 rack_set_state(tp, rack);
16890         }
16891         if ((rack->r_fast_output) &&
16892             (doing_tlp == 0) &&
16893             (tp->rcv_numsacks == 0)) {
16894                 int ret;
16895
16896                 error = 0;
16897                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16898                 if (ret >= 0)
16899                         return(ret);
16900                 else if (error) {
16901                         inp = rack->rc_inp;
16902                         so = inp->inp_socket;
16903                         sb = &so->so_snd;
16904                         goto nomore;
16905                 }
16906         }
16907         inp = rack->rc_inp;
16908         /*
16909          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16910          * only allow the initial SYN or SYN|ACK and those sent
16911          * by the retransmit timer.
16912          */
16913         if (IS_FASTOPEN(tp->t_flags) &&
16914             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16915              (tp->t_state == TCPS_SYN_SENT)) &&
16916             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16917             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16918                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16919                 so = inp->inp_socket;
16920                 sb = &so->so_snd;
16921                 goto just_return_nolock;
16922         }
16923         /*
16924          * Determine length of data that should be transmitted, and flags
16925          * that will be used. If there is some data or critical controls
16926          * (SYN, RST) to send, then transmit; otherwise, investigate
16927          * further.
16928          */
16929         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16930         if (tp->t_idle_reduce) {
16931                 if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16932                         rack_cc_after_idle(rack, tp);
16933         }
16934         tp->t_flags &= ~TF_LASTIDLE;
16935         if (idle) {
16936                 if (tp->t_flags & TF_MORETOCOME) {
16937                         tp->t_flags |= TF_LASTIDLE;
16938                         idle = 0;
16939                 }
16940         }
16941         if ((tp->snd_una == tp->snd_max) &&
16942             rack->r_ctl.rc_went_idle_time &&
16943             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16944                 idle = cts - rack->r_ctl.rc_went_idle_time;
16945                 if (idle > rack_min_probertt_hold) {
16946                         /* Count as a probe rtt */
16947                         if (rack->in_probe_rtt == 0) {
16948                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16949                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16950                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16951                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16952                         } else {
16953                                 rack_exit_probertt(rack, cts);
16954                         }
16955                 }
16956                 idle = 0;
16957         }
16958         if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16959                 rack_init_fsb_block(tp, rack);
16960 again:
16961         /*
16962          * If we've recently taken a timeout, snd_max will be greater than
16963          * snd_nxt.  There may be SACK information that allows us to avoid
16964          * resending already delivered data.  Adjust snd_nxt accordingly.
16965          */
16966         sendalot = 0;
16967         cts = tcp_get_usecs(&tv);
16968         ms_cts = tcp_tv_to_mssectick(&tv);
16969         tso = 0;
16970         mtu = 0;
16971         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16972         minseg = segsiz;
16973         if (rack->r_ctl.rc_pace_max_segs == 0)
16974                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16975         else
16976                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16977         sb_offset = tp->snd_max - tp->snd_una;
16978         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16979         flags = tcp_outflags[tp->t_state];
16980         while (rack->rc_free_cnt < rack_free_cache) {
16981                 rsm = rack_alloc(rack);
16982                 if (rsm == NULL) {
16983                         if (inp->inp_hpts_calls)
16984                                 /* Retry in a ms */
16985                                 slot = (1 * HPTS_USEC_IN_MSEC);
16986                         so = inp->inp_socket;
16987                         sb = &so->so_snd;
16988                         goto just_return_nolock;
16989                 }
16990                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16991                 rack->rc_free_cnt++;
16992                 rsm = NULL;
16993         }
16994         if (inp->inp_hpts_calls)
16995                 inp->inp_hpts_calls = 0;
16996         sack_rxmit = 0;
16997         len = 0;
16998         rsm = NULL;
16999         if (flags & TH_RST) {
17000                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
17001                 so = inp->inp_socket;
17002                 sb = &so->so_snd;
17003                 goto send;
17004         }
17005         if (rack->r_ctl.rc_resend) {
17006                 /* Retransmit timer */
17007                 rsm = rack->r_ctl.rc_resend;
17008                 rack->r_ctl.rc_resend = NULL;
17009                 len = rsm->r_end - rsm->r_start;
17010                 sack_rxmit = 1;
17011                 sendalot = 0;
17012                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17013                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17014                          __func__, __LINE__,
17015                          rsm->r_start, tp->snd_una, tp, rack, rsm));
17016                 sb_offset = rsm->r_start - tp->snd_una;
17017                 if (len >= segsiz)
17018                         len = segsiz;
17019         } else if (rack->r_collapse_point_valid &&
17020                    ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
17021                 /*
17022                  * If an RSM is returned then enough time has passed
17023                  * for us to retransmit it. Move up the collapse point,
17024                  * since this rsm has its chance to retransmit now.
17025                  */
17026                 rack_trace_point(rack, RACK_TP_COLLAPSED_RXT);
17027                 rack->r_ctl.last_collapse_point = rsm->r_end;
17028                 /* Are we done? */
17029                 if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
17030                             rack->r_ctl.high_collapse_point))
17031                         rack->r_collapse_point_valid = 0;
17032                 sack_rxmit = 1;
17033                 /* We are not doing a TLP */
17034                 doing_tlp = 0;
17035                 len = rsm->r_end - rsm->r_start;
17036                 sb_offset = rsm->r_start - tp->snd_una;
17037                 sendalot = 0;
17038                 if ((rack->full_size_rxt == 0) &&
17039                     (rack->shape_rxt_to_pacing_min == 0) &&
17040                     (len >= segsiz))
17041                         len = segsiz;
17042         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
17043                 /* We have a retransmit that takes precedence */
17044                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
17045                     ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
17046                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
17047                         /* Enter recovery if not induced by a time-out */
17048                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
17049                 }
17050 #ifdef INVARIANTS
17051                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
17052                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
17053                               tp, rack, rsm, rsm->r_start, tp->snd_una);
17054                 }
17055 #endif
17056                 len = rsm->r_end - rsm->r_start;
17057                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17058                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17059                          __func__, __LINE__,
17060                          rsm->r_start, tp->snd_una, tp, rack, rsm));
17061                 sb_offset = rsm->r_start - tp->snd_una;
17062                 sendalot = 0;
17063                 if (len >= segsiz)
17064                         len = segsiz;
17065                 if (len > 0) {
17066                         sack_rxmit = 1;
17067                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
17068                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
17069                             min(len, segsiz));
17070                 }
17071         } else if (rack->r_ctl.rc_tlpsend) {
17072                 /* Tail loss probe */
17073                 long cwin;
17074                 long tlen;
17075
17076                 /*
17077                  * Check if we can do a TLP with a RACK'd packet
17078                  * this can happen if we are not doing the rack
17079                  * cheat and we skipped to a TLP and it
17080                  * went off.
17081                  */
17082                 rsm = rack->r_ctl.rc_tlpsend;
17083                 /* We are doing a TLP make sure the flag is preent */
17084                 rsm->r_flags |= RACK_TLP;
17085                 rack->r_ctl.rc_tlpsend = NULL;
17086                 sack_rxmit = 1;
17087                 tlen = rsm->r_end - rsm->r_start;
17088                 if (tlen > segsiz)
17089                         tlen = segsiz;
17090                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17091                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17092                          __func__, __LINE__,
17093                          rsm->r_start, tp->snd_una, tp, rack, rsm));
17094                 sb_offset = rsm->r_start - tp->snd_una;
17095                 cwin = min(tp->snd_wnd, tlen);
17096                 len = cwin;
17097         }
17098         if (rack->r_must_retran &&
17099             (doing_tlp == 0) &&
17100             (SEQ_GT(tp->snd_max, tp->snd_una)) &&
17101             (rsm == NULL)) {
17102                 /*
17103                  * There are two different ways that we
17104                  * can get into this block:
17105                  * a) This is a non-sack connection, we had a time-out
17106                  *    and thus r_must_retran was set and everything
17107                  *    left outstanding as been marked for retransmit.
17108                  * b) The MTU of the path shrank, so that everything
17109                  *    was marked to be retransmitted with the smaller
17110                  *    mtu and r_must_retran was set.
17111                  *
17112                  * This means that we expect the sendmap (outstanding)
17113                  * to all be marked must. We can use the tmap to
17114                  * look at them.
17115                  *
17116                  */
17117                 int sendwin, flight;
17118
17119                 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
17120                 flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
17121                 if (flight >= sendwin) {
17122                         /*
17123                          * We can't send yet.
17124                          */
17125                         so = inp->inp_socket;
17126                         sb = &so->so_snd;
17127                         goto just_return_nolock;
17128                 }
17129                 /*
17130                  * This is the case a/b mentioned above. All
17131                  * outstanding/not-acked should be marked.
17132                  * We can use the tmap to find them.
17133                  */
17134                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17135                 if (rsm == NULL) {
17136                         /* TSNH */
17137                         rack->r_must_retran = 0;
17138                         rack->r_ctl.rc_out_at_rto = 0;
17139                         so = inp->inp_socket;
17140                         sb = &so->so_snd;
17141                         goto just_return_nolock;
17142                 }
17143                 if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
17144                         /*
17145                          * The first one does not have the flag, did we collapse
17146                          * further up in our list?
17147                          */
17148                         rack->r_must_retran = 0;
17149                         rack->r_ctl.rc_out_at_rto = 0;
17150                         rsm = NULL;
17151                         sack_rxmit = 0;
17152                 } else {
17153                         sack_rxmit = 1;
17154                         len = rsm->r_end - rsm->r_start;
17155                         sb_offset = rsm->r_start - tp->snd_una;
17156                         sendalot = 0;
17157                         if ((rack->full_size_rxt == 0) &&
17158                             (rack->shape_rxt_to_pacing_min == 0) &&
17159                             (len >= segsiz))
17160                                 len = segsiz;
17161                         /*
17162                          * Delay removing the flag RACK_MUST_RXT so
17163                          * that the fastpath for retransmit will
17164                          * work with this rsm.
17165                          */
17166                 }
17167         }
17168         /*
17169          * Enforce a connection sendmap count limit if set
17170          * as long as we are not retransmiting.
17171          */
17172         if ((rsm == NULL) &&
17173             (rack->do_detection == 0) &&
17174             (V_tcp_map_entries_limit > 0) &&
17175             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
17176                 counter_u64_add(rack_to_alloc_limited, 1);
17177                 if (!rack->alloc_limit_reported) {
17178                         rack->alloc_limit_reported = 1;
17179                         counter_u64_add(rack_alloc_limited_conns, 1);
17180                 }
17181                 so = inp->inp_socket;
17182                 sb = &so->so_snd;
17183                 goto just_return_nolock;
17184         }
17185         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
17186                 /* we are retransmitting the fin */
17187                 len--;
17188                 if (len) {
17189                         /*
17190                          * When retransmitting data do *not* include the
17191                          * FIN. This could happen from a TLP probe.
17192                          */
17193                         flags &= ~TH_FIN;
17194                 }
17195         }
17196         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
17197             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
17198                 int ret;
17199
17200                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
17201                 if (ret == 0)
17202                         return (0);
17203         }
17204         so = inp->inp_socket;
17205         sb = &so->so_snd;
17206         if (do_a_prefetch == 0) {
17207                 kern_prefetch(sb, &do_a_prefetch);
17208                 do_a_prefetch = 1;
17209         }
17210 #ifdef NETFLIX_SHARED_CWND
17211         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
17212             rack->rack_enable_scwnd) {
17213                 /* We are doing cwnd sharing */
17214                 if (rack->gp_ready &&
17215                     (rack->rack_attempted_scwnd == 0) &&
17216                     (rack->r_ctl.rc_scw == NULL) &&
17217                     tp->t_lib) {
17218                         /* The pcbid is in, lets make an attempt */
17219                         counter_u64_add(rack_try_scwnd, 1);
17220                         rack->rack_attempted_scwnd = 1;
17221                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
17222                                                                    &rack->r_ctl.rc_scw_index,
17223                                                                    segsiz);
17224                 }
17225                 if (rack->r_ctl.rc_scw &&
17226                     (rack->rack_scwnd_is_idle == 1) &&
17227                     sbavail(&so->so_snd)) {
17228                         /* we are no longer out of data */
17229                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17230                         rack->rack_scwnd_is_idle = 0;
17231                 }
17232                 if (rack->r_ctl.rc_scw) {
17233                         /* First lets update and get the cwnd */
17234                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
17235                                                                     rack->r_ctl.rc_scw_index,
17236                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
17237                 }
17238         }
17239 #endif
17240         /*
17241          * Get standard flags, and add SYN or FIN if requested by 'hidden'
17242          * state flags.
17243          */
17244         if (tp->t_flags & TF_NEEDFIN)
17245                 flags |= TH_FIN;
17246         if (tp->t_flags & TF_NEEDSYN)
17247                 flags |= TH_SYN;
17248         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17249                 void *end_rsm;
17250                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17251                 if (end_rsm)
17252                         kern_prefetch(end_rsm, &prefetch_rsm);
17253                 prefetch_rsm = 1;
17254         }
17255         SOCKBUF_LOCK(sb);
17256         /*
17257          * If snd_nxt == snd_max and we have transmitted a FIN, the
17258          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17259          * negative length.  This can also occur when TCP opens up its
17260          * congestion window while receiving additional duplicate acks after
17261          * fast-retransmit because TCP will reset snd_nxt to snd_max after
17262          * the fast-retransmit.
17263          *
17264          * In the normal retransmit-FIN-only case, however, snd_nxt will be
17265          * set to snd_una, the sb_offset will be 0, and the length may wind
17266          * up 0.
17267          *
17268          * If sack_rxmit is true we are retransmitting from the scoreboard
17269          * in which case len is already set.
17270          */
17271         if ((sack_rxmit == 0) &&
17272             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17273                 uint32_t avail;
17274
17275                 avail = sbavail(sb);
17276                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17277                         sb_offset = tp->snd_nxt - tp->snd_una;
17278                 else
17279                         sb_offset = 0;
17280                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17281                         if (rack->r_ctl.rc_tlp_new_data) {
17282                                 /* TLP is forcing out new data */
17283                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17284                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17285                                 }
17286                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17287                                         if (tp->snd_wnd > sb_offset)
17288                                                 len = tp->snd_wnd - sb_offset;
17289                                         else
17290                                                 len = 0;
17291                                 } else {
17292                                         len = rack->r_ctl.rc_tlp_new_data;
17293                                 }
17294                                 rack->r_ctl.rc_tlp_new_data = 0;
17295                         }  else {
17296                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17297                         }
17298                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17299                                 /*
17300                                  * For prr=off, we need to send only 1 MSS
17301                                  * at a time. We do this because another sack could
17302                                  * be arriving that causes us to send retransmits and
17303                                  * we don't want to be on a long pace due to a larger send
17304                                  * that keeps us from sending out the retransmit.
17305                                  */
17306                                 len = segsiz;
17307                         }
17308                 } else {
17309                         uint32_t outstanding;
17310                         /*
17311                          * We are inside of a Fast recovery episode, this
17312                          * is caused by a SACK or 3 dup acks. At this point
17313                          * we have sent all the retransmissions and we rely
17314                          * on PRR to dictate what we will send in the form of
17315                          * new data.
17316                          */
17317
17318                         outstanding = tp->snd_max - tp->snd_una;
17319                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17320                                 if (tp->snd_wnd > outstanding) {
17321                                         len = tp->snd_wnd - outstanding;
17322                                         /* Check to see if we have the data */
17323                                         if ((sb_offset + len) > avail) {
17324                                                 /* It does not all fit */
17325                                                 if (avail > sb_offset)
17326                                                         len = avail - sb_offset;
17327                                                 else
17328                                                         len = 0;
17329                                         }
17330                                 } else {
17331                                         len = 0;
17332                                 }
17333                         } else if (avail > sb_offset) {
17334                                 len = avail - sb_offset;
17335                         } else {
17336                                 len = 0;
17337                         }
17338                         if (len > 0) {
17339                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
17340                                         len = rack->r_ctl.rc_prr_sndcnt;
17341                                 }
17342                                 if (len > 0) {
17343                                         sub_from_prr = 1;
17344                                 }
17345                         }
17346                         if (len > segsiz) {
17347                                 /*
17348                                  * We should never send more than a MSS when
17349                                  * retransmitting or sending new data in prr
17350                                  * mode unless the override flag is on. Most
17351                                  * likely the PRR algorithm is not going to
17352                                  * let us send a lot as well :-)
17353                                  */
17354                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
17355                                         len = segsiz;
17356                                 }
17357                         } else if (len < segsiz) {
17358                                 /*
17359                                  * Do we send any? The idea here is if the
17360                                  * send empty's the socket buffer we want to
17361                                  * do it. However if not then lets just wait
17362                                  * for our prr_sndcnt to get bigger.
17363                                  */
17364                                 long leftinsb;
17365
17366                                 leftinsb = sbavail(sb) - sb_offset;
17367                                 if (leftinsb > len) {
17368                                         /* This send does not empty the sb */
17369                                         len = 0;
17370                                 }
17371                         }
17372                 }
17373         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17374                 /*
17375                  * If you have not established
17376                  * and are not doing FAST OPEN
17377                  * no data please.
17378                  */
17379                 if ((sack_rxmit == 0) &&
17380                     (!IS_FASTOPEN(tp->t_flags))){
17381                         len = 0;
17382                         sb_offset = 0;
17383                 }
17384         }
17385         if (prefetch_so_done == 0) {
17386                 kern_prefetch(so, &prefetch_so_done);
17387                 prefetch_so_done = 1;
17388         }
17389         /*
17390          * Lop off SYN bit if it has already been sent.  However, if this is
17391          * SYN-SENT state and if segment contains data and if we don't know
17392          * that foreign host supports TAO, suppress sending segment.
17393          */
17394         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17395             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17396                 /*
17397                  * When sending additional segments following a TFO SYN|ACK,
17398                  * do not include the SYN bit.
17399                  */
17400                 if (IS_FASTOPEN(tp->t_flags) &&
17401                     (tp->t_state == TCPS_SYN_RECEIVED))
17402                         flags &= ~TH_SYN;
17403         }
17404         /*
17405          * Be careful not to send data and/or FIN on SYN segments. This
17406          * measure is needed to prevent interoperability problems with not
17407          * fully conformant TCP implementations.
17408          */
17409         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17410                 len = 0;
17411                 flags &= ~TH_FIN;
17412         }
17413         /*
17414          * On TFO sockets, ensure no data is sent in the following cases:
17415          *
17416          *  - When retransmitting SYN|ACK on a passively-created socket
17417          *
17418          *  - When retransmitting SYN on an actively created socket
17419          *
17420          *  - When sending a zero-length cookie (cookie request) on an
17421          *    actively created socket
17422          *
17423          *  - When the socket is in the CLOSED state (RST is being sent)
17424          */
17425         if (IS_FASTOPEN(tp->t_flags) &&
17426             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17427              ((tp->t_state == TCPS_SYN_SENT) &&
17428               (tp->t_tfo_client_cookie_len == 0)) ||
17429              (flags & TH_RST))) {
17430                 sack_rxmit = 0;
17431                 len = 0;
17432         }
17433         /* Without fast-open there should never be data sent on a SYN */
17434         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17435                 tp->snd_nxt = tp->iss;
17436                 len = 0;
17437         }
17438         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17439                 /* We only send 1 MSS if we have a DSACK block */
17440                 add_flag |= RACK_SENT_W_DSACK;
17441                 len = segsiz;
17442         }
17443         orig_len = len;
17444         if (len <= 0) {
17445                 /*
17446                  * If FIN has been sent but not acked, but we haven't been
17447                  * called to retransmit, len will be < 0.  Otherwise, window
17448                  * shrank after we sent into it.  If window shrank to 0,
17449                  * cancel pending retransmit, pull snd_nxt back to (closed)
17450                  * window, and set the persist timer if it isn't already
17451                  * going.  If the window didn't close completely, just wait
17452                  * for an ACK.
17453                  *
17454                  * We also do a general check here to ensure that we will
17455                  * set the persist timer when we have data to send, but a
17456                  * 0-byte window. This makes sure the persist timer is set
17457                  * even if the packet hits one of the "goto send" lines
17458                  * below.
17459                  */
17460                 len = 0;
17461                 if ((tp->snd_wnd == 0) &&
17462                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17463                     (tp->snd_una == tp->snd_max) &&
17464                     (sb_offset < (int)sbavail(sb))) {
17465                         rack_enter_persist(tp, rack, cts);
17466                 }
17467         } else if ((rsm == NULL) &&
17468                    (doing_tlp == 0) &&
17469                    (len < pace_max_seg)) {
17470                 /*
17471                  * We are not sending a maximum sized segment for
17472                  * some reason. Should we not send anything (think
17473                  * sws or persists)?
17474                  */
17475                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17476                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17477                     (len < minseg) &&
17478                     (len < (int)(sbavail(sb) - sb_offset))) {
17479                         /*
17480                          * Here the rwnd is less than
17481                          * the minimum pacing size, this is not a retransmit,
17482                          * we are established and
17483                          * the send is not the last in the socket buffer
17484                          * we send nothing, and we may enter persists
17485                          * if nothing is outstanding.
17486                          */
17487                         len = 0;
17488                         if (tp->snd_max == tp->snd_una) {
17489                                 /*
17490                                  * Nothing out we can
17491                                  * go into persists.
17492                                  */
17493                                 rack_enter_persist(tp, rack, cts);
17494                         }
17495                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17496                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17497                            (len < (int)(sbavail(sb) - sb_offset)) &&
17498                            (len < minseg)) {
17499                         /*
17500                          * Here we are not retransmitting, and
17501                          * the cwnd is not so small that we could
17502                          * not send at least a min size (rxt timer
17503                          * not having gone off), We have 2 segments or
17504                          * more already in flight, its not the tail end
17505                          * of the socket buffer  and the cwnd is blocking
17506                          * us from sending out a minimum pacing segment size.
17507                          * Lets not send anything.
17508                          */
17509                         len = 0;
17510                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17511                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17512                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17513                            (len < (int)(sbavail(sb) - sb_offset)) &&
17514                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
17515                         /*
17516                          * Here we have a send window but we have
17517                          * filled it up and we can't send another pacing segment.
17518                          * We also have in flight more than 2 segments
17519                          * and we are not completing the sb i.e. we allow
17520                          * the last bytes of the sb to go out even if
17521                          * its not a full pacing segment.
17522                          */
17523                         len = 0;
17524                 } else if ((rack->r_ctl.crte != NULL) &&
17525                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17526                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17527                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17528                            (len < (int)(sbavail(sb) - sb_offset))) {
17529                         /*
17530                          * Here we are doing hardware pacing, this is not a TLP,
17531                          * we are not sending a pace max segment size, there is rwnd
17532                          * room to send at least N pace_max_seg, the cwnd is greater
17533                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
17534                          * more segments in flight and its not the tail of the socket buffer.
17535                          *
17536                          * We don't want to send instead we need to get more ack's in to
17537                          * allow us to send a full pacing segment. Normally, if we are pacing
17538                          * about the right speed, we should have finished our pacing
17539                          * send as most of the acks have come back if we are at the
17540                          * right rate. This is a bit fuzzy since return path delay
17541                          * can delay the acks, which is why we want to make sure we
17542                          * have cwnd space to have a bit more than a max pace segments in flight.
17543                          *
17544                          * If we have not gotten our acks back we are pacing at too high a
17545                          * rate delaying will not hurt and will bring our GP estimate down by
17546                          * injecting the delay. If we don't do this we will send
17547                          * 2 MSS out in response to the acks being clocked in which
17548                          * defeats the point of hw-pacing (i.e. to help us get
17549                          * larger TSO's out).
17550                          */
17551                         len = 0;
17552
17553                 }
17554
17555         }
17556         /* len will be >= 0 after this point. */
17557         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17558         rack_sndbuf_autoscale(rack);
17559         /*
17560          * Decide if we can use TCP Segmentation Offloading (if supported by
17561          * hardware).
17562          *
17563          * TSO may only be used if we are in a pure bulk sending state.  The
17564          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17565          * options prevent using TSO.  With TSO the TCP header is the same
17566          * (except for the sequence number) for all generated packets.  This
17567          * makes it impossible to transmit any options which vary per
17568          * generated segment or packet.
17569          *
17570          * IPv4 handling has a clear separation of ip options and ip header
17571          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17572          * the right thing below to provide length of just ip options and thus
17573          * checking for ipoptlen is enough to decide if ip options are present.
17574          */
17575         ipoptlen = 0;
17576 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17577         /*
17578          * Pre-calculate here as we save another lookup into the darknesses
17579          * of IPsec that way and can actually decide if TSO is ok.
17580          */
17581 #ifdef INET6
17582         if (isipv6 && IPSEC_ENABLED(ipv6))
17583                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
17584 #ifdef INET
17585         else
17586 #endif
17587 #endif                          /* INET6 */
17588 #ifdef INET
17589                 if (IPSEC_ENABLED(ipv4))
17590                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
17591 #endif                          /* INET */
17592 #endif
17593
17594 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17595         ipoptlen += ipsec_optlen;
17596 #endif
17597         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17598             (tp->t_port == 0) &&
17599             ((tp->t_flags & TF_SIGNATURE) == 0) &&
17600             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17601             ipoptlen == 0)
17602                 tso = 1;
17603         {
17604                 uint32_t outstanding __unused;
17605
17606                 outstanding = tp->snd_max - tp->snd_una;
17607                 if (tp->t_flags & TF_SENTFIN) {
17608                         /*
17609                          * If we sent a fin, snd_max is 1 higher than
17610                          * snd_una
17611                          */
17612                         outstanding--;
17613                 }
17614                 if (sack_rxmit) {
17615                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17616                                 flags &= ~TH_FIN;
17617                 } else {
17618                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17619                                    sbused(sb)))
17620                                 flags &= ~TH_FIN;
17621                 }
17622         }
17623         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17624             (long)TCP_MAXWIN << tp->rcv_scale);
17625
17626         /*
17627          * Sender silly window avoidance.   We transmit under the following
17628          * conditions when len is non-zero:
17629          *
17630          * - We have a full segment (or more with TSO) - This is the last
17631          * buffer in a write()/send() and we are either idle or running
17632          * NODELAY - we've timed out (e.g. persist timer) - we have more
17633          * then 1/2 the maximum send window's worth of data (receiver may be
17634          * limited the window size) - we need to retransmit
17635          */
17636         if (len) {
17637                 if (len >= segsiz) {
17638                         goto send;
17639                 }
17640                 /*
17641                  * NOTE! on localhost connections an 'ack' from the remote
17642                  * end may occur synchronously with the output and cause us
17643                  * to flush a buffer queued with moretocome.  XXX
17644                  *
17645                  */
17646                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
17647                     (idle || (tp->t_flags & TF_NODELAY)) &&
17648                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17649                     (tp->t_flags & TF_NOPUSH) == 0) {
17650                         pass = 2;
17651                         goto send;
17652                 }
17653                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
17654                         pass = 22;
17655                         goto send;
17656                 }
17657                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17658                         pass = 4;
17659                         goto send;
17660                 }
17661                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
17662                         pass = 5;
17663                         goto send;
17664                 }
17665                 if (sack_rxmit) {
17666                         pass = 6;
17667                         goto send;
17668                 }
17669                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17670                     (ctf_outstanding(tp) < (segsiz * 2))) {
17671                         /*
17672                          * We have less than two MSS outstanding (delayed ack)
17673                          * and our rwnd will not let us send a full sized
17674                          * MSS. Lets go ahead and let this small segment
17675                          * out because we want to try to have at least two
17676                          * packets inflight to not be caught by delayed ack.
17677                          */
17678                         pass = 12;
17679                         goto send;
17680                 }
17681         }
17682         /*
17683          * Sending of standalone window updates.
17684          *
17685          * Window updates are important when we close our window due to a
17686          * full socket buffer and are opening it again after the application
17687          * reads data from it.  Once the window has opened again and the
17688          * remote end starts to send again the ACK clock takes over and
17689          * provides the most current window information.
17690          *
17691          * We must avoid the silly window syndrome whereas every read from
17692          * the receive buffer, no matter how small, causes a window update
17693          * to be sent.  We also should avoid sending a flurry of window
17694          * updates when the socket buffer had queued a lot of data and the
17695          * application is doing small reads.
17696          *
17697          * Prevent a flurry of pointless window updates by only sending an
17698          * update when we can increase the advertized window by more than
17699          * 1/4th of the socket buffer capacity.  When the buffer is getting
17700          * full or is very small be more aggressive and send an update
17701          * whenever we can increase by two mss sized segments. In all other
17702          * situations the ACK's to new incoming data will carry further
17703          * window increases.
17704          *
17705          * Don't send an independent window update if a delayed ACK is
17706          * pending (it will get piggy-backed on it) or the remote side
17707          * already has done a half-close and won't send more data.  Skip
17708          * this if the connection is in T/TCP half-open state.
17709          */
17710         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17711             !(tp->t_flags & TF_DELACK) &&
17712             !TCPS_HAVERCVDFIN(tp->t_state)) {
17713                 /*
17714                  * "adv" is the amount we could increase the window, taking
17715                  * into account that we are limited by TCP_MAXWIN <<
17716                  * tp->rcv_scale.
17717                  */
17718                 int32_t adv;
17719                 int oldwin;
17720
17721                 adv = recwin;
17722                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17723                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
17724                         if (adv > oldwin)
17725                             adv -= oldwin;
17726                         else {
17727                                 /* We can't increase the window */
17728                                 adv = 0;
17729                         }
17730                 } else
17731                         oldwin = 0;
17732
17733                 /*
17734                  * If the new window size ends up being the same as or less
17735                  * than the old size when it is scaled, then don't force
17736                  * a window update.
17737                  */
17738                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17739                         goto dontupdate;
17740
17741                 if (adv >= (int32_t)(2 * segsiz) &&
17742                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17743                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17744                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17745                         pass = 7;
17746                         goto send;
17747                 }
17748                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17749                         pass = 23;
17750                         goto send;
17751                 }
17752         }
17753 dontupdate:
17754
17755         /*
17756          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17757          * is also a catch-all for the retransmit timer timeout case.
17758          */
17759         if (tp->t_flags & TF_ACKNOW) {
17760                 pass = 8;
17761                 goto send;
17762         }
17763         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17764                 pass = 9;
17765                 goto send;
17766         }
17767         /*
17768          * If our state indicates that FIN should be sent and we have not
17769          * yet done so, then we need to send.
17770          */
17771         if ((flags & TH_FIN) &&
17772             (tp->snd_nxt == tp->snd_una)) {
17773                 pass = 11;
17774                 goto send;
17775         }
17776         /*
17777          * No reason to send a segment, just return.
17778          */
17779 just_return:
17780         SOCKBUF_UNLOCK(sb);
17781 just_return_nolock:
17782         {
17783                 int app_limited = CTF_JR_SENT_DATA;
17784
17785                 if (tot_len_this_send > 0) {
17786                         /* Make sure snd_nxt is up to max */
17787                         rack->r_ctl.fsb.recwin = recwin;
17788                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17789                         if ((error == 0) &&
17790                             rack_use_rfo &&
17791                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
17792                             (ipoptlen == 0) &&
17793                             (tp->snd_nxt == tp->snd_max) &&
17794                             (tp->rcv_numsacks == 0) &&
17795                             rack->r_fsb_inited &&
17796                             TCPS_HAVEESTABLISHED(tp->t_state) &&
17797                             (rack->r_must_retran == 0) &&
17798                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
17799                             (len > 0) && (orig_len > 0) &&
17800                             (orig_len > len) &&
17801                             ((orig_len - len) >= segsiz) &&
17802                             ((optlen == 0) ||
17803                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17804                                 /* We can send at least one more MSS using our fsb */
17805
17806                                 rack->r_fast_output = 1;
17807                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17808                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17809                                 rack->r_ctl.fsb.tcp_flags = flags;
17810                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
17811                                 if (hw_tls)
17812                                         rack->r_ctl.fsb.hw_tls = 1;
17813                                 else
17814                                         rack->r_ctl.fsb.hw_tls = 0;
17815                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17816                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
17817                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17818                                          (tp->snd_max - tp->snd_una)));
17819                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
17820                                         rack->r_fast_output = 0;
17821                                 else {
17822                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17823                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
17824                                         else
17825                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
17826                                 }
17827                         } else
17828                                 rack->r_fast_output = 0;
17829
17830
17831                         rack_log_fsb(rack, tp, so, flags,
17832                                      ipoptlen, orig_len, len, 0,
17833                                      1, optlen, __LINE__, 1);
17834                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17835                                 tp->snd_nxt = tp->snd_max;
17836                 } else {
17837                         int end_window = 0;
17838                         uint32_t seq = tp->gput_ack;
17839
17840                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17841                         if (rsm) {
17842                                 /*
17843                                  * Mark the last sent that we just-returned (hinting
17844                                  * that delayed ack may play a role in any rtt measurement).
17845                                  */
17846                                 rsm->r_just_ret = 1;
17847                         }
17848                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17849                         rack->r_ctl.rc_agg_delayed = 0;
17850                         rack->r_early = 0;
17851                         rack->r_late = 0;
17852                         rack->r_ctl.rc_agg_early = 0;
17853                         if ((ctf_outstanding(tp) +
17854                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17855                                  minseg)) >= tp->snd_wnd) {
17856                                 /* We are limited by the rwnd */
17857                                 app_limited = CTF_JR_RWND_LIMITED;
17858                                 if (IN_FASTRECOVERY(tp->t_flags))
17859                                     rack->r_ctl.rc_prr_sndcnt = 0;
17860                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17861                                 /* We are limited by whats available -- app limited */
17862                                 app_limited = CTF_JR_APP_LIMITED;
17863                                 if (IN_FASTRECOVERY(tp->t_flags))
17864                                     rack->r_ctl.rc_prr_sndcnt = 0;
17865                         } else if ((idle == 0) &&
17866                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17867                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17868                                    (len < segsiz)) {
17869                                 /*
17870                                  * No delay is not on and the
17871                                  * user is sending less than 1MSS. This
17872                                  * brings out SWS avoidance so we
17873                                  * don't send. Another app-limited case.
17874                                  */
17875                                 app_limited = CTF_JR_APP_LIMITED;
17876                         } else if (tp->t_flags & TF_NOPUSH) {
17877                                 /*
17878                                  * The user has requested no push of
17879                                  * the last segment and we are
17880                                  * at the last segment. Another app
17881                                  * limited case.
17882                                  */
17883                                 app_limited = CTF_JR_APP_LIMITED;
17884                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17885                                 /* Its the cwnd */
17886                                 app_limited = CTF_JR_CWND_LIMITED;
17887                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17888                                    (rack->rack_no_prr == 0) &&
17889                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17890                                 app_limited = CTF_JR_PRR;
17891                         } else {
17892                                 /* Now why here are we not sending? */
17893 #ifdef NOW
17894 #ifdef INVARIANTS
17895                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17896 #endif
17897 #endif
17898                                 app_limited = CTF_JR_ASSESSING;
17899                         }
17900                         /*
17901                          * App limited in some fashion, for our pacing GP
17902                          * measurements we don't want any gap (even cwnd).
17903                          * Close  down the measurement window.
17904                          */
17905                         if (rack_cwnd_block_ends_measure &&
17906                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17907                              (app_limited == CTF_JR_PRR))) {
17908                                 /*
17909                                  * The reason we are not sending is
17910                                  * the cwnd (or prr). We have been configured
17911                                  * to end the measurement window in
17912                                  * this case.
17913                                  */
17914                                 end_window = 1;
17915                         } else if (rack_rwnd_block_ends_measure &&
17916                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17917                                 /*
17918                                  * We are rwnd limited and have been
17919                                  * configured to end the measurement
17920                                  * window in this case.
17921                                  */
17922                                 end_window = 1;
17923                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17924                                 /*
17925                                  * A true application limited period, we have
17926                                  * ran out of data.
17927                                  */
17928                                 end_window = 1;
17929                         } else if (app_limited == CTF_JR_ASSESSING) {
17930                                 /*
17931                                  * In the assessing case we hit the end of
17932                                  * the if/else and had no known reason
17933                                  * This will panic us under invariants..
17934                                  *
17935                                  * If we get this out in logs we need to
17936                                  * investagate which reason we missed.
17937                                  */
17938                                 end_window = 1;
17939                         }
17940                         if (end_window) {
17941                                 uint8_t log = 0;
17942
17943                                 /* Adjust the Gput measurement */
17944                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17945                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17946                                         tp->gput_ack = tp->snd_max;
17947                                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17948                                                 /*
17949                                                  * There is not enough to measure.
17950                                                  */
17951                                                 tp->t_flags &= ~TF_GPUTINPROG;
17952                                                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17953                                                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
17954                                                                            tp->gput_seq,
17955                                                                            0, 0, 18, __LINE__, NULL, 0);
17956                                         } else
17957                                                 log = 1;
17958                                 }
17959                                 /* Mark the last packet has app limited */
17960                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17961                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17962                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17963                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17964                                         else {
17965                                                 /*
17966                                                  * Go out to the end app limited and mark
17967                                                  * this new one as next and move the end_appl up
17968                                                  * to this guy.
17969                                                  */
17970                                                 if (rack->r_ctl.rc_end_appl)
17971                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17972                                                 rack->r_ctl.rc_end_appl = rsm;
17973                                         }
17974                                         rsm->r_flags |= RACK_APP_LIMITED;
17975                                         rack->r_ctl.rc_app_limited_cnt++;
17976                                 }
17977                                 if (log)
17978                                         rack_log_pacing_delay_calc(rack,
17979                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17980                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17981                         }
17982                 }
17983                 /* Check if we need to go into persists or not */
17984                 if ((tp->snd_max == tp->snd_una) &&
17985                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17986                     sbavail(sb) &&
17987                     (sbavail(sb) > tp->snd_wnd) &&
17988                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17989                         /* Yes lets make sure to move to persist before timer-start */
17990                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17991                 }
17992                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17993                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17994         }
17995 #ifdef NETFLIX_SHARED_CWND
17996         if ((sbavail(sb) == 0) &&
17997             rack->r_ctl.rc_scw) {
17998                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17999                 rack->rack_scwnd_is_idle = 1;
18000         }
18001 #endif
18002 #ifdef TCP_ACCOUNTING
18003         if (tot_len_this_send > 0) {
18004                 crtsc = get_cyclecount();
18005                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18006                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
18007                 }
18008                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
18009                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18010                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
18011                 }
18012                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
18013                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18014                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
18015                 }
18016                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
18017         } else {
18018                 crtsc = get_cyclecount();
18019                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18020                         tp->tcp_cnt_counters[SND_LIMITED]++;
18021                 }
18022                 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
18023                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18024                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
18025                 }
18026                 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
18027         }
18028         sched_unpin();
18029 #endif
18030         return (0);
18031
18032 send:
18033         if (rsm || sack_rxmit)
18034                 counter_u64_add(rack_nfto_resend, 1);
18035         else
18036                 counter_u64_add(rack_non_fto_send, 1);
18037         if ((flags & TH_FIN) &&
18038             sbavail(sb)) {
18039                 /*
18040                  * We do not transmit a FIN
18041                  * with data outstanding. We
18042                  * need to make it so all data
18043                  * is acked first.
18044                  */
18045                 flags &= ~TH_FIN;
18046         }
18047         /* Enforce stack imposed max seg size if we have one */
18048         if (rack->r_ctl.rc_pace_max_segs &&
18049             (len > rack->r_ctl.rc_pace_max_segs)) {
18050                 mark = 1;
18051                 len = rack->r_ctl.rc_pace_max_segs;
18052         }
18053         SOCKBUF_LOCK_ASSERT(sb);
18054         if (len > 0) {
18055                 if (len >= segsiz)
18056                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
18057                 else
18058                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
18059         }
18060         /*
18061          * Before ESTABLISHED, force sending of initial options unless TCP
18062          * set not to do any options. NOTE: we assume that the IP/TCP header
18063          * plus TCP options always fit in a single mbuf, leaving room for a
18064          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
18065          * + optlen <= MCLBYTES
18066          */
18067         optlen = 0;
18068 #ifdef INET6
18069         if (isipv6)
18070                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18071         else
18072 #endif
18073                 hdrlen = sizeof(struct tcpiphdr);
18074
18075         /*
18076          * Compute options for segment. We only have to care about SYN and
18077          * established connection segments.  Options for SYN-ACK segments
18078          * are handled in TCP syncache.
18079          */
18080         to.to_flags = 0;
18081         if ((tp->t_flags & TF_NOOPT) == 0) {
18082                 /* Maximum segment size. */
18083                 if (flags & TH_SYN) {
18084                         tp->snd_nxt = tp->iss;
18085                         to.to_mss = tcp_mssopt(&inp->inp_inc);
18086                         if (tp->t_port)
18087                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
18088                         to.to_flags |= TOF_MSS;
18089
18090                         /*
18091                          * On SYN or SYN|ACK transmits on TFO connections,
18092                          * only include the TFO option if it is not a
18093                          * retransmit, as the presence of the TFO option may
18094                          * have caused the original SYN or SYN|ACK to have
18095                          * been dropped by a middlebox.
18096                          */
18097                         if (IS_FASTOPEN(tp->t_flags) &&
18098                             (tp->t_rxtshift == 0)) {
18099                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
18100                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
18101                                         to.to_tfo_cookie =
18102                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
18103                                         to.to_flags |= TOF_FASTOPEN;
18104                                         wanted_cookie = 1;
18105                                 } else if (tp->t_state == TCPS_SYN_SENT) {
18106                                         to.to_tfo_len =
18107                                                 tp->t_tfo_client_cookie_len;
18108                                         to.to_tfo_cookie =
18109                                                 tp->t_tfo_cookie.client;
18110                                         to.to_flags |= TOF_FASTOPEN;
18111                                         wanted_cookie = 1;
18112                                         /*
18113                                          * If we wind up having more data to
18114                                          * send with the SYN than can fit in
18115                                          * one segment, don't send any more
18116                                          * until the SYN|ACK comes back from
18117                                          * the other end.
18118                                          */
18119                                         sendalot = 0;
18120                                 }
18121                         }
18122                 }
18123                 /* Window scaling. */
18124                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
18125                         to.to_wscale = tp->request_r_scale;
18126                         to.to_flags |= TOF_SCALE;
18127                 }
18128                 /* Timestamps. */
18129                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
18130                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
18131                         to.to_tsval = ms_cts + tp->ts_offset;
18132                         to.to_tsecr = tp->ts_recent;
18133                         to.to_flags |= TOF_TS;
18134                 }
18135                 /* Set receive buffer autosizing timestamp. */
18136                 if (tp->rfbuf_ts == 0 &&
18137                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
18138                         tp->rfbuf_ts = tcp_ts_getticks();
18139                 /* Selective ACK's. */
18140                 if (tp->t_flags & TF_SACK_PERMIT) {
18141                         if (flags & TH_SYN)
18142                                 to.to_flags |= TOF_SACKPERM;
18143                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18144                                  tp->rcv_numsacks > 0) {
18145                                 to.to_flags |= TOF_SACK;
18146                                 to.to_nsacks = tp->rcv_numsacks;
18147                                 to.to_sacks = (u_char *)tp->sackblks;
18148                         }
18149                 }
18150 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18151                 /* TCP-MD5 (RFC2385). */
18152                 if (tp->t_flags & TF_SIGNATURE)
18153                         to.to_flags |= TOF_SIGNATURE;
18154 #endif                          /* TCP_SIGNATURE */
18155
18156                 /* Processing the options. */
18157                 hdrlen += optlen = tcp_addoptions(&to, opt);
18158                 /*
18159                  * If we wanted a TFO option to be added, but it was unable
18160                  * to fit, ensure no data is sent.
18161                  */
18162                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
18163                     !(to.to_flags & TOF_FASTOPEN))
18164                         len = 0;
18165         }
18166         if (tp->t_port) {
18167                 if (V_tcp_udp_tunneling_port == 0) {
18168                         /* The port was removed?? */
18169                         SOCKBUF_UNLOCK(&so->so_snd);
18170 #ifdef TCP_ACCOUNTING
18171                         crtsc = get_cyclecount();
18172                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18173                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18174                         }
18175                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18176                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18177                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18178                         }
18179                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18180                         sched_unpin();
18181 #endif
18182                         return (EHOSTUNREACH);
18183                 }
18184                 hdrlen += sizeof(struct udphdr);
18185         }
18186 #ifdef INET6
18187         if (isipv6)
18188                 ipoptlen = ip6_optlen(tp->t_inpcb);
18189         else
18190 #endif
18191                 if (tp->t_inpcb->inp_options)
18192                         ipoptlen = tp->t_inpcb->inp_options->m_len -
18193                                 offsetof(struct ipoption, ipopt_list);
18194                 else
18195                         ipoptlen = 0;
18196 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18197         ipoptlen += ipsec_optlen;
18198 #endif
18199
18200         /*
18201          * Adjust data length if insertion of options will bump the packet
18202          * length beyond the t_maxseg length. Clear the FIN bit because we
18203          * cut off the tail of the segment.
18204          */
18205         if (len + optlen + ipoptlen > tp->t_maxseg) {
18206                 if (tso) {
18207                         uint32_t if_hw_tsomax;
18208                         uint32_t moff;
18209                         int32_t max_len;
18210
18211                         /* extract TSO information */
18212                         if_hw_tsomax = tp->t_tsomax;
18213                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18214                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18215                         KASSERT(ipoptlen == 0,
18216                                 ("%s: TSO can't do IP options", __func__));
18217
18218                         /*
18219                          * Check if we should limit by maximum payload
18220                          * length:
18221                          */
18222                         if (if_hw_tsomax != 0) {
18223                                 /* compute maximum TSO length */
18224                                 max_len = (if_hw_tsomax - hdrlen -
18225                                            max_linkhdr);
18226                                 if (max_len <= 0) {
18227                                         len = 0;
18228                                 } else if (len > max_len) {
18229                                         sendalot = 1;
18230                                         len = max_len;
18231                                         mark = 2;
18232                                 }
18233                         }
18234                         /*
18235                          * Prevent the last segment from being fractional
18236                          * unless the send sockbuf can be emptied:
18237                          */
18238                         max_len = (tp->t_maxseg - optlen);
18239                         if ((sb_offset + len) < sbavail(sb)) {
18240                                 moff = len % (u_int)max_len;
18241                                 if (moff != 0) {
18242                                         mark = 3;
18243                                         len -= moff;
18244                                 }
18245                         }
18246                         /*
18247                          * In case there are too many small fragments don't
18248                          * use TSO:
18249                          */
18250                         if (len <= segsiz) {
18251                                 mark = 4;
18252                                 tso = 0;
18253                         }
18254                         /*
18255                          * Send the FIN in a separate segment after the bulk
18256                          * sending is done. We don't trust the TSO
18257                          * implementations to clear the FIN flag on all but
18258                          * the last segment.
18259                          */
18260                         if (tp->t_flags & TF_NEEDFIN) {
18261                                 sendalot = 4;
18262                         }
18263                 } else {
18264                         mark = 5;
18265                         if (optlen + ipoptlen >= tp->t_maxseg) {
18266                                 /*
18267                                  * Since we don't have enough space to put
18268                                  * the IP header chain and the TCP header in
18269                                  * one packet as required by RFC 7112, don't
18270                                  * send it. Also ensure that at least one
18271                                  * byte of the payload can be put into the
18272                                  * TCP segment.
18273                                  */
18274                                 SOCKBUF_UNLOCK(&so->so_snd);
18275                                 error = EMSGSIZE;
18276                                 sack_rxmit = 0;
18277                                 goto out;
18278                         }
18279                         len = tp->t_maxseg - optlen - ipoptlen;
18280                         sendalot = 5;
18281                 }
18282         } else {
18283                 tso = 0;
18284                 mark = 6;
18285         }
18286         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18287                 ("%s: len > IP_MAXPACKET", __func__));
18288 #ifdef DIAGNOSTIC
18289 #ifdef INET6
18290         if (max_linkhdr + hdrlen > MCLBYTES)
18291 #else
18292                 if (max_linkhdr + hdrlen > MHLEN)
18293 #endif
18294                         panic("tcphdr too big");
18295 #endif
18296
18297         /*
18298          * This KASSERT is here to catch edge cases at a well defined place.
18299          * Before, those had triggered (random) panic conditions further
18300          * down.
18301          */
18302         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18303         if ((len == 0) &&
18304             (flags & TH_FIN) &&
18305             (sbused(sb))) {
18306                 /*
18307                  * We have outstanding data, don't send a fin by itself!.
18308                  */
18309                 goto just_return;
18310         }
18311         /*
18312          * Grab a header mbuf, attaching a copy of data to be transmitted,
18313          * and initialize the header from the template for sends on this
18314          * connection.
18315          */
18316         hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
18317         if (len) {
18318                 uint32_t max_val;
18319                 uint32_t moff;
18320
18321                 if (rack->r_ctl.rc_pace_max_segs)
18322                         max_val = rack->r_ctl.rc_pace_max_segs;
18323                 else if (rack->rc_user_set_max_segs)
18324                         max_val = rack->rc_user_set_max_segs * segsiz;
18325                 else
18326                         max_val = len;
18327                 /*
18328                  * We allow a limit on sending with hptsi.
18329                  */
18330                 if (len > max_val) {
18331                         mark = 7;
18332                         len = max_val;
18333                 }
18334 #ifdef INET6
18335                 if (MHLEN < hdrlen + max_linkhdr)
18336                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18337                 else
18338 #endif
18339                         m = m_gethdr(M_NOWAIT, MT_DATA);
18340
18341                 if (m == NULL) {
18342                         SOCKBUF_UNLOCK(sb);
18343                         error = ENOBUFS;
18344                         sack_rxmit = 0;
18345                         goto out;
18346                 }
18347                 m->m_data += max_linkhdr;
18348                 m->m_len = hdrlen;
18349
18350                 /*
18351                  * Start the m_copy functions from the closest mbuf to the
18352                  * sb_offset in the socket buffer chain.
18353                  */
18354                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
18355                 s_mb = mb;
18356                 s_moff = moff;
18357                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18358                         m_copydata(mb, moff, (int)len,
18359                                    mtod(m, caddr_t)+hdrlen);
18360                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18361                                 sbsndptr_adv(sb, mb, len);
18362                         m->m_len += len;
18363                 } else {
18364                         struct sockbuf *msb;
18365
18366                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18367                                 msb = NULL;
18368                         else
18369                                 msb = sb;
18370                         m->m_next = tcp_m_copym(
18371                                 mb, moff, &len,
18372                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18373                                 ((rsm == NULL) ? hw_tls : 0)
18374 #ifdef NETFLIX_COPY_ARGS
18375                                 , &s_mb, &s_moff
18376 #endif
18377                                 );
18378                         if (len <= (tp->t_maxseg - optlen)) {
18379                                 /*
18380                                  * Must have ran out of mbufs for the copy
18381                                  * shorten it to no longer need tso. Lets
18382                                  * not put on sendalot since we are low on
18383                                  * mbufs.
18384                                  */
18385                                 tso = 0;
18386                         }
18387                         if (m->m_next == NULL) {
18388                                 SOCKBUF_UNLOCK(sb);
18389                                 (void)m_free(m);
18390                                 error = ENOBUFS;
18391                                 sack_rxmit = 0;
18392                                 goto out;
18393                         }
18394                 }
18395                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18396                         if (rsm && (rsm->r_flags & RACK_TLP)) {
18397                                 /*
18398                                  * TLP should not count in retran count, but
18399                                  * in its own bin
18400                                  */
18401                                 counter_u64_add(rack_tlp_retran, 1);
18402                                 counter_u64_add(rack_tlp_retran_bytes, len);
18403                         } else {
18404                                 tp->t_sndrexmitpack++;
18405                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18406                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18407                         }
18408 #ifdef STATS
18409                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18410                                                  len);
18411 #endif
18412                 } else {
18413                         KMOD_TCPSTAT_INC(tcps_sndpack);
18414                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18415 #ifdef STATS
18416                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18417                                                  len);
18418 #endif
18419                 }
18420                 /*
18421                  * If we're sending everything we've got, set PUSH. (This
18422                  * will keep happy those implementations which only give
18423                  * data to the user when a buffer fills or a PUSH comes in.)
18424                  */
18425                 if (sb_offset + len == sbused(sb) &&
18426                     sbused(sb) &&
18427                     !(flags & TH_SYN)) {
18428                         flags |= TH_PUSH;
18429                         add_flag |= RACK_HAD_PUSH;
18430                 }
18431
18432                 SOCKBUF_UNLOCK(sb);
18433         } else {
18434                 SOCKBUF_UNLOCK(sb);
18435                 if (tp->t_flags & TF_ACKNOW)
18436                         KMOD_TCPSTAT_INC(tcps_sndacks);
18437                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
18438                         KMOD_TCPSTAT_INC(tcps_sndctrl);
18439                 else
18440                         KMOD_TCPSTAT_INC(tcps_sndwinup);
18441
18442                 m = m_gethdr(M_NOWAIT, MT_DATA);
18443                 if (m == NULL) {
18444                         error = ENOBUFS;
18445                         sack_rxmit = 0;
18446                         goto out;
18447                 }
18448 #ifdef INET6
18449                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18450                     MHLEN >= hdrlen) {
18451                         M_ALIGN(m, hdrlen);
18452                 } else
18453 #endif
18454                         m->m_data += max_linkhdr;
18455                 m->m_len = hdrlen;
18456         }
18457         SOCKBUF_UNLOCK_ASSERT(sb);
18458         m->m_pkthdr.rcvif = (struct ifnet *)0;
18459 #ifdef MAC
18460         mac_inpcb_create_mbuf(inp, m);
18461 #endif
18462         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18463 #ifdef INET6
18464                 if (isipv6)
18465                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18466                 else
18467 #endif                          /* INET6 */
18468                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18469                 th = rack->r_ctl.fsb.th;
18470                 udp = rack->r_ctl.fsb.udp;
18471                 if (udp) {
18472 #ifdef INET6
18473                         if (isipv6)
18474                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18475                         else
18476 #endif                          /* INET6 */
18477                                 ulen = hdrlen + len - sizeof(struct ip);
18478                         udp->uh_ulen = htons(ulen);
18479                 }
18480         } else {
18481 #ifdef INET6
18482                 if (isipv6) {
18483                         ip6 = mtod(m, struct ip6_hdr *);
18484                         if (tp->t_port) {
18485                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18486                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18487                                 udp->uh_dport = tp->t_port;
18488                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
18489                                 udp->uh_ulen = htons(ulen);
18490                                 th = (struct tcphdr *)(udp + 1);
18491                         } else
18492                                 th = (struct tcphdr *)(ip6 + 1);
18493                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
18494                 } else
18495 #endif                          /* INET6 */
18496                 {
18497                         ip = mtod(m, struct ip *);
18498                         if (tp->t_port) {
18499                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18500                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18501                                 udp->uh_dport = tp->t_port;
18502                                 ulen = hdrlen + len - sizeof(struct ip);
18503                                 udp->uh_ulen = htons(ulen);
18504                                 th = (struct tcphdr *)(udp + 1);
18505                         } else
18506                                 th = (struct tcphdr *)(ip + 1);
18507                         tcpip_fillheaders(inp, tp->t_port, ip, th);
18508                 }
18509         }
18510         /*
18511          * Fill in fields, remembering maximum advertised window for use in
18512          * delaying messages about window sizes. If resending a FIN, be sure
18513          * not to use a new sequence number.
18514          */
18515         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18516             tp->snd_nxt == tp->snd_max)
18517                 tp->snd_nxt--;
18518         /*
18519          * If we are starting a connection, send ECN setup SYN packet. If we
18520          * are on a retransmit, we may resend those bits a number of times
18521          * as per RFC 3168.
18522          */
18523         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
18524                 flags |= tcp_ecn_output_syn_sent(tp);
18525         }
18526         /* Also handle parallel SYN for ECN */
18527         if (TCPS_HAVERCVDSYN(tp->t_state) &&
18528             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18529                 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
18530                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18531                     (tp->t_flags2 & TF2_ECN_SND_ECE))
18532                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18533 #ifdef INET6
18534                 if (isipv6) {
18535                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18536                         ip6->ip6_flow |= htonl(ect << 20);
18537                 }
18538                 else
18539 #endif
18540                 {
18541                         ip->ip_tos &= ~IPTOS_ECN_MASK;
18542                         ip->ip_tos |= ect;
18543                 }
18544         }
18545         /*
18546          * If we are doing retransmissions, then snd_nxt will not reflect
18547          * the first unsent octet.  For ACK only packets, we do not want the
18548          * sequence number of the retransmitted packet, we want the sequence
18549          * number of the next unsent octet.  So, if there is no data (and no
18550          * SYN or FIN), use snd_max instead of snd_nxt when filling in
18551          * ti_seq.  But if we are in persist state, snd_max might reflect
18552          * one byte beyond the right edge of the window, so use snd_nxt in
18553          * that case, since we know we aren't doing a retransmission.
18554          * (retransmit and persist are mutually exclusive...)
18555          */
18556         if (sack_rxmit == 0) {
18557                 if (len || (flags & (TH_SYN | TH_FIN))) {
18558                         th->th_seq = htonl(tp->snd_nxt);
18559                         rack_seq = tp->snd_nxt;
18560                 } else {
18561                         th->th_seq = htonl(tp->snd_max);
18562                         rack_seq = tp->snd_max;
18563                 }
18564         } else {
18565                 th->th_seq = htonl(rsm->r_start);
18566                 rack_seq = rsm->r_start;
18567         }
18568         th->th_ack = htonl(tp->rcv_nxt);
18569         tcp_set_flags(th, flags);
18570         /*
18571          * Calculate receive window.  Don't shrink window, but avoid silly
18572          * window syndrome.
18573          * If a RST segment is sent, advertise a window of zero.
18574          */
18575         if (flags & TH_RST) {
18576                 recwin = 0;
18577         } else {
18578                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18579                     recwin < (long)segsiz) {
18580                         recwin = 0;
18581                 }
18582                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18583                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18584                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18585         }
18586
18587         /*
18588          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18589          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18590          * handled in syncache.
18591          */
18592         if (flags & TH_SYN)
18593                 th->th_win = htons((u_short)
18594                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18595         else {
18596                 /* Avoid shrinking window with window scaling. */
18597                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
18598                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18599         }
18600         /*
18601          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18602          * window.  This may cause the remote transmitter to stall.  This
18603          * flag tells soreceive() to disable delayed acknowledgements when
18604          * draining the buffer.  This can occur if the receiver is
18605          * attempting to read more data than can be buffered prior to
18606          * transmitting on the connection.
18607          */
18608         if (th->th_win == 0) {
18609                 tp->t_sndzerowin++;
18610                 tp->t_flags |= TF_RXWIN0SENT;
18611         } else
18612                 tp->t_flags &= ~TF_RXWIN0SENT;
18613         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
18614         /* Now are we using fsb?, if so copy the template data to the mbuf */
18615         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18616                 uint8_t *cpto;
18617
18618                 cpto = mtod(m, uint8_t *);
18619                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18620                 /*
18621                  * We have just copied in:
18622                  * IP/IP6
18623                  * <optional udphdr>
18624                  * tcphdr (no options)
18625                  *
18626                  * We need to grab the correct pointers into the mbuf
18627                  * for both the tcp header, and possibly the udp header (if tunneling).
18628                  * We do this by using the offset in the copy buffer and adding it
18629                  * to the mbuf base pointer (cpto).
18630                  */
18631 #ifdef INET6
18632                 if (isipv6)
18633                         ip6 = mtod(m, struct ip6_hdr *);
18634                 else
18635 #endif                          /* INET6 */
18636                         ip = mtod(m, struct ip *);
18637                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18638                 /* If we have a udp header lets set it into the mbuf as well */
18639                 if (udp)
18640                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18641         }
18642 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18643         if (to.to_flags & TOF_SIGNATURE) {
18644                 /*
18645                  * Calculate MD5 signature and put it into the place
18646                  * determined before.
18647                  * NOTE: since TCP options buffer doesn't point into
18648                  * mbuf's data, calculate offset and use it.
18649                  */
18650                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18651                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18652                         /*
18653                          * Do not send segment if the calculation of MD5
18654                          * digest has failed.
18655                          */
18656                         goto out;
18657                 }
18658         }
18659 #endif
18660         if (optlen) {
18661                 bcopy(opt, th + 1, optlen);
18662                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18663         }
18664         /*
18665          * Put TCP length in extended header, and then checksum extended
18666          * header and data.
18667          */
18668         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
18669 #ifdef INET6
18670         if (isipv6) {
18671                 /*
18672                  * ip6_plen is not need to be filled now, and will be filled
18673                  * in ip6_output.
18674                  */
18675                 if (tp->t_port) {
18676                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18677                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18678                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18679                         th->th_sum = htons(0);
18680                         UDPSTAT_INC(udps_opackets);
18681                 } else {
18682                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18683                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18684                         th->th_sum = in6_cksum_pseudo(ip6,
18685                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18686                                                       0);
18687                 }
18688         }
18689 #endif
18690 #if defined(INET6) && defined(INET)
18691         else
18692 #endif
18693 #ifdef INET
18694         {
18695                 if (tp->t_port) {
18696                         m->m_pkthdr.csum_flags = CSUM_UDP;
18697                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18698                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18699                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18700                         th->th_sum = htons(0);
18701                         UDPSTAT_INC(udps_opackets);
18702                 } else {
18703                         m->m_pkthdr.csum_flags = CSUM_TCP;
18704                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18705                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
18706                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18707                                                                         IPPROTO_TCP + len + optlen));
18708                 }
18709                 /* IP version must be set here for ipv4/ipv6 checking later */
18710                 KASSERT(ip->ip_v == IPVERSION,
18711                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
18712         }
18713 #endif
18714         /*
18715          * Enable TSO and specify the size of the segments. The TCP pseudo
18716          * header checksum is always provided. XXX: Fixme: This is currently
18717          * not the case for IPv6.
18718          */
18719         if (tso) {
18720                 KASSERT(len > tp->t_maxseg - optlen,
18721                         ("%s: len <= tso_segsz", __func__));
18722                 m->m_pkthdr.csum_flags |= CSUM_TSO;
18723                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18724         }
18725         KASSERT(len + hdrlen == m_length(m, NULL),
18726                 ("%s: mbuf chain different than expected: %d + %u != %u",
18727                  __func__, len, hdrlen, m_length(m, NULL)));
18728
18729 #ifdef TCP_HHOOK
18730         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18731         hhook_run_tcp_est_out(tp, th, &to, len, tso);
18732 #endif
18733         /* We're getting ready to send; log now. */
18734         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18735                 union tcp_log_stackspecific log;
18736
18737                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18738                 log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18739                 if (rack->rack_no_prr)
18740                         log.u_bbr.flex1 = 0;
18741                 else
18742                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18743                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18744                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18745                 log.u_bbr.flex4 = orig_len;
18746                 /* Save off the early/late values */
18747                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18748                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18749                 log.u_bbr.bw_inuse = rack_get_bw(rack);
18750                 log.u_bbr.flex8 = 0;
18751                 if (rsm) {
18752                         if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18753                                 rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18754                                 counter_u64_add(rack_collapsed_win_rxt, 1);
18755                                 counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18756                         }
18757                         if (doing_tlp)
18758                                 log.u_bbr.flex8 = 2;
18759                         else
18760                                 log.u_bbr.flex8 = 1;
18761                 } else {
18762                         if (doing_tlp)
18763                                 log.u_bbr.flex8 = 3;
18764                         else
18765                                 log.u_bbr.flex8 = 0;
18766                 }
18767                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18768                 log.u_bbr.flex7 = mark;
18769                 log.u_bbr.flex7 <<= 8;
18770                 log.u_bbr.flex7 |= pass;
18771                 log.u_bbr.pkts_out = tp->t_maxseg;
18772                 log.u_bbr.timeStamp = cts;
18773                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18774                 log.u_bbr.lt_epoch = cwnd_to_use;
18775                 log.u_bbr.delivered = sendalot;
18776                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18777                                      len, &log, false, NULL, NULL, 0, &tv);
18778         } else
18779                 lgb = NULL;
18780
18781         /*
18782          * Fill in IP length and desired time to live and send to IP level.
18783          * There should be a better way to handle ttl and tos; we could keep
18784          * them in the template, but need a way to checksum without them.
18785          */
18786         /*
18787          * m->m_pkthdr.len should have been set before cksum calcuration,
18788          * because in6_cksum() need it.
18789          */
18790 #ifdef INET6
18791         if (isipv6) {
18792                 /*
18793                  * we separately set hoplimit for every segment, since the
18794                  * user might want to change the value via setsockopt. Also,
18795                  * desired default hop limit might be changed via Neighbor
18796                  * Discovery.
18797                  */
18798                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18799
18800                 /*
18801                  * Set the packet size here for the benefit of DTrace
18802                  * probes. ip6_output() will set it properly; it's supposed
18803                  * to include the option header lengths as well.
18804                  */
18805                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18806
18807                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18808                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18809                 else
18810                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18811
18812                 if (tp->t_state == TCPS_SYN_SENT)
18813                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18814
18815                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18816                 /* TODO: IPv6 IP6TOS_ECT bit on */
18817                 error = ip6_output(m,
18818 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18819                                    inp->in6p_outputopts,
18820 #else
18821                                    NULL,
18822 #endif
18823                                    &inp->inp_route6,
18824                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18825                                    NULL, NULL, inp);
18826
18827                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18828                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18829         }
18830 #endif                          /* INET6 */
18831 #if defined(INET) && defined(INET6)
18832         else
18833 #endif
18834 #ifdef INET
18835         {
18836                 ip->ip_len = htons(m->m_pkthdr.len);
18837 #ifdef INET6
18838                 if (inp->inp_vflag & INP_IPV6PROTO)
18839                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18840 #endif                          /* INET6 */
18841                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18842                 /*
18843                  * If we do path MTU discovery, then we set DF on every
18844                  * packet. This might not be the best thing to do according
18845                  * to RFC3390 Section 2. However the tcp hostcache migitates
18846                  * the problem so it affects only the first tcp connection
18847                  * with a host.
18848                  *
18849                  * NB: Don't set DF on small MTU/MSS to have a safe
18850                  * fallback.
18851                  */
18852                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18853                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18854                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18855                                 ip->ip_off |= htons(IP_DF);
18856                         }
18857                 } else {
18858                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18859                 }
18860
18861                 if (tp->t_state == TCPS_SYN_SENT)
18862                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18863
18864                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18865
18866                 error = ip_output(m,
18867 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18868                                   inp->inp_options,
18869 #else
18870                                   NULL,
18871 #endif
18872                                   &inp->inp_route,
18873                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18874                                   inp);
18875                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18876                         mtu = inp->inp_route.ro_nh->nh_mtu;
18877         }
18878 #endif                          /* INET */
18879
18880 out:
18881         if (lgb) {
18882                 lgb->tlb_errno = error;
18883                 lgb = NULL;
18884         }
18885         /*
18886          * In transmit state, time the transmission and arrange for the
18887          * retransmit.  In persist state, just set snd_max.
18888          */
18889         if (error == 0) {
18890                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18891                 if (rsm && doing_tlp) {
18892                         rack->rc_last_sent_tlp_past_cumack = 0;
18893                         rack->rc_last_sent_tlp_seq_valid = 1;
18894                         rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18895                         rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18896                 }
18897                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18898                 if (rsm && (doing_tlp == 0)) {
18899                         /* Set we retransmitted */
18900                         rack->rc_gp_saw_rec = 1;
18901                 } else {
18902                         if (cwnd_to_use > tp->snd_ssthresh) {
18903                                 /* Set we sent in CA */
18904                                 rack->rc_gp_saw_ca = 1;
18905                         } else {
18906                                 /* Set we sent in SS */
18907                                 rack->rc_gp_saw_ss = 1;
18908                         }
18909                 }
18910                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18911                     (tp->t_flags & TF_SACK_PERMIT) &&
18912                     tp->rcv_numsacks > 0)
18913                         tcp_clean_dsack_blocks(tp);
18914                 tot_len_this_send += len;
18915                 if (len == 0)
18916                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18917                 else if (len == 1) {
18918                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18919                 } else if (len > 1) {
18920                         int idx;
18921
18922                         idx = (len / segsiz) + 3;
18923                         if (idx >= TCP_MSS_ACCT_ATIMER)
18924                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18925                         else
18926                                 counter_u64_add(rack_out_size[idx], 1);
18927                 }
18928         }
18929         if ((rack->rack_no_prr == 0) &&
18930             sub_from_prr &&
18931             (error == 0)) {
18932                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18933                         rack->r_ctl.rc_prr_sndcnt -= len;
18934                 else
18935                         rack->r_ctl.rc_prr_sndcnt = 0;
18936         }
18937         sub_from_prr = 0;
18938         if (doing_tlp) {
18939                 /* Make sure the TLP is added */
18940                 add_flag |= RACK_TLP;
18941         } else if (rsm) {
18942                 /* If its a resend without TLP then it must not have the flag */
18943                 rsm->r_flags &= ~RACK_TLP;
18944         }
18945         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18946                         rack_to_usec_ts(&tv),
18947                         rsm, add_flag, s_mb, s_moff, hw_tls);
18948
18949
18950         if ((error == 0) &&
18951             (len > 0) &&
18952             (tp->snd_una == tp->snd_max))
18953                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18954         {
18955                 tcp_seq startseq = tp->snd_nxt;
18956
18957                 /* Track our lost count */
18958                 if (rsm && (doing_tlp == 0))
18959                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18960                 /*
18961                  * Advance snd_nxt over sequence space of this segment.
18962                  */
18963                 if (error)
18964                         /* We don't log or do anything with errors */
18965                         goto nomore;
18966                 if (doing_tlp == 0) {
18967                         if (rsm == NULL) {
18968                                 /*
18969                                  * Not a retransmission of some
18970                                  * sort, new data is going out so
18971                                  * clear our TLP count and flag.
18972                                  */
18973                                 rack->rc_tlp_in_progress = 0;
18974                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18975                         }
18976                 } else {
18977                         /*
18978                          * We have just sent a TLP, mark that it is true
18979                          * and make sure our in progress is set so we
18980                          * continue to check the count.
18981                          */
18982                         rack->rc_tlp_in_progress = 1;
18983                         rack->r_ctl.rc_tlp_cnt_out++;
18984                 }
18985                 if (flags & (TH_SYN | TH_FIN)) {
18986                         if (flags & TH_SYN)
18987                                 tp->snd_nxt++;
18988                         if (flags & TH_FIN) {
18989                                 tp->snd_nxt++;
18990                                 tp->t_flags |= TF_SENTFIN;
18991                         }
18992                 }
18993                 /* In the ENOBUFS case we do *not* update snd_max */
18994                 if (sack_rxmit)
18995                         goto nomore;
18996
18997                 tp->snd_nxt += len;
18998                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18999                         if (tp->snd_una == tp->snd_max) {
19000                                 /*
19001                                  * Update the time we just added data since
19002                                  * none was outstanding.
19003                                  */
19004                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19005                                 tp->t_acktime = ticks;
19006                         }
19007                         tp->snd_max = tp->snd_nxt;
19008                         /*
19009                          * Time this transmission if not a retransmission and
19010                          * not currently timing anything.
19011                          * This is only relevant in case of switching back to
19012                          * the base stack.
19013                          */
19014                         if (tp->t_rtttime == 0) {
19015                                 tp->t_rtttime = ticks;
19016                                 tp->t_rtseq = startseq;
19017                                 KMOD_TCPSTAT_INC(tcps_segstimed);
19018                         }
19019                         if (len &&
19020                             ((tp->t_flags & TF_GPUTINPROG) == 0))
19021                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
19022                 }
19023                 /*
19024                  * If we are doing FO we need to update the mbuf position and subtract
19025                  * this happens when the peer sends us duplicate information and
19026                  * we thus want to send a DSACK.
19027                  *
19028                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
19029                  * turned off? If not then we are going to echo multiple DSACK blocks
19030                  * out (with the TSO), which we should not be doing.
19031                  */
19032                 if (rack->r_fast_output && len) {
19033                         if (rack->r_ctl.fsb.left_to_send > len)
19034                                 rack->r_ctl.fsb.left_to_send -= len;
19035                         else
19036                                 rack->r_ctl.fsb.left_to_send = 0;
19037                         if (rack->r_ctl.fsb.left_to_send < segsiz)
19038                                 rack->r_fast_output = 0;
19039                         if (rack->r_fast_output) {
19040                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19041                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19042                         }
19043                 }
19044         }
19045 nomore:
19046         if (error) {
19047                 rack->r_ctl.rc_agg_delayed = 0;
19048                 rack->r_early = 0;
19049                 rack->r_late = 0;
19050                 rack->r_ctl.rc_agg_early = 0;
19051                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
19052                 /*
19053                  * Failures do not advance the seq counter above. For the
19054                  * case of ENOBUFS we will fall out and retry in 1ms with
19055                  * the hpts. Everything else will just have to retransmit
19056                  * with the timer.
19057                  *
19058                  * In any case, we do not want to loop around for another
19059                  * send without a good reason.
19060                  */
19061                 sendalot = 0;
19062                 switch (error) {
19063                 case EPERM:
19064                         tp->t_softerror = error;
19065 #ifdef TCP_ACCOUNTING
19066                         crtsc = get_cyclecount();
19067                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19068                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19069                         }
19070                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19071                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19072                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19073                         }
19074                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19075                         sched_unpin();
19076 #endif
19077                         return (error);
19078                 case ENOBUFS:
19079                         /*
19080                          * Pace us right away to retry in a some
19081                          * time
19082                          */
19083                         if (rack->r_ctl.crte != NULL) {
19084                                 rack_trace_point(rack, RACK_TP_HWENOBUF);
19085                         } else
19086                                 rack_trace_point(rack, RACK_TP_ENOBUF);
19087                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
19088                         if (rack->rc_enobuf < 0x7f)
19089                                 rack->rc_enobuf++;
19090                         if (slot < (10 * HPTS_USEC_IN_MSEC))
19091                                 slot = 10 * HPTS_USEC_IN_MSEC;
19092                         if (rack->r_ctl.crte != NULL) {
19093                                 counter_u64_add(rack_saw_enobuf_hw, 1);
19094                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
19095                         }
19096                         counter_u64_add(rack_saw_enobuf, 1);
19097                         goto enobufs;
19098                 case EMSGSIZE:
19099                         /*
19100                          * For some reason the interface we used initially
19101                          * to send segments changed to another or lowered
19102                          * its MTU. If TSO was active we either got an
19103                          * interface without TSO capabilits or TSO was
19104                          * turned off. If we obtained mtu from ip_output()
19105                          * then update it and try again.
19106                          */
19107                         if (tso)
19108                                 tp->t_flags &= ~TF_TSO;
19109                         if (mtu != 0) {
19110                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
19111                                 goto again;
19112                         }
19113                         slot = 10 * HPTS_USEC_IN_MSEC;
19114                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19115 #ifdef TCP_ACCOUNTING
19116                         crtsc = get_cyclecount();
19117                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19118                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19119                         }
19120                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19121                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19122                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19123                         }
19124                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19125                         sched_unpin();
19126 #endif
19127                         return (error);
19128                 case ENETUNREACH:
19129                         counter_u64_add(rack_saw_enetunreach, 1);
19130                 case EHOSTDOWN:
19131                 case EHOSTUNREACH:
19132                 case ENETDOWN:
19133                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
19134                                 tp->t_softerror = error;
19135                         }
19136                         /* FALLTHROUGH */
19137                 default:
19138                         slot = 10 * HPTS_USEC_IN_MSEC;
19139                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19140 #ifdef TCP_ACCOUNTING
19141                         crtsc = get_cyclecount();
19142                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19143                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19144                         }
19145                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19146                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19147                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19148                         }
19149                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19150                         sched_unpin();
19151 #endif
19152                         return (error);
19153                 }
19154         } else {
19155                 rack->rc_enobuf = 0;
19156                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19157                         rack->r_ctl.retran_during_recovery += len;
19158         }
19159         KMOD_TCPSTAT_INC(tcps_sndtotal);
19160
19161         /*
19162          * Data sent (as far as we can tell). If this advertises a larger
19163          * window than any other segment, then remember the size of the
19164          * advertised window. Any pending ACK has now been sent.
19165          */
19166         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
19167                 tp->rcv_adv = tp->rcv_nxt + recwin;
19168
19169         tp->last_ack_sent = tp->rcv_nxt;
19170         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19171 enobufs:
19172         if (sendalot) {
19173                 /* Do we need to turn off sendalot? */
19174                 if (rack->r_ctl.rc_pace_max_segs &&
19175                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
19176                         /* We hit our max. */
19177                         sendalot = 0;
19178                 } else if ((rack->rc_user_set_max_segs) &&
19179                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
19180                         /* We hit the user defined max */
19181                         sendalot = 0;
19182                 }
19183         }
19184         if ((error == 0) && (flags & TH_FIN))
19185                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
19186         if (flags & TH_RST) {
19187                 /*
19188                  * We don't send again after sending a RST.
19189                  */
19190                 slot = 0;
19191                 sendalot = 0;
19192                 if (error == 0)
19193                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
19194         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
19195                 /*
19196                  * Get our pacing rate, if an error
19197                  * occurred in sending (ENOBUF) we would
19198                  * hit the else if with slot preset. Other
19199                  * errors return.
19200                  */
19201                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
19202         }
19203         if (rsm &&
19204             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
19205             rack->use_rack_rr) {
19206                 /* Its a retransmit and we use the rack cheat? */
19207                 if ((slot == 0) ||
19208                     (rack->rc_always_pace == 0) ||
19209                     (rack->r_rr_config == 1)) {
19210                         /*
19211                          * We have no pacing set or we
19212                          * are using old-style rack or
19213                          * we are overridden to use the old 1ms pacing.
19214                          */
19215                         slot = rack->r_ctl.rc_min_to;
19216                 }
19217         }
19218         /* We have sent clear the flag */
19219         rack->r_ent_rec_ns = 0;
19220         if (rack->r_must_retran) {
19221                 if (rsm) {
19222                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
19223                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
19224                                 /*
19225                                  * We have retransmitted all.
19226                                  */
19227                                 rack->r_must_retran = 0;
19228                                 rack->r_ctl.rc_out_at_rto = 0;
19229                         }
19230                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19231                         /*
19232                          * Sending new data will also kill
19233                          * the loop.
19234                          */
19235                         rack->r_must_retran = 0;
19236                         rack->r_ctl.rc_out_at_rto = 0;
19237                 }
19238         }
19239         rack->r_ctl.fsb.recwin = recwin;
19240         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19241             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19242                 /*
19243                  * We hit an RTO and now have past snd_max at the RTO
19244                  * clear all the WAS flags.
19245                  */
19246                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19247         }
19248         if (slot) {
19249                 /* set the rack tcb into the slot N */
19250                 if ((error == 0) &&
19251                     rack_use_rfo &&
19252                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
19253                     (rsm == NULL) &&
19254                     (tp->snd_nxt == tp->snd_max) &&
19255                     (ipoptlen == 0) &&
19256                     (tp->rcv_numsacks == 0) &&
19257                     rack->r_fsb_inited &&
19258                     TCPS_HAVEESTABLISHED(tp->t_state) &&
19259                     (rack->r_must_retran == 0) &&
19260                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
19261                     (len > 0) && (orig_len > 0) &&
19262                     (orig_len > len) &&
19263                     ((orig_len - len) >= segsiz) &&
19264                     ((optlen == 0) ||
19265                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19266                         /* We can send at least one more MSS using our fsb */
19267
19268                         rack->r_fast_output = 1;
19269                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19270                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19271                         rack->r_ctl.fsb.tcp_flags = flags;
19272                         rack->r_ctl.fsb.left_to_send = orig_len - len;
19273                         if (hw_tls)
19274                                 rack->r_ctl.fsb.hw_tls = 1;
19275                         else
19276                                 rack->r_ctl.fsb.hw_tls = 0;
19277                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19278                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19279                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19280                                  (tp->snd_max - tp->snd_una)));
19281                         if (rack->r_ctl.fsb.left_to_send < segsiz)
19282                                 rack->r_fast_output = 0;
19283                         else {
19284                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19285                                         rack->r_ctl.fsb.rfo_apply_push = 1;
19286                                 else
19287                                         rack->r_ctl.fsb.rfo_apply_push = 0;
19288                         }
19289                 } else
19290                         rack->r_fast_output = 0;
19291                 rack_log_fsb(rack, tp, so, flags,
19292                              ipoptlen, orig_len, len, error,
19293                              (rsm == NULL), optlen, __LINE__, 2);
19294         } else if (sendalot) {
19295                 int ret;
19296
19297                 sack_rxmit = 0;
19298                 if ((error == 0) &&
19299                     rack_use_rfo &&
19300                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
19301                     (rsm == NULL) &&
19302                     (ipoptlen == 0) &&
19303                     (tp->rcv_numsacks == 0) &&
19304                     (tp->snd_nxt == tp->snd_max) &&
19305                     (rack->r_must_retran == 0) &&
19306                     rack->r_fsb_inited &&
19307                     TCPS_HAVEESTABLISHED(tp->t_state) &&
19308                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
19309                     (len > 0) && (orig_len > 0) &&
19310                     (orig_len > len) &&
19311                     ((orig_len - len) >= segsiz) &&
19312                     ((optlen == 0) ||
19313                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19314                         /* we can use fast_output for more */
19315
19316                         rack->r_fast_output = 1;
19317                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19318                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19319                         rack->r_ctl.fsb.tcp_flags = flags;
19320                         rack->r_ctl.fsb.left_to_send = orig_len - len;
19321                         if (hw_tls)
19322                                 rack->r_ctl.fsb.hw_tls = 1;
19323                         else
19324                                 rack->r_ctl.fsb.hw_tls = 0;
19325                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19326                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
19327                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19328                                  (tp->snd_max - tp->snd_una)));
19329                         if (rack->r_ctl.fsb.left_to_send < segsiz) {
19330                                 rack->r_fast_output = 0;
19331                         }
19332                         if (rack->r_fast_output) {
19333                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19334                                         rack->r_ctl.fsb.rfo_apply_push = 1;
19335                                 else
19336                                         rack->r_ctl.fsb.rfo_apply_push = 0;
19337                                 rack_log_fsb(rack, tp, so, flags,
19338                                              ipoptlen, orig_len, len, error,
19339                                              (rsm == NULL), optlen, __LINE__, 3);
19340                                 error = 0;
19341                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19342                                 if (ret >= 0)
19343                                         return (ret);
19344                                 else if (error)
19345                                         goto nomore;
19346
19347                         }
19348                 }
19349                 goto again;
19350         }
19351         /* Assure when we leave that snd_nxt will point to top */
19352         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19353                 tp->snd_nxt = tp->snd_max;
19354         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19355 #ifdef TCP_ACCOUNTING
19356         crtsc = get_cyclecount() - ts_val;
19357         if (tot_len_this_send) {
19358                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19359                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
19360                 }
19361                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19362                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19363                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19364                 }
19365                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19366                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19367                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19368                 }
19369                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19370         } else {
19371                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19372                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
19373                 }
19374                 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19375                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19376                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19377                 }
19378                 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19379         }
19380         sched_unpin();
19381 #endif
19382         if (error == ENOBUFS)
19383                 error = 0;
19384         return (error);
19385 }
19386
19387 static void
19388 rack_update_seg(struct tcp_rack *rack)
19389 {
19390         uint32_t orig_val;
19391
19392         orig_val = rack->r_ctl.rc_pace_max_segs;
19393         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19394         if (orig_val != rack->r_ctl.rc_pace_max_segs)
19395                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19396 }
19397
19398 static void
19399 rack_mtu_change(struct tcpcb *tp)
19400 {
19401         /*
19402          * The MSS may have changed
19403          */
19404         struct tcp_rack *rack;
19405         struct rack_sendmap *rsm;
19406
19407         rack = (struct tcp_rack *)tp->t_fb_ptr;
19408         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19409                 /*
19410                  * The MTU has changed we need to resend everything
19411                  * since all we have sent is lost. We first fix
19412                  * up the mtu though.
19413                  */
19414                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19415                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
19416                 rack_remxt_tmr(tp);
19417                 rack->r_fast_output = 0;
19418                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19419                                                 rack->r_ctl.rc_sacked);
19420                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19421                 rack->r_must_retran = 1;
19422                 /* Mark all inflight to needing to be rxt'd */
19423                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19424                         rsm->r_flags |= RACK_MUST_RXT;
19425                 }
19426         }
19427         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19428         /* We don't use snd_nxt to retransmit */
19429         tp->snd_nxt = tp->snd_max;
19430 }
19431
19432 static int
19433 rack_set_profile(struct tcp_rack *rack, int prof)
19434 {
19435         int err = EINVAL;
19436         if (prof == 1) {
19437                 /* pace_always=1 */
19438                 if (rack->rc_always_pace == 0) {
19439                         if (tcp_can_enable_pacing() == 0)
19440                                 return (EBUSY);
19441                 }
19442                 rack->rc_always_pace = 1;
19443                 if (rack->use_fixed_rate || rack->gp_ready)
19444                         rack_set_cc_pacing(rack);
19445                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19446                 rack->rack_attempt_hdwr_pace = 0;
19447                 /* cmpack=1 */
19448                 if (rack_use_cmp_acks)
19449                         rack->r_use_cmp_ack = 1;
19450                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19451                     rack->r_use_cmp_ack)
19452                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19453                 /* scwnd=1 */
19454                 rack->rack_enable_scwnd = 1;
19455                 /* dynamic=100 */
19456                 rack->rc_gp_dyn_mul = 1;
19457                 /* gp_inc_ca */
19458                 rack->r_ctl.rack_per_of_gp_ca = 100;
19459                 /* rrr_conf=3 */
19460                 rack->r_rr_config = 3;
19461                 /* npush=2 */
19462                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19463                 /* fillcw=1 */
19464                 rack->rc_pace_to_cwnd = 1;
19465                 rack->rc_pace_fill_if_rttin_range = 0;
19466                 rack->rtt_limit_mul = 0;
19467                 /* noprr=1 */
19468                 rack->rack_no_prr = 1;
19469                 /* lscwnd=1 */
19470                 rack->r_limit_scw = 1;
19471                 /* gp_inc_rec */
19472                 rack->r_ctl.rack_per_of_gp_rec = 90;
19473                 err = 0;
19474
19475         } else if (prof == 3) {
19476                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19477                 /* pace_always=1 */
19478                 if (rack->rc_always_pace == 0) {
19479                         if (tcp_can_enable_pacing() == 0)
19480                                 return (EBUSY);
19481                 }
19482                 rack->rc_always_pace = 1;
19483                 if (rack->use_fixed_rate || rack->gp_ready)
19484                         rack_set_cc_pacing(rack);
19485                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19486                 rack->rack_attempt_hdwr_pace = 0;
19487                 /* cmpack=1 */
19488                 if (rack_use_cmp_acks)
19489                         rack->r_use_cmp_ack = 1;
19490                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19491                     rack->r_use_cmp_ack)
19492                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19493                 /* scwnd=1 */
19494                 rack->rack_enable_scwnd = 1;
19495                 /* dynamic=100 */
19496                 rack->rc_gp_dyn_mul = 1;
19497                 /* gp_inc_ca */
19498                 rack->r_ctl.rack_per_of_gp_ca = 100;
19499                 /* rrr_conf=3 */
19500                 rack->r_rr_config = 3;
19501                 /* npush=2 */
19502                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19503                 /* fillcw=2 */
19504                 rack->rc_pace_to_cwnd = 1;
19505                 rack->r_fill_less_agg = 1;
19506                 rack->rc_pace_fill_if_rttin_range = 0;
19507                 rack->rtt_limit_mul = 0;
19508                 /* noprr=1 */
19509                 rack->rack_no_prr = 1;
19510                 /* lscwnd=1 */
19511                 rack->r_limit_scw = 1;
19512                 /* gp_inc_rec */
19513                 rack->r_ctl.rack_per_of_gp_rec = 90;
19514                 err = 0;
19515
19516
19517         } else if (prof == 2) {
19518                 /* cmpack=1 */
19519                 if (rack->rc_always_pace == 0) {
19520                         if (tcp_can_enable_pacing() == 0)
19521                                 return (EBUSY);
19522                 }
19523                 rack->rc_always_pace = 1;
19524                 if (rack->use_fixed_rate || rack->gp_ready)
19525                         rack_set_cc_pacing(rack);
19526                 rack->r_use_cmp_ack = 1;
19527                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19528                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19529                 /* pace_always=1 */
19530                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19531                 /* scwnd=1 */
19532                 rack->rack_enable_scwnd = 1;
19533                 /* dynamic=100 */
19534                 rack->rc_gp_dyn_mul = 1;
19535                 rack->r_ctl.rack_per_of_gp_ca = 100;
19536                 /* rrr_conf=3 */
19537                 rack->r_rr_config = 3;
19538                 /* npush=2 */
19539                 rack->r_ctl.rc_no_push_at_mrtt = 2;
19540                 /* fillcw=1 */
19541                 rack->rc_pace_to_cwnd = 1;
19542                 rack->rc_pace_fill_if_rttin_range = 0;
19543                 rack->rtt_limit_mul = 0;
19544                 /* noprr=1 */
19545                 rack->rack_no_prr = 1;
19546                 /* lscwnd=0 */
19547                 rack->r_limit_scw = 0;
19548                 err = 0;
19549         } else if (prof == 0) {
19550                 /* This changes things back to the default settings */
19551                 err = 0;
19552                 if (rack->rc_always_pace) {
19553                         tcp_decrement_paced_conn();
19554                         rack_undo_cc_pacing(rack);
19555                         rack->rc_always_pace = 0;
19556                 }
19557                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19558                         rack->rc_always_pace = 1;
19559                         if (rack->use_fixed_rate || rack->gp_ready)
19560                                 rack_set_cc_pacing(rack);
19561                 } else
19562                         rack->rc_always_pace = 0;
19563                 if (rack_dsack_std_based & 0x1) {
19564                         /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19565                         rack->rc_rack_tmr_std_based = 1;
19566                 }
19567                 if (rack_dsack_std_based & 0x2) {
19568                         /* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19569                         rack->rc_rack_use_dsack = 1;
19570                 }
19571                 if (rack_use_cmp_acks)
19572                         rack->r_use_cmp_ack = 1;
19573                 else
19574                         rack->r_use_cmp_ack = 0;
19575                 if (rack_disable_prr)
19576                         rack->rack_no_prr = 1;
19577                 else
19578                         rack->rack_no_prr = 0;
19579                 if (rack_gp_no_rec_chg)
19580                         rack->rc_gp_no_rec_chg = 1;
19581                 else
19582                         rack->rc_gp_no_rec_chg = 0;
19583                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19584                         rack->r_mbuf_queue = 1;
19585                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19586                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19587                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19588                 } else {
19589                         rack->r_mbuf_queue = 0;
19590                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19591                 }
19592                 if (rack_enable_shared_cwnd)
19593                         rack->rack_enable_scwnd = 1;
19594                 else
19595                         rack->rack_enable_scwnd = 0;
19596                 if (rack_do_dyn_mul) {
19597                         /* When dynamic adjustment is on CA needs to start at 100% */
19598                         rack->rc_gp_dyn_mul = 1;
19599                         if (rack_do_dyn_mul >= 100)
19600                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19601                 } else {
19602                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19603                         rack->rc_gp_dyn_mul = 0;
19604                 }
19605                 rack->r_rr_config = 0;
19606                 rack->r_ctl.rc_no_push_at_mrtt = 0;
19607                 rack->rc_pace_to_cwnd = 0;
19608                 rack->rc_pace_fill_if_rttin_range = 0;
19609                 rack->rtt_limit_mul = 0;
19610
19611                 if (rack_enable_hw_pacing)
19612                         rack->rack_hdw_pace_ena = 1;
19613                 else
19614                         rack->rack_hdw_pace_ena = 0;
19615                 if (rack_disable_prr)
19616                         rack->rack_no_prr = 1;
19617                 else
19618                         rack->rack_no_prr = 0;
19619                 if (rack_limits_scwnd)
19620                         rack->r_limit_scw  = 1;
19621                 else
19622                         rack->r_limit_scw  = 0;
19623                 err = 0;
19624         }
19625         return (err);
19626 }
19627
19628 static int
19629 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19630 {
19631         struct deferred_opt_list *dol;
19632
19633         dol = malloc(sizeof(struct deferred_opt_list),
19634                      M_TCPFSB, M_NOWAIT|M_ZERO);
19635         if (dol == NULL) {
19636                 /*
19637                  * No space yikes -- fail out..
19638                  */
19639                 return (0);
19640         }
19641         dol->optname = sopt_name;
19642         dol->optval = loptval;
19643         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19644         return (1);
19645 }
19646
19647 static int
19648 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19649                     uint32_t optval, uint64_t loptval)
19650 {
19651         struct epoch_tracker et;
19652         struct sockopt sopt;
19653         struct cc_newreno_opts opt;
19654         uint64_t val;
19655         int error = 0;
19656         uint16_t ca, ss;
19657
19658         switch (sopt_name) {
19659
19660         case TCP_RACK_DSACK_OPT:
19661                 RACK_OPTS_INC(tcp_rack_dsack_opt);
19662                 if (optval & 0x1) {
19663                         rack->rc_rack_tmr_std_based = 1;
19664                 } else {
19665                         rack->rc_rack_tmr_std_based = 0;
19666                 }
19667                 if (optval & 0x2) {
19668                         rack->rc_rack_use_dsack = 1;
19669                 } else {
19670                         rack->rc_rack_use_dsack = 0;
19671                 }
19672                 rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19673                 break;
19674         case TCP_RACK_PACING_BETA:
19675                 RACK_OPTS_INC(tcp_rack_beta);
19676                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19677                         /* This only works for newreno. */
19678                         error = EINVAL;
19679                         break;
19680                 }
19681                 if (rack->rc_pacing_cc_set) {
19682                         /*
19683                          * Set them into the real CC module
19684                          * whats in the rack pcb is the old values
19685                          * to be used on restoral/
19686                          */
19687                         sopt.sopt_dir = SOPT_SET;
19688                         opt.name = CC_NEWRENO_BETA;
19689                         opt.val = optval;
19690                         if (CC_ALGO(tp)->ctl_output != NULL)
19691                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19692                         else {
19693                                 error = ENOENT;
19694                                 break;
19695                         }
19696                 } else {
19697                         /*
19698                          * Not pacing yet so set it into our local
19699                          * rack pcb storage.
19700                          */
19701                         rack->r_ctl.rc_saved_beta.beta = optval;
19702                 }
19703                 break;
19704         case TCP_RACK_TIMER_SLOP:
19705                 RACK_OPTS_INC(tcp_rack_timer_slop);
19706                 rack->r_ctl.timer_slop = optval;
19707                 if (rack->rc_tp->t_srtt) {
19708                         /*
19709                          * If we have an SRTT lets update t_rxtcur
19710                          * to have the new slop.
19711                          */
19712                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19713                                            rack_rto_min, rack_rto_max,
19714                                            rack->r_ctl.timer_slop);
19715                 }
19716                 break;
19717         case TCP_RACK_PACING_BETA_ECN:
19718                 RACK_OPTS_INC(tcp_rack_beta_ecn);
19719                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19720                         /* This only works for newreno. */
19721                         error = EINVAL;
19722                         break;
19723                 }
19724                 if (rack->rc_pacing_cc_set) {
19725                         /*
19726                          * Set them into the real CC module
19727                          * whats in the rack pcb is the old values
19728                          * to be used on restoral/
19729                          */
19730                         sopt.sopt_dir = SOPT_SET;
19731                         opt.name = CC_NEWRENO_BETA_ECN;
19732                         opt.val = optval;
19733                         if (CC_ALGO(tp)->ctl_output != NULL)
19734                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19735                         else
19736                                 error = ENOENT;
19737                 } else {
19738                         /*
19739                          * Not pacing yet so set it into our local
19740                          * rack pcb storage.
19741                          */
19742                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19743                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19744                 }
19745                 break;
19746         case TCP_DEFER_OPTIONS:
19747                 RACK_OPTS_INC(tcp_defer_opt);
19748                 if (optval) {
19749                         if (rack->gp_ready) {
19750                                 /* Too late */
19751                                 error = EINVAL;
19752                                 break;
19753                         }
19754                         rack->defer_options = 1;
19755                 } else
19756                         rack->defer_options = 0;
19757                 break;
19758         case TCP_RACK_MEASURE_CNT:
19759                 RACK_OPTS_INC(tcp_rack_measure_cnt);
19760                 if (optval && (optval <= 0xff)) {
19761                         rack->r_ctl.req_measurements = optval;
19762                 } else
19763                         error = EINVAL;
19764                 break;
19765         case TCP_REC_ABC_VAL:
19766                 RACK_OPTS_INC(tcp_rec_abc_val);
19767                 if (optval > 0)
19768                         rack->r_use_labc_for_rec = 1;
19769                 else
19770                         rack->r_use_labc_for_rec = 0;
19771                 break;
19772         case TCP_RACK_ABC_VAL:
19773                 RACK_OPTS_INC(tcp_rack_abc_val);
19774                 if ((optval > 0) && (optval < 255))
19775                         rack->rc_labc = optval;
19776                 else
19777                         error = EINVAL;
19778                 break;
19779         case TCP_HDWR_UP_ONLY:
19780                 RACK_OPTS_INC(tcp_pacing_up_only);
19781                 if (optval)
19782                         rack->r_up_only = 1;
19783                 else
19784                         rack->r_up_only = 0;
19785                 break;
19786         case TCP_PACING_RATE_CAP:
19787                 RACK_OPTS_INC(tcp_pacing_rate_cap);
19788                 rack->r_ctl.bw_rate_cap = loptval;
19789                 break;
19790         case TCP_RACK_PROFILE:
19791                 RACK_OPTS_INC(tcp_profile);
19792                 error = rack_set_profile(rack, optval);
19793                 break;
19794         case TCP_USE_CMP_ACKS:
19795                 RACK_OPTS_INC(tcp_use_cmp_acks);
19796                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19797                         /* You can't turn it off once its on! */
19798                         error = EINVAL;
19799                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19800                         rack->r_use_cmp_ack = 1;
19801                         rack->r_mbuf_queue = 1;
19802                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19803                 }
19804                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19805                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19806                 break;
19807         case TCP_SHARED_CWND_TIME_LIMIT:
19808                 RACK_OPTS_INC(tcp_lscwnd);
19809                 if (optval)
19810                         rack->r_limit_scw = 1;
19811                 else
19812                         rack->r_limit_scw = 0;
19813                 break;
19814         case TCP_RACK_PACE_TO_FILL:
19815                 RACK_OPTS_INC(tcp_fillcw);
19816                 if (optval == 0)
19817                         rack->rc_pace_to_cwnd = 0;
19818                 else {
19819                         rack->rc_pace_to_cwnd = 1;
19820                         if (optval > 1)
19821                                 rack->r_fill_less_agg = 1;
19822                 }
19823                 if ((optval >= rack_gp_rtt_maxmul) &&
19824                     rack_gp_rtt_maxmul &&
19825                     (optval < 0xf)) {
19826                         rack->rc_pace_fill_if_rttin_range = 1;
19827                         rack->rtt_limit_mul = optval;
19828                 } else {
19829                         rack->rc_pace_fill_if_rttin_range = 0;
19830                         rack->rtt_limit_mul = 0;
19831                 }
19832                 break;
19833         case TCP_RACK_NO_PUSH_AT_MAX:
19834                 RACK_OPTS_INC(tcp_npush);
19835                 if (optval == 0)
19836                         rack->r_ctl.rc_no_push_at_mrtt = 0;
19837                 else if (optval < 0xff)
19838                         rack->r_ctl.rc_no_push_at_mrtt = optval;
19839                 else
19840                         error = EINVAL;
19841                 break;
19842         case TCP_SHARED_CWND_ENABLE:
19843                 RACK_OPTS_INC(tcp_rack_scwnd);
19844                 if (optval == 0)
19845                         rack->rack_enable_scwnd = 0;
19846                 else
19847                         rack->rack_enable_scwnd = 1;
19848                 break;
19849         case TCP_RACK_MBUF_QUEUE:
19850                 /* Now do we use the LRO mbuf-queue feature */
19851                 RACK_OPTS_INC(tcp_rack_mbufq);
19852                 if (optval || rack->r_use_cmp_ack)
19853                         rack->r_mbuf_queue = 1;
19854                 else
19855                         rack->r_mbuf_queue = 0;
19856                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19857                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19858                 else
19859                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19860                 break;
19861         case TCP_RACK_NONRXT_CFG_RATE:
19862                 RACK_OPTS_INC(tcp_rack_cfg_rate);
19863                 if (optval == 0)
19864                         rack->rack_rec_nonrxt_use_cr = 0;
19865                 else
19866                         rack->rack_rec_nonrxt_use_cr = 1;
19867                 break;
19868         case TCP_NO_PRR:
19869                 RACK_OPTS_INC(tcp_rack_noprr);
19870                 if (optval == 0)
19871                         rack->rack_no_prr = 0;
19872                 else if (optval == 1)
19873                         rack->rack_no_prr = 1;
19874                 else if (optval == 2)
19875                         rack->no_prr_addback = 1;
19876                 else
19877                         error = EINVAL;
19878                 break;
19879         case TCP_TIMELY_DYN_ADJ:
19880                 RACK_OPTS_INC(tcp_timely_dyn);
19881                 if (optval == 0)
19882                         rack->rc_gp_dyn_mul = 0;
19883                 else {
19884                         rack->rc_gp_dyn_mul = 1;
19885                         if (optval >= 100) {
19886                                 /*
19887                                  * If the user sets something 100 or more
19888                                  * its the gp_ca value.
19889                                  */
19890                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19891                         }
19892                 }
19893                 break;
19894         case TCP_RACK_DO_DETECTION:
19895                 RACK_OPTS_INC(tcp_rack_do_detection);
19896                 if (optval == 0)
19897                         rack->do_detection = 0;
19898                 else
19899                         rack->do_detection = 1;
19900                 break;
19901         case TCP_RACK_TLP_USE:
19902                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19903                         error = EINVAL;
19904                         break;
19905                 }
19906                 RACK_OPTS_INC(tcp_tlp_use);
19907                 rack->rack_tlp_threshold_use = optval;
19908                 break;
19909         case TCP_RACK_TLP_REDUCE:
19910                 /* RACK TLP cwnd reduction (bool) */
19911                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19912                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19913                 break;
19914         /*  Pacing related ones */
19915         case TCP_RACK_PACE_ALWAYS:
19916                 /*
19917                  * zero is old rack method, 1 is new
19918                  * method using a pacing rate.
19919                  */
19920                 RACK_OPTS_INC(tcp_rack_pace_always);
19921                 if (optval > 0) {
19922                         if (rack->rc_always_pace) {
19923                                 error = EALREADY;
19924                                 break;
19925                         } else if (tcp_can_enable_pacing()) {
19926                                 rack->rc_always_pace = 1;
19927                                 if (rack->use_fixed_rate || rack->gp_ready)
19928                                         rack_set_cc_pacing(rack);
19929                         }
19930                         else {
19931                                 error = ENOSPC;
19932                                 break;
19933                         }
19934                 } else {
19935                         if (rack->rc_always_pace) {
19936                                 tcp_decrement_paced_conn();
19937                                 rack->rc_always_pace = 0;
19938                                 rack_undo_cc_pacing(rack);
19939                         }
19940                 }
19941                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19942                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19943                 else
19944                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19945                 /* A rate may be set irate or other, if so set seg size */
19946                 rack_update_seg(rack);
19947                 break;
19948         case TCP_BBR_RACK_INIT_RATE:
19949                 RACK_OPTS_INC(tcp_initial_rate);
19950                 val = optval;
19951                 /* Change from kbits per second to bytes per second */
19952                 val *= 1000;
19953                 val /= 8;
19954                 rack->r_ctl.init_rate = val;
19955                 if (rack->rc_init_win != rack_default_init_window) {
19956                         uint32_t win, snt;
19957
19958                         /*
19959                          * Options don't always get applied
19960                          * in the order you think. So in order
19961                          * to assure we update a cwnd we need
19962                          * to check and see if we are still
19963                          * where we should raise the cwnd.
19964                          */
19965                         win = rc_init_window(rack);
19966                         if (SEQ_GT(tp->snd_max, tp->iss))
19967                                 snt = tp->snd_max - tp->iss;
19968                         else
19969                                 snt = 0;
19970                         if ((snt < win) &&
19971                             (tp->snd_cwnd < win))
19972                                 tp->snd_cwnd = win;
19973                 }
19974                 if (rack->rc_always_pace)
19975                         rack_update_seg(rack);
19976                 break;
19977         case TCP_BBR_IWINTSO:
19978                 RACK_OPTS_INC(tcp_initial_win);
19979                 if (optval && (optval <= 0xff)) {
19980                         uint32_t win, snt;
19981
19982                         rack->rc_init_win = optval;
19983                         win = rc_init_window(rack);
19984                         if (SEQ_GT(tp->snd_max, tp->iss))
19985                                 snt = tp->snd_max - tp->iss;
19986                         else
19987                                 snt = 0;
19988                         if ((snt < win) &&
19989                             (tp->t_srtt |
19990 #ifdef NETFLIX_PEAKRATE
19991                              tp->t_maxpeakrate |
19992 #endif
19993                              rack->r_ctl.init_rate)) {
19994                                 /*
19995                                  * We are not past the initial window
19996                                  * and we have some bases for pacing,
19997                                  * so we need to possibly adjust up
19998                                  * the cwnd. Note even if we don't set
19999                                  * the cwnd, its still ok to raise the rc_init_win
20000                                  * which can be used coming out of idle when we
20001                                  * would have a rate.
20002                                  */
20003                                 if (tp->snd_cwnd < win)
20004                                         tp->snd_cwnd = win;
20005                         }
20006                         if (rack->rc_always_pace)
20007                                 rack_update_seg(rack);
20008                 } else
20009                         error = EINVAL;
20010                 break;
20011         case TCP_RACK_FORCE_MSEG:
20012                 RACK_OPTS_INC(tcp_rack_force_max_seg);
20013                 if (optval)
20014                         rack->rc_force_max_seg = 1;
20015                 else
20016                         rack->rc_force_max_seg = 0;
20017                 break;
20018         case TCP_RACK_PACE_MAX_SEG:
20019                 /* Max segments size in a pace in bytes */
20020                 RACK_OPTS_INC(tcp_rack_max_seg);
20021                 rack->rc_user_set_max_segs = optval;
20022                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
20023                 break;
20024         case TCP_RACK_PACE_RATE_REC:
20025                 /* Set the fixed pacing rate in Bytes per second ca */
20026                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
20027                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20028                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
20029                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20030                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
20031                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20032                 rack->use_fixed_rate = 1;
20033                 if (rack->rc_always_pace)
20034                         rack_set_cc_pacing(rack);
20035                 rack_log_pacing_delay_calc(rack,
20036                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
20037                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
20038                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20039                                            __LINE__, NULL,0);
20040                 break;
20041
20042         case TCP_RACK_PACE_RATE_SS:
20043                 /* Set the fixed pacing rate in Bytes per second ca */
20044                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
20045                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20046                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
20047                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20048                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
20049                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20050                 rack->use_fixed_rate = 1;
20051                 if (rack->rc_always_pace)
20052                         rack_set_cc_pacing(rack);
20053                 rack_log_pacing_delay_calc(rack,
20054                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
20055                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
20056                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20057                                            __LINE__, NULL, 0);
20058                 break;
20059
20060         case TCP_RACK_PACE_RATE_CA:
20061                 /* Set the fixed pacing rate in Bytes per second ca */
20062                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
20063                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20064                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
20065                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20066                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
20067                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20068                 rack->use_fixed_rate = 1;
20069                 if (rack->rc_always_pace)
20070                         rack_set_cc_pacing(rack);
20071                 rack_log_pacing_delay_calc(rack,
20072                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
20073                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
20074                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20075                                            __LINE__, NULL, 0);
20076                 break;
20077         case TCP_RACK_GP_INCREASE_REC:
20078                 RACK_OPTS_INC(tcp_gp_inc_rec);
20079                 rack->r_ctl.rack_per_of_gp_rec = optval;
20080                 rack_log_pacing_delay_calc(rack,
20081                                            rack->r_ctl.rack_per_of_gp_ss,
20082                                            rack->r_ctl.rack_per_of_gp_ca,
20083                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20084                                            __LINE__, NULL, 0);
20085                 break;
20086         case TCP_RACK_GP_INCREASE_CA:
20087                 RACK_OPTS_INC(tcp_gp_inc_ca);
20088                 ca = optval;
20089                 if (ca < 100) {
20090                         /*
20091                          * We don't allow any reduction
20092                          * over the GP b/w.
20093                          */
20094                         error = EINVAL;
20095                         break;
20096                 }
20097                 rack->r_ctl.rack_per_of_gp_ca = ca;
20098                 rack_log_pacing_delay_calc(rack,
20099                                            rack->r_ctl.rack_per_of_gp_ss,
20100                                            rack->r_ctl.rack_per_of_gp_ca,
20101                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20102                                            __LINE__, NULL, 0);
20103                 break;
20104         case TCP_RACK_GP_INCREASE_SS:
20105                 RACK_OPTS_INC(tcp_gp_inc_ss);
20106                 ss = optval;
20107                 if (ss < 100) {
20108                         /*
20109                          * We don't allow any reduction
20110                          * over the GP b/w.
20111                          */
20112                         error = EINVAL;
20113                         break;
20114                 }
20115                 rack->r_ctl.rack_per_of_gp_ss = ss;
20116                 rack_log_pacing_delay_calc(rack,
20117                                            rack->r_ctl.rack_per_of_gp_ss,
20118                                            rack->r_ctl.rack_per_of_gp_ca,
20119                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20120                                            __LINE__, NULL, 0);
20121                 break;
20122         case TCP_RACK_RR_CONF:
20123                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
20124                 if (optval && optval <= 3)
20125                         rack->r_rr_config = optval;
20126                 else
20127                         rack->r_rr_config = 0;
20128                 break;
20129         case TCP_HDWR_RATE_CAP:
20130                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
20131                 if (optval) {
20132                         if (rack->r_rack_hw_rate_caps == 0)
20133                                 rack->r_rack_hw_rate_caps = 1;
20134                         else
20135                                 error = EALREADY;
20136                 } else {
20137                         rack->r_rack_hw_rate_caps = 0;
20138                 }
20139                 break;
20140         case TCP_BBR_HDWR_PACE:
20141                 RACK_OPTS_INC(tcp_hdwr_pacing);
20142                 if (optval){
20143                         if (rack->rack_hdrw_pacing == 0) {
20144                                 rack->rack_hdw_pace_ena = 1;
20145                                 rack->rack_attempt_hdwr_pace = 0;
20146                         } else
20147                                 error = EALREADY;
20148                 } else {
20149                         rack->rack_hdw_pace_ena = 0;
20150 #ifdef RATELIMIT
20151                         if (rack->r_ctl.crte != NULL) {
20152                                 rack->rack_hdrw_pacing = 0;
20153                                 rack->rack_attempt_hdwr_pace = 0;
20154                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
20155                                 rack->r_ctl.crte = NULL;
20156                         }
20157 #endif
20158                 }
20159                 break;
20160         /*  End Pacing related ones */
20161         case TCP_RACK_PRR_SENDALOT:
20162                 /* Allow PRR to send more than one seg */
20163                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
20164                 rack->r_ctl.rc_prr_sendalot = optval;
20165                 break;
20166         case TCP_RACK_MIN_TO:
20167                 /* Minimum time between rack t-o's in ms */
20168                 RACK_OPTS_INC(tcp_rack_min_to);
20169                 rack->r_ctl.rc_min_to = optval;
20170                 break;
20171         case TCP_RACK_EARLY_SEG:
20172                 /* If early recovery max segments */
20173                 RACK_OPTS_INC(tcp_rack_early_seg);
20174                 rack->r_ctl.rc_early_recovery_segs = optval;
20175                 break;
20176         case TCP_RACK_ENABLE_HYSTART:
20177         {
20178                 if (optval) {
20179                         tp->ccv->flags |= CCF_HYSTART_ALLOWED;
20180                         if (rack_do_hystart > RACK_HYSTART_ON)
20181                                 tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND;
20182                         if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
20183                                 tp->ccv->flags |= CCF_HYSTART_CONS_SSTH;
20184                 } else {
20185                         tp->ccv->flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
20186                 }
20187         }
20188         break;
20189         case TCP_RACK_REORD_THRESH:
20190                 /* RACK reorder threshold (shift amount) */
20191                 RACK_OPTS_INC(tcp_rack_reord_thresh);
20192                 if ((optval > 0) && (optval < 31))
20193                         rack->r_ctl.rc_reorder_shift = optval;
20194                 else
20195                         error = EINVAL;
20196                 break;
20197         case TCP_RACK_REORD_FADE:
20198                 /* Does reordering fade after ms time */
20199                 RACK_OPTS_INC(tcp_rack_reord_fade);
20200                 rack->r_ctl.rc_reorder_fade = optval;
20201                 break;
20202         case TCP_RACK_TLP_THRESH:
20203                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
20204                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
20205                 if (optval)
20206                         rack->r_ctl.rc_tlp_threshold = optval;
20207                 else
20208                         error = EINVAL;
20209                 break;
20210         case TCP_BBR_USE_RACK_RR:
20211                 RACK_OPTS_INC(tcp_rack_rr);
20212                 if (optval)
20213                         rack->use_rack_rr = 1;
20214                 else
20215                         rack->use_rack_rr = 0;
20216                 break;
20217         case TCP_FAST_RSM_HACK:
20218                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
20219                 if (optval)
20220                         rack->fast_rsm_hack = 1;
20221                 else
20222                         rack->fast_rsm_hack = 0;
20223                 break;
20224         case TCP_RACK_PKT_DELAY:
20225                 /* RACK added ms i.e. rack-rtt + reord + N */
20226                 RACK_OPTS_INC(tcp_rack_pkt_delay);
20227                 rack->r_ctl.rc_pkt_delay = optval;
20228                 break;
20229         case TCP_DELACK:
20230                 RACK_OPTS_INC(tcp_rack_delayed_ack);
20231                 if (optval == 0)
20232                         tp->t_delayed_ack = 0;
20233                 else
20234                         tp->t_delayed_ack = 1;
20235                 if (tp->t_flags & TF_DELACK) {
20236                         tp->t_flags &= ~TF_DELACK;
20237                         tp->t_flags |= TF_ACKNOW;
20238                         NET_EPOCH_ENTER(et);
20239                         rack_output(tp);
20240                         NET_EPOCH_EXIT(et);
20241                 }
20242                 break;
20243
20244         case TCP_BBR_RACK_RTT_USE:
20245                 RACK_OPTS_INC(tcp_rack_rtt_use);
20246                 if ((optval != USE_RTT_HIGH) &&
20247                     (optval != USE_RTT_LOW) &&
20248                     (optval != USE_RTT_AVG))
20249                         error = EINVAL;
20250                 else
20251                         rack->r_ctl.rc_rate_sample_method = optval;
20252                 break;
20253         case TCP_DATA_AFTER_CLOSE:
20254                 RACK_OPTS_INC(tcp_data_after_close);
20255                 if (optval)
20256                         rack->rc_allow_data_af_clo = 1;
20257                 else
20258                         rack->rc_allow_data_af_clo = 0;
20259                 break;
20260         default:
20261                 break;
20262         }
20263 #ifdef NETFLIX_STATS
20264         tcp_log_socket_option(tp, sopt_name, optval, error);
20265 #endif
20266         return (error);
20267 }
20268
20269
20270 static void
20271 rack_apply_deferred_options(struct tcp_rack *rack)
20272 {
20273         struct deferred_opt_list *dol, *sdol;
20274         uint32_t s_optval;
20275
20276         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20277                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20278                 /* Disadvantage of deferal is you loose the error return */
20279                 s_optval = (uint32_t)dol->optval;
20280                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20281                 free(dol, M_TCPDO);
20282         }
20283 }
20284
20285 static void
20286 rack_hw_tls_change(struct tcpcb *tp, int chg)
20287 {
20288         /*
20289          * HW tls state has changed.. fix all
20290          * rsm's in flight.
20291          */
20292         struct tcp_rack *rack;
20293         struct rack_sendmap *rsm;
20294
20295         rack = (struct tcp_rack *)tp->t_fb_ptr;
20296         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
20297                 if (chg)
20298                         rsm->r_hw_tls = 1;
20299                 else
20300                         rsm->r_hw_tls = 0;
20301         }
20302         if (chg)
20303                 rack->r_ctl.fsb.hw_tls = 1;
20304         else
20305                 rack->r_ctl.fsb.hw_tls = 0;
20306 }
20307
20308 static int
20309 rack_pru_options(struct tcpcb *tp, int flags)
20310 {
20311         if (flags & PRUS_OOB)
20312                 return (EOPNOTSUPP);
20313         return (0);
20314 }
20315
20316 static struct tcp_function_block __tcp_rack = {
20317         .tfb_tcp_block_name = __XSTRING(STACKNAME),
20318         .tfb_tcp_output = rack_output,
20319         .tfb_do_queued_segments = ctf_do_queued_segments,
20320         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
20321         .tfb_tcp_do_segment = rack_do_segment,
20322         .tfb_tcp_ctloutput = rack_ctloutput,
20323         .tfb_tcp_fb_init = rack_init,
20324         .tfb_tcp_fb_fini = rack_fini,
20325         .tfb_tcp_timer_stop_all = rack_stopall,
20326         .tfb_tcp_timer_activate = rack_timer_activate,
20327         .tfb_tcp_timer_active = rack_timer_active,
20328         .tfb_tcp_timer_stop = rack_timer_stop,
20329         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20330         .tfb_tcp_handoff_ok = rack_handoff_ok,
20331         .tfb_tcp_mtu_chg = rack_mtu_change,
20332         .tfb_pru_options = rack_pru_options,
20333         .tfb_hwtls_change = rack_hw_tls_change,
20334         .tfb_compute_pipe = rack_compute_pipe,
20335         .tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20336 };
20337
20338 /*
20339  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20340  * socket option arguments.  When it re-acquires the lock after the copy, it
20341  * has to revalidate that the connection is still valid for the socket
20342  * option.
20343  */
20344 static int
20345 rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
20346 {
20347 #ifdef INET6
20348         struct ip6_hdr *ip6;
20349 #endif
20350 #ifdef INET
20351         struct ip *ip;
20352 #endif
20353         struct tcpcb *tp;
20354         struct tcp_rack *rack;
20355         uint64_t loptval;
20356         int32_t error = 0, optval;
20357
20358         tp = intotcpcb(inp);
20359         rack = (struct tcp_rack *)tp->t_fb_ptr;
20360         if (rack == NULL) {
20361                 INP_WUNLOCK(inp);
20362                 return (EINVAL);
20363         }
20364 #ifdef INET6
20365         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20366 #endif
20367 #ifdef INET
20368         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20369 #endif
20370
20371         switch (sopt->sopt_level) {
20372 #ifdef INET6
20373         case IPPROTO_IPV6:
20374                 MPASS(inp->inp_vflag & INP_IPV6PROTO);
20375                 switch (sopt->sopt_name) {
20376                 case IPV6_USE_MIN_MTU:
20377                         tcp6_use_min_mtu(tp);
20378                         break;
20379                 case IPV6_TCLASS:
20380                         /*
20381                          * The DSCP codepoint has changed, update the fsb.
20382                          */
20383                         ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20384                             (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20385                         break;
20386                 }
20387                 INP_WUNLOCK(inp);
20388                 return (0);
20389 #endif
20390 #ifdef INET
20391         case IPPROTO_IP:
20392                 switch (sopt->sopt_name) {
20393                 case IP_TOS:
20394                         /*
20395                          * The DSCP codepoint has changed, update the fsb.
20396                          */
20397                         ip->ip_tos = rack->rc_inp->inp_ip_tos;
20398                         break;
20399                 case IP_TTL:
20400                         /*
20401                          * The TTL has changed, update the fsb.
20402                          */
20403                         ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20404                         break;
20405                 }
20406                 INP_WUNLOCK(inp);
20407                 return (0);
20408 #endif
20409         }
20410
20411         switch (sopt->sopt_name) {
20412         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
20413         /*  Pacing related ones */
20414         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
20415         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
20416         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
20417         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
20418         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
20419         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
20420         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
20421         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
20422         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
20423         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
20424         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
20425         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
20426         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
20427         case TCP_HDWR_RATE_CAP:                 /*  URL:hdwrcap boolean */
20428         case TCP_PACING_RATE_CAP:               /*  URL:cap  -- used by side-channel */
20429         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
20430        /* End pacing related */
20431         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
20432         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20433         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
20434         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
20435         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
20436         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
20437         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
20438         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
20439         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
20440         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
20441         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
20442         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
20443         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
20444         case TCP_NO_PRR:                        /*  URL:noprr */
20445         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
20446         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
20447         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
20448         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
20449         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
20450         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
20451         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
20452         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
20453         case TCP_RACK_PROFILE:                  /*  URL:profile */
20454         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
20455         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
20456         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
20457         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
20458         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
20459         case TCP_RACK_DSACK_OPT:                /*  URL:dsack */
20460         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
20461         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
20462         case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
20463         case TCP_RACK_ENABLE_HYSTART:           /*  URL:hystart */
20464                 break;
20465         default:
20466                 /* Filter off all unknown options to the base stack */
20467                 return (tcp_default_ctloutput(inp, sopt));
20468                 break;
20469         }
20470         INP_WUNLOCK(inp);
20471         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20472                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20473                 /*
20474                  * We truncate it down to 32 bits for the socket-option trace this
20475                  * means rates > 34Gbps won't show right, but thats probably ok.
20476                  */
20477                 optval = (uint32_t)loptval;
20478         } else {
20479                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20480                 /* Save it in 64 bit form too */
20481                 loptval = optval;
20482         }
20483         if (error)
20484                 return (error);
20485         INP_WLOCK(inp);
20486         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
20487                 INP_WUNLOCK(inp);
20488                 return (ECONNRESET);
20489         }
20490         if (tp->t_fb != &__tcp_rack) {
20491                 INP_WUNLOCK(inp);
20492                 return (ENOPROTOOPT);
20493         }
20494         if (rack->defer_options && (rack->gp_ready == 0) &&
20495             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20496             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20497             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20498             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20499                 /* Options are beind deferred */
20500                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20501                         INP_WUNLOCK(inp);
20502                         return (0);
20503                 } else {
20504                         /* No memory to defer, fail */
20505                         INP_WUNLOCK(inp);
20506                         return (ENOMEM);
20507                 }
20508         }
20509         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20510         INP_WUNLOCK(inp);
20511         return (error);
20512 }
20513
20514 static void
20515 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20516 {
20517
20518         INP_WLOCK_ASSERT(tp->t_inpcb);
20519         bzero(ti, sizeof(*ti));
20520
20521         ti->tcpi_state = tp->t_state;
20522         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20523                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20524         if (tp->t_flags & TF_SACK_PERMIT)
20525                 ti->tcpi_options |= TCPI_OPT_SACK;
20526         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20527                 ti->tcpi_options |= TCPI_OPT_WSCALE;
20528                 ti->tcpi_snd_wscale = tp->snd_scale;
20529                 ti->tcpi_rcv_wscale = tp->rcv_scale;
20530         }
20531         if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
20532                 ti->tcpi_options |= TCPI_OPT_ECN;
20533         if (tp->t_flags & TF_FASTOPEN)
20534                 ti->tcpi_options |= TCPI_OPT_TFO;
20535         /* still kept in ticks is t_rcvtime */
20536         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20537         /* Since we hold everything in precise useconds this is easy */
20538         ti->tcpi_rtt = tp->t_srtt;
20539         ti->tcpi_rttvar = tp->t_rttvar;
20540         ti->tcpi_rto = tp->t_rxtcur;
20541         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20542         ti->tcpi_snd_cwnd = tp->snd_cwnd;
20543         /*
20544          * FreeBSD-specific extension fields for tcp_info.
20545          */
20546         ti->tcpi_rcv_space = tp->rcv_wnd;
20547         ti->tcpi_rcv_nxt = tp->rcv_nxt;
20548         ti->tcpi_snd_wnd = tp->snd_wnd;
20549         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
20550         ti->tcpi_snd_nxt = tp->snd_nxt;
20551         ti->tcpi_snd_mss = tp->t_maxseg;
20552         ti->tcpi_rcv_mss = tp->t_maxseg;
20553         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20554         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20555         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20556 #ifdef NETFLIX_STATS
20557         ti->tcpi_total_tlp = tp->t_sndtlppack;
20558         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20559         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20560 #endif
20561 #ifdef TCP_OFFLOAD
20562         if (tp->t_flags & TF_TOE) {
20563                 ti->tcpi_options |= TCPI_OPT_TOE;
20564                 tcp_offload_tcp_info(tp, ti);
20565         }
20566 #endif
20567 }
20568
20569 static int
20570 rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
20571 {
20572         struct tcpcb *tp;
20573         struct tcp_rack *rack;
20574         int32_t error, optval;
20575         uint64_t val, loptval;
20576         struct  tcp_info ti;
20577         /*
20578          * Because all our options are either boolean or an int, we can just
20579          * pull everything into optval and then unlock and copy. If we ever
20580          * add a option that is not a int, then this will have quite an
20581          * impact to this routine.
20582          */
20583         error = 0;
20584         tp = intotcpcb(inp);
20585         rack = (struct tcp_rack *)tp->t_fb_ptr;
20586         if (rack == NULL) {
20587                 INP_WUNLOCK(inp);
20588                 return (EINVAL);
20589         }
20590         switch (sopt->sopt_name) {
20591         case TCP_INFO:
20592                 /* First get the info filled */
20593                 rack_fill_info(tp, &ti);
20594                 /* Fix up the rtt related fields if needed */
20595                 INP_WUNLOCK(inp);
20596                 error = sooptcopyout(sopt, &ti, sizeof ti);
20597                 return (error);
20598         /*
20599          * Beta is the congestion control value for NewReno that influences how
20600          * much of a backoff happens when loss is detected. It is normally set
20601          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20602          * when you exit recovery.
20603          */
20604         case TCP_RACK_PACING_BETA:
20605                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20606                         error = EINVAL;
20607                 else if (rack->rc_pacing_cc_set == 0)
20608                         optval = rack->r_ctl.rc_saved_beta.beta;
20609                 else {
20610                         /*
20611                          * Reach out into the CC data and report back what
20612                          * I have previously set. Yeah it looks hackish but
20613                          * we don't want to report the saved values.
20614                          */
20615                         if (tp->ccv->cc_data)
20616                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta;
20617                         else
20618                                 error = EINVAL;
20619                 }
20620                 break;
20621                 /*
20622                  * Beta_ecn is the congestion control value for NewReno that influences how
20623                  * much of a backoff happens when a ECN mark is detected. It is normally set
20624                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20625                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
20626                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
20627                  */
20628
20629         case TCP_RACK_PACING_BETA_ECN:
20630                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20631                         error = EINVAL;
20632                 else if (rack->rc_pacing_cc_set == 0)
20633                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20634                 else {
20635                         /*
20636                          * Reach out into the CC data and report back what
20637                          * I have previously set. Yeah it looks hackish but
20638                          * we don't want to report the saved values.
20639                          */
20640                         if (tp->ccv->cc_data)
20641                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
20642                         else
20643                                 error = EINVAL;
20644                 }
20645                 break;
20646         case TCP_RACK_DSACK_OPT:
20647                 optval = 0;
20648                 if (rack->rc_rack_tmr_std_based) {
20649                         optval |= 1;
20650                 }
20651                 if (rack->rc_rack_use_dsack) {
20652                         optval |= 2;
20653                 }
20654                 break;
20655         case TCP_RACK_ENABLE_HYSTART:
20656         {
20657                 if (tp->ccv->flags & CCF_HYSTART_ALLOWED) {
20658                         optval = RACK_HYSTART_ON;
20659                         if (tp->ccv->flags & CCF_HYSTART_CAN_SH_CWND)
20660                                 optval = RACK_HYSTART_ON_W_SC;
20661                         if (tp->ccv->flags & CCF_HYSTART_CONS_SSTH)
20662                                 optval = RACK_HYSTART_ON_W_SC_C;
20663                 } else {
20664                         optval = RACK_HYSTART_OFF;
20665                 }
20666         }
20667         break;
20668         case TCP_FAST_RSM_HACK:
20669                 optval = rack->fast_rsm_hack;
20670                 break;
20671         case TCP_DEFER_OPTIONS:
20672                 optval = rack->defer_options;
20673                 break;
20674         case TCP_RACK_MEASURE_CNT:
20675                 optval = rack->r_ctl.req_measurements;
20676                 break;
20677         case TCP_REC_ABC_VAL:
20678                 optval = rack->r_use_labc_for_rec;
20679                 break;
20680         case TCP_RACK_ABC_VAL:
20681                 optval = rack->rc_labc;
20682                 break;
20683         case TCP_HDWR_UP_ONLY:
20684                 optval= rack->r_up_only;
20685                 break;
20686         case TCP_PACING_RATE_CAP:
20687                 loptval = rack->r_ctl.bw_rate_cap;
20688                 break;
20689         case TCP_RACK_PROFILE:
20690                 /* You cannot retrieve a profile, its write only */
20691                 error = EINVAL;
20692                 break;
20693         case TCP_USE_CMP_ACKS:
20694                 optval = rack->r_use_cmp_ack;
20695                 break;
20696         case TCP_RACK_PACE_TO_FILL:
20697                 optval = rack->rc_pace_to_cwnd;
20698                 if (optval && rack->r_fill_less_agg)
20699                         optval++;
20700                 break;
20701         case TCP_RACK_NO_PUSH_AT_MAX:
20702                 optval = rack->r_ctl.rc_no_push_at_mrtt;
20703                 break;
20704         case TCP_SHARED_CWND_ENABLE:
20705                 optval = rack->rack_enable_scwnd;
20706                 break;
20707         case TCP_RACK_NONRXT_CFG_RATE:
20708                 optval = rack->rack_rec_nonrxt_use_cr;
20709                 break;
20710         case TCP_NO_PRR:
20711                 if (rack->rack_no_prr  == 1)
20712                         optval = 1;
20713                 else if (rack->no_prr_addback == 1)
20714                         optval = 2;
20715                 else
20716                         optval = 0;
20717                 break;
20718         case TCP_RACK_DO_DETECTION:
20719                 optval = rack->do_detection;
20720                 break;
20721         case TCP_RACK_MBUF_QUEUE:
20722                 /* Now do we use the LRO mbuf-queue feature */
20723                 optval = rack->r_mbuf_queue;
20724                 break;
20725         case TCP_TIMELY_DYN_ADJ:
20726                 optval = rack->rc_gp_dyn_mul;
20727                 break;
20728         case TCP_BBR_IWINTSO:
20729                 optval = rack->rc_init_win;
20730                 break;
20731         case TCP_RACK_TLP_REDUCE:
20732                 /* RACK TLP cwnd reduction (bool) */
20733                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20734                 break;
20735         case TCP_BBR_RACK_INIT_RATE:
20736                 val = rack->r_ctl.init_rate;
20737                 /* convert to kbits per sec */
20738                 val *= 8;
20739                 val /= 1000;
20740                 optval = (uint32_t)val;
20741                 break;
20742         case TCP_RACK_FORCE_MSEG:
20743                 optval = rack->rc_force_max_seg;
20744                 break;
20745         case TCP_RACK_PACE_MAX_SEG:
20746                 /* Max segments in a pace */
20747                 optval = rack->rc_user_set_max_segs;
20748                 break;
20749         case TCP_RACK_PACE_ALWAYS:
20750                 /* Use the always pace method */
20751                 optval = rack->rc_always_pace;
20752                 break;
20753         case TCP_RACK_PRR_SENDALOT:
20754                 /* Allow PRR to send more than one seg */
20755                 optval = rack->r_ctl.rc_prr_sendalot;
20756                 break;
20757         case TCP_RACK_MIN_TO:
20758                 /* Minimum time between rack t-o's in ms */
20759                 optval = rack->r_ctl.rc_min_to;
20760                 break;
20761         case TCP_RACK_EARLY_SEG:
20762                 /* If early recovery max segments */
20763                 optval = rack->r_ctl.rc_early_recovery_segs;
20764                 break;
20765         case TCP_RACK_REORD_THRESH:
20766                 /* RACK reorder threshold (shift amount) */
20767                 optval = rack->r_ctl.rc_reorder_shift;
20768                 break;
20769         case TCP_RACK_REORD_FADE:
20770                 /* Does reordering fade after ms time */
20771                 optval = rack->r_ctl.rc_reorder_fade;
20772                 break;
20773         case TCP_BBR_USE_RACK_RR:
20774                 /* Do we use the rack cheat for rxt */
20775                 optval = rack->use_rack_rr;
20776                 break;
20777         case TCP_RACK_RR_CONF:
20778                 optval = rack->r_rr_config;
20779                 break;
20780         case TCP_HDWR_RATE_CAP:
20781                 optval = rack->r_rack_hw_rate_caps;
20782                 break;
20783         case TCP_BBR_HDWR_PACE:
20784                 optval = rack->rack_hdw_pace_ena;
20785                 break;
20786         case TCP_RACK_TLP_THRESH:
20787                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
20788                 optval = rack->r_ctl.rc_tlp_threshold;
20789                 break;
20790         case TCP_RACK_PKT_DELAY:
20791                 /* RACK added ms i.e. rack-rtt + reord + N */
20792                 optval = rack->r_ctl.rc_pkt_delay;
20793                 break;
20794         case TCP_RACK_TLP_USE:
20795                 optval = rack->rack_tlp_threshold_use;
20796                 break;
20797         case TCP_RACK_PACE_RATE_CA:
20798                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20799                 break;
20800         case TCP_RACK_PACE_RATE_SS:
20801                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20802                 break;
20803         case TCP_RACK_PACE_RATE_REC:
20804                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20805                 break;
20806         case TCP_RACK_GP_INCREASE_SS:
20807                 optval = rack->r_ctl.rack_per_of_gp_ca;
20808                 break;
20809         case TCP_RACK_GP_INCREASE_CA:
20810                 optval = rack->r_ctl.rack_per_of_gp_ss;
20811                 break;
20812         case TCP_BBR_RACK_RTT_USE:
20813                 optval = rack->r_ctl.rc_rate_sample_method;
20814                 break;
20815         case TCP_DELACK:
20816                 optval = tp->t_delayed_ack;
20817                 break;
20818         case TCP_DATA_AFTER_CLOSE:
20819                 optval = rack->rc_allow_data_af_clo;
20820                 break;
20821         case TCP_SHARED_CWND_TIME_LIMIT:
20822                 optval = rack->r_limit_scw;
20823                 break;
20824         case TCP_RACK_TIMER_SLOP:
20825                 optval = rack->r_ctl.timer_slop;
20826                 break;
20827         default:
20828                 return (tcp_default_ctloutput(inp, sopt));
20829                 break;
20830         }
20831         INP_WUNLOCK(inp);
20832         if (error == 0) {
20833                 if (TCP_PACING_RATE_CAP)
20834                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
20835                 else
20836                         error = sooptcopyout(sopt, &optval, sizeof optval);
20837         }
20838         return (error);
20839 }
20840
20841 static int
20842 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt)
20843 {
20844         if (sopt->sopt_dir == SOPT_SET) {
20845                 return (rack_set_sockopt(inp, sopt));
20846         } else if (sopt->sopt_dir == SOPT_GET) {
20847                 return (rack_get_sockopt(inp, sopt));
20848         } else {
20849                 panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
20850         }
20851 }
20852
20853 static const char *rack_stack_names[] = {
20854         __XSTRING(STACKNAME),
20855 #ifdef STACKALIAS
20856         __XSTRING(STACKALIAS),
20857 #endif
20858 };
20859
20860 static int
20861 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20862 {
20863         memset(mem, 0, size);
20864         return (0);
20865 }
20866
20867 static void
20868 rack_dtor(void *mem, int32_t size, void *arg)
20869 {
20870
20871 }
20872
20873 static bool rack_mod_inited = false;
20874
20875 static int
20876 tcp_addrack(module_t mod, int32_t type, void *data)
20877 {
20878         int32_t err = 0;
20879         int num_stacks;
20880
20881         switch (type) {
20882         case MOD_LOAD:
20883                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20884                     sizeof(struct rack_sendmap),
20885                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20886
20887                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20888                     sizeof(struct tcp_rack),
20889                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20890
20891                 sysctl_ctx_init(&rack_sysctl_ctx);
20892                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20893                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20894                     OID_AUTO,
20895 #ifdef STACKALIAS
20896                     __XSTRING(STACKALIAS),
20897 #else
20898                     __XSTRING(STACKNAME),
20899 #endif
20900                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20901                     "");
20902                 if (rack_sysctl_root == NULL) {
20903                         printf("Failed to add sysctl node\n");
20904                         err = EFAULT;
20905                         goto free_uma;
20906                 }
20907                 rack_init_sysctls();
20908                 num_stacks = nitems(rack_stack_names);
20909                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20910                     rack_stack_names, &num_stacks);
20911                 if (err) {
20912                         printf("Failed to register %s stack name for "
20913                             "%s module\n", rack_stack_names[num_stacks],
20914                             __XSTRING(MODNAME));
20915                         sysctl_ctx_free(&rack_sysctl_ctx);
20916 free_uma:
20917                         uma_zdestroy(rack_zone);
20918                         uma_zdestroy(rack_pcb_zone);
20919                         rack_counter_destroy();
20920                         printf("Failed to register rack module -- err:%d\n", err);
20921                         return (err);
20922                 }
20923                 tcp_lro_reg_mbufq();
20924                 rack_mod_inited = true;
20925                 break;
20926         case MOD_QUIESCE:
20927                 err = deregister_tcp_functions(&__tcp_rack, true, false);
20928                 break;
20929         case MOD_UNLOAD:
20930                 err = deregister_tcp_functions(&__tcp_rack, false, true);
20931                 if (err == EBUSY)
20932                         break;
20933                 if (rack_mod_inited) {
20934                         uma_zdestroy(rack_zone);
20935                         uma_zdestroy(rack_pcb_zone);
20936                         sysctl_ctx_free(&rack_sysctl_ctx);
20937                         rack_counter_destroy();
20938                         rack_mod_inited = false;
20939                 }
20940                 tcp_lro_dereg_mbufq();
20941                 err = 0;
20942                 break;
20943         default:
20944                 return (EOPNOTSUPP);
20945         }
20946         return (err);
20947 }
20948
20949 static moduledata_t tcp_rack = {
20950         .name = __XSTRING(MODNAME),
20951         .evhand = tcp_addrack,
20952         .priv = 0
20953 };
20954
20955 MODULE_VERSION(MODNAME, 1);
20956 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20957 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);