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_hpts.h>
  97 #include <netinet/tcp_ratelimit.h>
  98 #include <netinet/tcp_accounting.h>
  99 #include <netinet/tcpip.h>
 100 #include <netinet/cc/cc.h>
 101 #include <netinet/cc/cc_newreno.h>
 102 #include <netinet/tcp_fastopen.h>
 103 #include <netinet/tcp_lro.h>
 104 #ifdef NETFLIX_SHARED_CWND
 105 #include <netinet/tcp_shared_cwnd.h>
 106 #endif
 107 #ifdef TCPDEBUG
 108 #include <netinet/tcp_debug.h>
 109 #endif                          /* TCPDEBUG */
 110 #ifdef TCP_OFFLOAD
 111 #include <netinet/tcp_offload.h>
 112 #endif
 113 #ifdef INET6
 114 #include <netinet6/tcp6_var.h>
 115 #endif
 116
 117 #include <netipsec/ipsec_support.h>
 118
 119 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 120 #include <netipsec/ipsec.h>
 121 #include <netipsec/ipsec6.h>
 122 #endif                          /* IPSEC */
 123
 124 #include <netinet/udp.h>
 125 #include <netinet/udp_var.h>
 126 #include <machine/in_cksum.h>
 127
 128 #ifdef MAC
 129 #include <security/mac/mac_framework.h>
 130 #endif
 131 #include "sack_filter.h"
 132 #include "tcp_rack.h"
 133 #include "rack_bbr_common.h"
 134
 135 uma_zone_t rack_zone;
 136 uma_zone_t rack_pcb_zone;
 137
 138 #ifndef TICKS2SBT
 139 #define TICKS2SBT(__t)  (tick_sbt * ((sbintime_t)(__t)))
 140 #endif
 141
 142 VNET_DECLARE(uint32_t, newreno_beta);
 143 VNET_DECLARE(uint32_t, newreno_beta_ecn);
 144 #define V_newreno_beta VNET(newreno_beta)
 145 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
 146
 147
 148 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
 149 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
 150
 151 struct sysctl_ctx_list rack_sysctl_ctx;
 152 struct sysctl_oid *rack_sysctl_root;
 153
 154 #define CUM_ACKED 1
 155 #define SACKED 2
 156
 157 /*
 158  * The RACK module incorporates a number of
 159  * TCP ideas that have been put out into the IETF
 160  * over the last few years:
 161  * - Matt Mathis's Rate Halving which slowly drops
 162  *    the congestion window so that the ack clock can
 163  *    be maintained during a recovery.
 164  * - Yuchung Cheng's RACK TCP (for which its named) that
 165  *    will stop us using the number of dup acks and instead
 166  *    use time as the gage of when we retransmit.
 167  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 168  *    of Dukkipati et.al.
 169  * RACK depends on SACK, so if an endpoint arrives that
 170  * cannot do SACK the state machine below will shuttle the
 171  * connection back to using the "default" TCP stack that is
 172  * in FreeBSD.
 173  *
 174  * To implement RACK the original TCP stack was first decomposed
 175  * into a functional state machine with individual states
 176  * for each of the possible TCP connection states. The do_segement
 177  * functions role in life is to mandate the connection supports SACK
 178  * initially and then assure that the RACK state matches the conenction
 179  * state before calling the states do_segment function. Each
 180  * state is simplified due to the fact that the original do_segment
 181  * has been decomposed and we *know* what state we are in (no
 182  * switches on the state) and all tests for SACK are gone. This
 183  * greatly simplifies what each state does.
 184  *
 185  * TCP output is also over-written with a new version since it
 186  * must maintain the new rack scoreboard.
 187  *
 188  */
 189 static int32_t rack_tlp_thresh = 1;
 190 static int32_t rack_tlp_limit = 2;      /* No more than 2 TLPs w-out new data */
 191 static int32_t rack_tlp_use_greater = 1;
 192 static int32_t rack_reorder_thresh = 2;
 193 static int32_t rack_reorder_fade = 60000000;    /* 0 - never fade, def 60,000,000
 194                                                  * - 60 seconds */
 195 static uint8_t rack_req_measurements = 1;
 196 /* Attack threshold detections */
 197 static uint32_t rack_highest_sack_thresh_seen = 0;
 198 static uint32_t rack_highest_move_thresh_seen = 0;
 199 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
 200 static int32_t rack_hw_pace_extra_slots = 2;    /* 2 extra MSS time betweens */
 201 static int32_t rack_hw_rate_caps = 1; /* 1; */
 202 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
 203 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
 204 static int32_t rack_hw_up_only = 1;
 205 static int32_t rack_stats_gets_ms_rtt = 1;
 206 static int32_t rack_prr_addbackmax = 2;
 207
 208 static int32_t rack_pkt_delay = 1000;
 209 static int32_t rack_send_a_lot_in_prr = 1;
 210 static int32_t rack_min_to = 1000;      /* Number of microsecond  min timeout */
 211 static int32_t rack_verbose_logging = 0;
 212 static int32_t rack_ignore_data_after_close = 1;
 213 static int32_t rack_enable_shared_cwnd = 1;
 214 static int32_t rack_use_cmp_acks = 1;
 215 static int32_t rack_use_fsb = 1;
 216 static int32_t rack_use_rfo = 1;
 217 static int32_t rack_use_rsm_rfo = 1;
 218 static int32_t rack_max_abc_post_recovery = 2;
 219 static int32_t rack_client_low_buf = 0;
 220 #ifdef TCP_ACCOUNTING
 221 static int32_t rack_tcp_accounting = 0;
 222 #endif
 223 static int32_t rack_limits_scwnd = 1;
 224 static int32_t rack_enable_mqueue_for_nonpaced = 0;
 225 static int32_t rack_disable_prr = 0;
 226 static int32_t use_rack_rr = 1;
 227 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
 228 static int32_t rack_persist_min = 250000;       /* 250usec */
 229 static int32_t rack_persist_max = 2000000;      /* 2 Second in usec's */
 230 static int32_t rack_sack_not_required = 1;      /* set to one to allow non-sack to use rack */
 231 static int32_t rack_default_init_window = 0;    /* Use system default */
 232 static int32_t rack_limit_time_with_srtt = 0;
 233 static int32_t rack_autosndbuf_inc = 20;        /* In percentage form */
 234 static int32_t rack_enobuf_hw_boost_mult = 2;   /* How many times the hw rate we boost slot using time_between */
 235 static int32_t rack_enobuf_hw_max = 12000;      /* 12 ms in usecs */
 236 static int32_t rack_enobuf_hw_min = 10000;      /* 10 ms in usecs */
 237 static int32_t rack_hw_rwnd_factor = 2;         /* How many max_segs the rwnd must be before we hold off sending */
 238 /*
 239  * Currently regular tcp has a rto_min of 30ms
 240  * the backoff goes 12 times so that ends up
 241  * being a total of 122.850 seconds before a
 242  * connection is killed.
 243  */
 244 static uint32_t rack_def_data_window = 20;
 245 static uint32_t rack_goal_bdp = 2;
 246 static uint32_t rack_min_srtts = 1;
 247 static uint32_t rack_min_measure_usec = 0;
 248 static int32_t rack_tlp_min = 10000;    /* 10ms */
 249 static int32_t rack_rto_min = 30000;    /* 30,000 usec same as main freebsd */
 250 static int32_t rack_rto_max = 4000000;  /* 4 seconds in usec's */
 251 static const int32_t rack_free_cache = 2;
 252 static int32_t rack_hptsi_segments = 40;
 253 static int32_t rack_rate_sample_method = USE_RTT_LOW;
 254 static int32_t rack_pace_every_seg = 0;
 255 static int32_t rack_delayed_ack_time = 40000;   /* 40ms in usecs */
 256 static int32_t rack_slot_reduction = 4;
 257 static int32_t rack_wma_divisor = 8;            /* For WMA calculation */
 258 static int32_t rack_cwnd_block_ends_measure = 0;
 259 static int32_t rack_rwnd_block_ends_measure = 0;
 260 static int32_t rack_def_profile = 0;
 261
 262 static int32_t rack_lower_cwnd_at_tlp = 0;
 263 static int32_t rack_limited_retran = 0;
 264 static int32_t rack_always_send_oldest = 0;
 265 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
 266
 267 static uint16_t rack_per_of_gp_ss = 250;        /* 250 % slow-start */
 268 static uint16_t rack_per_of_gp_ca = 200;        /* 200 % congestion-avoidance */
 269 static uint16_t rack_per_of_gp_rec = 200;       /* 200 % of bw */
 270
 271 /* Probertt */
 272 static uint16_t rack_per_of_gp_probertt = 60;   /* 60% of bw */
 273 static uint16_t rack_per_of_gp_lowthresh = 40;  /* 40% is bottom */
 274 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
 275 static uint16_t rack_atexit_prtt_hbp = 130;     /* Clamp to 130% on exit prtt if highly buffered path */
 276 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
 277
 278 static uint32_t rack_max_drain_wait = 2;        /* How man gp srtt's before we give up draining */
 279 static uint32_t rack_must_drain = 1;            /* How many GP srtt's we *must* wait */
 280 static uint32_t rack_probertt_use_min_rtt_entry = 1;    /* Use the min to calculate the goal else gp_srtt */
 281 static uint32_t rack_probertt_use_min_rtt_exit = 0;
 282 static uint32_t rack_probe_rtt_sets_cwnd = 0;
 283 static uint32_t rack_probe_rtt_safety_val = 2000000;    /* No more than 2 sec in probe-rtt */
 284 static uint32_t rack_time_between_probertt = 9600000;   /* 9.6 sec in usecs */
 285 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;       /* How many srtt periods does probe-rtt last top fraction */
 286 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;       /* How many srtt periods does probe-rtt last bottom fraction */
 287 static uint32_t rack_min_probertt_hold = 40000;         /* Equal to delayed ack time */
 288 static uint32_t rack_probertt_filter_life = 10000000;
 289 static uint32_t rack_probertt_lower_within = 10;
 290 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds)  to count as a lowering */
 291 static int32_t rack_pace_one_seg = 0;           /* Shall we pace for less than 1.4Meg 1MSS at a time */
 292 static int32_t rack_probertt_clear_is = 1;
 293 static int32_t rack_max_drain_hbp = 1;          /* Extra drain times gpsrtt for highly buffered paths */
 294 static int32_t rack_hbp_thresh = 3;             /* what is the divisor max_rtt/min_rtt to decided a hbp */
 295
 296 /* Part of pacing */
 297 static int32_t rack_max_per_above = 30;         /* When we go to increment stop if above 100+this% */
 298
 299 /* Timely information */
 300 /* Combine these two gives the range of 'no change' to bw */
 301 /* ie the up/down provide the upper and lower bound */
 302 static int32_t rack_gp_per_bw_mul_up = 2;       /* 2% */
 303 static int32_t rack_gp_per_bw_mul_down = 4;     /* 4% */
 304 static int32_t rack_gp_rtt_maxmul = 3;          /* 3 x maxmin */
 305 static int32_t rack_gp_rtt_minmul = 1;          /* minrtt + (minrtt/mindiv) is lower rtt */
 306 static int32_t rack_gp_rtt_mindiv = 4;          /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
 307 static int32_t rack_gp_decrease_per = 20;       /* 20% decrease in multipler */
 308 static int32_t rack_gp_increase_per = 2;        /* 2% increase in multipler */
 309 static int32_t rack_per_lower_bound = 50;       /* Don't allow to drop below this multiplier */
 310 static int32_t rack_per_upper_bound_ss = 0;     /* Don't allow SS to grow above this */
 311 static int32_t rack_per_upper_bound_ca = 0;     /* Don't allow CA to grow above this */
 312 static int32_t rack_do_dyn_mul = 0;             /* Are the rack gp multipliers dynamic */
 313 static int32_t rack_gp_no_rec_chg = 1;          /* Prohibit recovery from reducing it's multiplier */
 314 static int32_t rack_timely_dec_clear = 6;       /* Do we clear decrement count at a value (6)? */
 315 static int32_t rack_timely_max_push_rise = 3;   /* One round of pushing */
 316 static int32_t rack_timely_max_push_drop = 3;   /* Three round of pushing */
 317 static int32_t rack_timely_min_segs = 4;        /* 4 segment minimum */
 318 static int32_t rack_use_max_for_nobackoff = 0;
 319 static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */
 320 static int32_t rack_timely_no_stopping = 0;
 321 static int32_t rack_down_raise_thresh = 100;
 322 static int32_t rack_req_segs = 1;
 323 static uint64_t rack_bw_rate_cap = 0;
 324
 325 /* Weird delayed ack mode */
 326 static int32_t rack_use_imac_dack = 0;
 327 /* Rack specific counters */
 328 counter_u64_t rack_badfr;
 329 counter_u64_t rack_badfr_bytes;
 330 counter_u64_t rack_rtm_prr_retran;
 331 counter_u64_t rack_rtm_prr_newdata;
 332 counter_u64_t rack_timestamp_mismatch;
 333 counter_u64_t rack_reorder_seen;
 334 counter_u64_t rack_paced_segments;
 335 counter_u64_t rack_unpaced_segments;
 336 counter_u64_t rack_calc_zero;
 337 counter_u64_t rack_calc_nonzero;
 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_per_timer_hole;
 342 counter_u64_t rack_large_ackcmp;
 343 counter_u64_t rack_small_ackcmp;
 344 #ifdef INVARIANTS
 345 counter_u64_t rack_adjust_map_bw;
 346 #endif
 347 /* Tail loss probe counters */
 348 counter_u64_t rack_tlp_tot;
 349 counter_u64_t rack_tlp_newdata;
 350 counter_u64_t rack_tlp_retran;
 351 counter_u64_t rack_tlp_retran_bytes;
 352 counter_u64_t rack_tlp_retran_fail;
 353 counter_u64_t rack_to_tot;
 354 counter_u64_t rack_to_arm_rack;
 355 counter_u64_t rack_to_arm_tlp;
 356 counter_u64_t rack_hot_alloc;
 357 counter_u64_t rack_to_alloc;
 358 counter_u64_t rack_to_alloc_hard;
 359 counter_u64_t rack_to_alloc_emerg;
 360 counter_u64_t rack_to_alloc_limited;
 361 counter_u64_t rack_alloc_limited_conns;
 362 counter_u64_t rack_split_limited;
 363
 364 #define MAX_NUM_OF_CNTS 13
 365 counter_u64_t rack_proc_comp_ack[MAX_NUM_OF_CNTS];
 366 counter_u64_t rack_multi_single_eq;
 367 counter_u64_t rack_proc_non_comp_ack;
 368
 369 counter_u64_t rack_fto_send;
 370 counter_u64_t rack_fto_rsm_send;
 371 counter_u64_t rack_nfto_resend;
 372 counter_u64_t rack_non_fto_send;
 373 counter_u64_t rack_extended_rfo;
 374
 375 counter_u64_t rack_sack_proc_all;
 376 counter_u64_t rack_sack_proc_short;
 377 counter_u64_t rack_sack_proc_restart;
 378 counter_u64_t rack_sack_attacks_detected;
 379 counter_u64_t rack_sack_attacks_reversed;
 380 counter_u64_t rack_sack_used_next_merge;
 381 counter_u64_t rack_sack_splits;
 382 counter_u64_t rack_sack_used_prev_merge;
 383 counter_u64_t rack_sack_skipped_acked;
 384 counter_u64_t rack_ack_total;
 385 counter_u64_t rack_express_sack;
 386 counter_u64_t rack_sack_total;
 387 counter_u64_t rack_move_none;
 388 counter_u64_t rack_move_some;
 389
 390 counter_u64_t rack_used_tlpmethod;
 391 counter_u64_t rack_used_tlpmethod2;
 392 counter_u64_t rack_enter_tlp_calc;
 393 counter_u64_t rack_input_idle_reduces;
 394 counter_u64_t rack_collapsed_win;
 395 counter_u64_t rack_tlp_does_nada;
 396 counter_u64_t rack_try_scwnd;
 397 counter_u64_t rack_hw_pace_init_fail;
 398 counter_u64_t rack_hw_pace_lost;
 399 counter_u64_t rack_sbsndptr_right;
 400 counter_u64_t rack_sbsndptr_wrong;
 401
 402 /* Temp CPU counters */
 403 counter_u64_t rack_find_high;
 404
 405 counter_u64_t rack_progress_drops;
 406 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
 407 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
 408
 409
 410 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
 411
 412 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {  \
 413         (tv) = (value) + slop;   \
 414         if ((u_long)(tv) < (u_long)(tvmin)) \
 415                 (tv) = (tvmin); \
 416         if ((u_long)(tv) > (u_long)(tvmax)) \
 417                 (tv) = (tvmax); \
 418 } while (0)
 419
 420 static void
 421 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
 422
 423 static int
 424 rack_process_ack(struct mbuf *m, struct tcphdr *th,
 425     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
 426     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
 427 static int
 428 rack_process_data(struct mbuf *m, struct tcphdr *th,
 429     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 430     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
 431 static void
 432 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
 433    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
 434 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
 435 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
 436     uint8_t limit_type);
 437 static struct rack_sendmap *
 438 rack_check_recovery_mode(struct tcpcb *tp,
 439     uint32_t tsused);
 440 static void
 441 rack_cong_signal(struct tcpcb *tp,
 442                  uint32_t type, uint32_t ack);
 443 static void rack_counter_destroy(void);
 444 static int
 445 rack_ctloutput(struct socket *so, struct sockopt *sopt,
 446     struct inpcb *inp, struct tcpcb *tp);
 447 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
 448 static void
 449 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
 450 static void
 451 rack_do_segment(struct mbuf *m, struct tcphdr *th,
 452     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 453     uint8_t iptos);
 454 static void rack_dtor(void *mem, int32_t size, void *arg);
 455 static void
 456 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
 457     uint32_t flex1, uint32_t flex2,
 458     uint32_t flex3, uint32_t flex4,
 459     uint32_t flex5, uint32_t flex6,
 460     uint16_t flex7, uint8_t mod);
 461 static void
 462 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
 463    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm);
 464 static struct rack_sendmap *
 465 rack_find_high_nonack(struct tcp_rack *rack,
 466     struct rack_sendmap *rsm);
 467 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
 468 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
 469 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
 470 static int
 471 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
 472     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
 473 static void
 474 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
 475                             tcp_seq th_ack, int line);
 476 static uint32_t
 477 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
 478 static int32_t rack_handoff_ok(struct tcpcb *tp);
 479 static int32_t rack_init(struct tcpcb *tp);
 480 static void rack_init_sysctls(void);
 481 static void
 482 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
 483     struct tcphdr *th, int entered_rec, int dup_ack_struck);
 484 static void
 485 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
 486     uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t ts,
 487     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
 488
 489 static void
 490 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
 491     struct rack_sendmap *rsm);
 492 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
 493 static int32_t rack_output(struct tcpcb *tp);
 494
 495 static uint32_t
 496 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
 497     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
 498     uint32_t cts, int *moved_two);
 499 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
 500 static void rack_remxt_tmr(struct tcpcb *tp);
 501 static int
 502 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
 503     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
 504 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
 505 static int32_t rack_stopall(struct tcpcb *tp);
 506 static void
 507 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
 508     uint32_t delta);
 509 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
 510 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
 511 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
 512 static uint32_t
 513 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
 514     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
 515 static void
 516 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
 517     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
 518 static int
 519 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
 520     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
 521 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
 522 static int
 523 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
 524     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 525     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 526 static int
 527 rack_do_closing(struct mbuf *m, struct tcphdr *th,
 528     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 529     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 530 static int
 531 rack_do_established(struct mbuf *m, struct tcphdr *th,
 532     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 533     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 534 static int
 535 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
 536     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 537     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
 538 static int
 539 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
 540     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 541     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 542 static int
 543 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
 544     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 545     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 546 static int
 547 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
 548     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 549     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 550 static int
 551 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
 552     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 553     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 554 static int
 555 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
 556     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 557     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 558 struct rack_sendmap *
 559 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
 560     uint32_t tsused);
 561 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
 562     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
 563 static void
 564      tcp_rack_partialack(struct tcpcb *tp);
 565 static int
 566 rack_set_profile(struct tcp_rack *rack, int prof);
 567 static void
 568 rack_apply_deferred_options(struct tcp_rack *rack);
 569
 570 int32_t rack_clear_counter=0;
 571
 572 static void
 573 rack_set_cc_pacing(struct tcp_rack *rack)
 574 {
 575         struct sockopt sopt;
 576         struct cc_newreno_opts opt;
 577         struct newreno old, *ptr;
 578         struct tcpcb *tp;
 579         int error;
 580
 581         if (rack->rc_pacing_cc_set)
 582                 return;
 583
 584         tp = rack->rc_tp;
 585         if (tp->cc_algo == NULL) {
 586                 /* Tcb is leaving */
 587                 printf("No cc algorithm?\n");
 588                 return;
 589         }
 590         rack->rc_pacing_cc_set = 1;
 591         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 592                 /* Not new-reno we can't play games with beta! */
 593                 goto out;
 594         }
 595         ptr = ((struct newreno *)tp->ccv->cc_data);
 596         if (CC_ALGO(tp)->ctl_output == NULL)  {
 597                 /* Huh, why does new_reno no longer have a set function? */
 598                 printf("no ctl_output for algo:%s\n", tp->cc_algo->name);
 599                 goto out;
 600         }
 601         if (ptr == NULL) {
 602                 /* Just the default values */
 603                 old.beta = V_newreno_beta_ecn;
 604                 old.beta_ecn = V_newreno_beta_ecn;
 605                 old.newreno_flags = 0;
 606         } else {
 607                 old.beta = ptr->beta;
 608                 old.beta_ecn = ptr->beta_ecn;
 609                 old.newreno_flags = ptr->newreno_flags;
 610         }
 611         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 612         sopt.sopt_dir = SOPT_SET;
 613         opt.name = CC_NEWRENO_BETA;
 614         opt.val = rack->r_ctl.rc_saved_beta.beta;
 615         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 616         if (error)  {
 617                 printf("Error returned by ctl_output %d\n", error);
 618                 goto out;
 619         }
 620         /*
 621          * Hack alert we need to set in our newreno_flags
 622          * so that Abe behavior is also applied.
 623          */
 624         ((struct newreno *)tp->ccv->cc_data)->newreno_flags = CC_NEWRENO_BETA_ECN;
 625         opt.name = CC_NEWRENO_BETA_ECN;
 626         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 627         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 628         if (error) {
 629                 printf("Error returned by ctl_output %d\n", error);
 630                 goto out;
 631         }
 632         /* Save off the original values for restoral */
 633         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 634 out:
 635         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 636                 union tcp_log_stackspecific log;
 637                 struct timeval tv;
 638
 639                 ptr = ((struct newreno *)tp->ccv->cc_data);
 640                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 641                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 642                 if (ptr) {
 643                         log.u_bbr.flex1 = ptr->beta;
 644                         log.u_bbr.flex2 = ptr->beta_ecn;
 645                         log.u_bbr.flex3 = ptr->newreno_flags;
 646                 }
 647                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 648                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 649                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 650                 log.u_bbr.flex7 = rack->gp_ready;
 651                 log.u_bbr.flex7 <<= 1;
 652                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 653                 log.u_bbr.flex7 <<= 1;
 654                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 655                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 656                 log.u_bbr.flex8 = 3;
 657                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 658                                0, &log, false, NULL, NULL, 0, &tv);
 659         }
 660 }
 661
 662 static void
 663 rack_undo_cc_pacing(struct tcp_rack *rack)
 664 {
 665         struct newreno old, *ptr;
 666         struct tcpcb *tp;
 667
 668         if (rack->rc_pacing_cc_set == 0)
 669                 return;
 670         tp = rack->rc_tp;
 671         rack->rc_pacing_cc_set = 0;
 672         if (tp->cc_algo == NULL)
 673                 /* Tcb is leaving */
 674                 return;
 675         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 676                 /* Not new-reno nothing to do! */
 677                 return;
 678         }
 679         ptr = ((struct newreno *)tp->ccv->cc_data);
 680         if (ptr == NULL) {
 681                 /*
 682                  * This happens at rack_fini() if the
 683                  * cc module gets freed on us. In that
 684                  * case we loose our "new" settings but
 685                  * thats ok, since the tcb is going away anyway.
 686                  */
 687                 return;
 688         }
 689         /* Grab out our set values */
 690         memcpy(&old, ptr, sizeof(struct newreno));
 691         /* Copy back in the original values */
 692         memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
 693         /* Now save back the values we had set in (for when pacing is restored) */
 694         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 695         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 696                 union tcp_log_stackspecific log;
 697                 struct timeval tv;
 698
 699                 ptr = ((struct newreno *)tp->ccv->cc_data);
 700                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 701                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 702                 log.u_bbr.flex1 = ptr->beta;
 703                 log.u_bbr.flex2 = ptr->beta_ecn;
 704                 log.u_bbr.flex3 = ptr->newreno_flags;
 705                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 706                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 707                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 708                 log.u_bbr.flex7 = rack->gp_ready;
 709                 log.u_bbr.flex7 <<= 1;
 710                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 711                 log.u_bbr.flex7 <<= 1;
 712                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 713                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 714                 log.u_bbr.flex8 = 4;
 715                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
 716                                0, &log, false, NULL, NULL, 0, &tv);
 717         }
 718 }
 719
 720 #ifdef NETFLIX_PEAKRATE
 721 static inline void
 722 rack_update_peakrate_thr(struct tcpcb *tp)
 723 {
 724         /* Keep in mind that t_maxpeakrate is in B/s. */
 725         uint64_t peak;
 726         peak = uqmax((tp->t_maxseg * 2),
 727                      (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
 728         tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
 729 }
 730 #endif
 731
 732 static int
 733 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 734 {
 735         uint32_t stat;
 736         int32_t error;
 737         int i;
 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_badfr);
 751                 counter_u64_zero(rack_badfr_bytes);
 752                 counter_u64_zero(rack_rtm_prr_retran);
 753                 counter_u64_zero(rack_rtm_prr_newdata);
 754                 counter_u64_zero(rack_timestamp_mismatch);
 755                 counter_u64_zero(rack_reorder_seen);
 756                 counter_u64_zero(rack_tlp_tot);
 757                 counter_u64_zero(rack_tlp_newdata);
 758                 counter_u64_zero(rack_tlp_retran);
 759                 counter_u64_zero(rack_tlp_retran_bytes);
 760                 counter_u64_zero(rack_tlp_retran_fail);
 761                 counter_u64_zero(rack_to_tot);
 762                 counter_u64_zero(rack_to_arm_rack);
 763                 counter_u64_zero(rack_to_arm_tlp);
 764                 counter_u64_zero(rack_paced_segments);
 765                 counter_u64_zero(rack_calc_zero);
 766                 counter_u64_zero(rack_calc_nonzero);
 767                 counter_u64_zero(rack_unpaced_segments);
 768                 counter_u64_zero(rack_saw_enobuf);
 769                 counter_u64_zero(rack_saw_enobuf_hw);
 770                 counter_u64_zero(rack_saw_enetunreach);
 771                 counter_u64_zero(rack_per_timer_hole);
 772                 counter_u64_zero(rack_large_ackcmp);
 773                 counter_u64_zero(rack_small_ackcmp);
 774 #ifdef INVARIANTS
 775                 counter_u64_zero(rack_adjust_map_bw);
 776 #endif
 777                 counter_u64_zero(rack_to_alloc_hard);
 778                 counter_u64_zero(rack_to_alloc_emerg);
 779                 counter_u64_zero(rack_sack_proc_all);
 780                 counter_u64_zero(rack_fto_send);
 781                 counter_u64_zero(rack_fto_rsm_send);
 782                 counter_u64_zero(rack_extended_rfo);
 783                 counter_u64_zero(rack_hw_pace_init_fail);
 784                 counter_u64_zero(rack_hw_pace_lost);
 785                 counter_u64_zero(rack_sbsndptr_wrong);
 786                 counter_u64_zero(rack_sbsndptr_right);
 787                 counter_u64_zero(rack_non_fto_send);
 788                 counter_u64_zero(rack_nfto_resend);
 789                 counter_u64_zero(rack_sack_proc_short);
 790                 counter_u64_zero(rack_sack_proc_restart);
 791                 counter_u64_zero(rack_to_alloc);
 792                 counter_u64_zero(rack_to_alloc_limited);
 793                 counter_u64_zero(rack_alloc_limited_conns);
 794                 counter_u64_zero(rack_split_limited);
 795                 for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
 796                         counter_u64_zero(rack_proc_comp_ack[i]);
 797                 }
 798                 counter_u64_zero(rack_multi_single_eq);
 799                 counter_u64_zero(rack_proc_non_comp_ack);
 800                 counter_u64_zero(rack_find_high);
 801                 counter_u64_zero(rack_sack_attacks_detected);
 802                 counter_u64_zero(rack_sack_attacks_reversed);
 803                 counter_u64_zero(rack_sack_used_next_merge);
 804                 counter_u64_zero(rack_sack_used_prev_merge);
 805                 counter_u64_zero(rack_sack_splits);
 806                 counter_u64_zero(rack_sack_skipped_acked);
 807                 counter_u64_zero(rack_ack_total);
 808                 counter_u64_zero(rack_express_sack);
 809                 counter_u64_zero(rack_sack_total);
 810                 counter_u64_zero(rack_move_none);
 811                 counter_u64_zero(rack_move_some);
 812                 counter_u64_zero(rack_used_tlpmethod);
 813                 counter_u64_zero(rack_used_tlpmethod2);
 814                 counter_u64_zero(rack_enter_tlp_calc);
 815                 counter_u64_zero(rack_progress_drops);
 816                 counter_u64_zero(rack_tlp_does_nada);
 817                 counter_u64_zero(rack_try_scwnd);
 818                 counter_u64_zero(rack_collapsed_win);
 819         }
 820         rack_clear_counter = 0;
 821         return (0);
 822 }
 823
 824 static void
 825 rack_init_sysctls(void)
 826 {
 827         int i;
 828         struct sysctl_oid *rack_counters;
 829         struct sysctl_oid *rack_attack;
 830         struct sysctl_oid *rack_pacing;
 831         struct sysctl_oid *rack_timely;
 832         struct sysctl_oid *rack_timers;
 833         struct sysctl_oid *rack_tlp;
 834         struct sysctl_oid *rack_misc;
 835         struct sysctl_oid *rack_measure;
 836         struct sysctl_oid *rack_probertt;
 837         struct sysctl_oid *rack_hw_pacing;
 838
 839         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 840             SYSCTL_CHILDREN(rack_sysctl_root),
 841             OID_AUTO,
 842             "sack_attack",
 843             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 844             "Rack Sack Attack Counters and Controls");
 845         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 846             SYSCTL_CHILDREN(rack_sysctl_root),
 847             OID_AUTO,
 848             "stats",
 849             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 850             "Rack Counters");
 851         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 852             SYSCTL_CHILDREN(rack_sysctl_root),
 853             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 854             &rack_rate_sample_method , USE_RTT_LOW,
 855             "What method should we use for rate sampling 0=high, 1=low ");
 856         /* Probe rtt related controls */
 857         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 858             SYSCTL_CHILDREN(rack_sysctl_root),
 859             OID_AUTO,
 860             "probertt",
 861             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 862             "ProbeRTT related Controls");
 863         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 864             SYSCTL_CHILDREN(rack_probertt),
 865             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 866             &rack_atexit_prtt_hbp, 130,
 867             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 868         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 869             SYSCTL_CHILDREN(rack_probertt),
 870             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 871             &rack_atexit_prtt, 130,
 872             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 873         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 874             SYSCTL_CHILDREN(rack_probertt),
 875             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 876             &rack_per_of_gp_probertt, 60,
 877             "What percentage of goodput do we pace at in probertt");
 878         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 879             SYSCTL_CHILDREN(rack_probertt),
 880             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 881             &rack_per_of_gp_probertt_reduce, 10,
 882             "What percentage of goodput do we reduce every gp_srtt");
 883         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 884             SYSCTL_CHILDREN(rack_probertt),
 885             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 886             &rack_per_of_gp_lowthresh, 40,
 887             "What percentage of goodput do we allow the multiplier to fall to");
 888         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 889             SYSCTL_CHILDREN(rack_probertt),
 890             OID_AUTO, "time_between", CTLFLAG_RW,
 891             & rack_time_between_probertt, 96000000,
 892             "How many useconds between the lowest rtt falling must past before we enter probertt");
 893         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 894             SYSCTL_CHILDREN(rack_probertt),
 895             OID_AUTO, "safety", CTLFLAG_RW,
 896             &rack_probe_rtt_safety_val, 2000000,
 897             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 898         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 899             SYSCTL_CHILDREN(rack_probertt),
 900             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 901             &rack_probe_rtt_sets_cwnd, 0,
 902             "Do we set the cwnd too (if always_lower is on)");
 903         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 904             SYSCTL_CHILDREN(rack_probertt),
 905             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 906             &rack_max_drain_wait, 2,
 907             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 908         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 909             SYSCTL_CHILDREN(rack_probertt),
 910             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 911             &rack_must_drain, 1,
 912             "We must drain this many gp_srtt's waiting for flight to reach goal");
 913         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 914             SYSCTL_CHILDREN(rack_probertt),
 915             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 916             &rack_probertt_use_min_rtt_entry, 1,
 917             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 918         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 919             SYSCTL_CHILDREN(rack_probertt),
 920             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 921             &rack_probertt_use_min_rtt_exit, 0,
 922             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 923         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 924             SYSCTL_CHILDREN(rack_probertt),
 925             OID_AUTO, "length_div", CTLFLAG_RW,
 926             &rack_probertt_gpsrtt_cnt_div, 0,
 927             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 928         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 929             SYSCTL_CHILDREN(rack_probertt),
 930             OID_AUTO, "length_mul", CTLFLAG_RW,
 931             &rack_probertt_gpsrtt_cnt_mul, 0,
 932             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 933         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 934             SYSCTL_CHILDREN(rack_probertt),
 935             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 936             &rack_min_probertt_hold, 200000,
 937             "What is the minimum time we hold probertt at target");
 938         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 939             SYSCTL_CHILDREN(rack_probertt),
 940             OID_AUTO, "filter_life", CTLFLAG_RW,
 941             &rack_probertt_filter_life, 10000000,
 942             "What is the time for the filters life in useconds");
 943         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 944             SYSCTL_CHILDREN(rack_probertt),
 945             OID_AUTO, "lower_within", CTLFLAG_RW,
 946             &rack_probertt_lower_within, 10,
 947             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 948         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 949             SYSCTL_CHILDREN(rack_probertt),
 950             OID_AUTO, "must_move", CTLFLAG_RW,
 951             &rack_min_rtt_movement, 250,
 952             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 953         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 954             SYSCTL_CHILDREN(rack_probertt),
 955             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 956             &rack_probertt_clear_is, 1,
 957             "Do we clear I/S counts on exiting probe-rtt");
 958         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 959             SYSCTL_CHILDREN(rack_probertt),
 960             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 961             &rack_max_drain_hbp, 1,
 962             "How many extra drain gpsrtt's do we get in highly buffered paths");
 963         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 964             SYSCTL_CHILDREN(rack_probertt),
 965             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 966             &rack_hbp_thresh, 3,
 967             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 968         /* Pacing related sysctls */
 969         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 970             SYSCTL_CHILDREN(rack_sysctl_root),
 971             OID_AUTO,
 972             "pacing",
 973             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 974             "Pacing related Controls");
 975         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 976             SYSCTL_CHILDREN(rack_pacing),
 977             OID_AUTO, "max_pace_over", CTLFLAG_RW,
 978             &rack_max_per_above, 30,
 979             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
 980         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 981             SYSCTL_CHILDREN(rack_pacing),
 982             OID_AUTO, "pace_to_one", CTLFLAG_RW,
 983             &rack_pace_one_seg, 0,
 984             "Do we allow low b/w pacing of 1MSS instead of two");
 985         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 986             SYSCTL_CHILDREN(rack_pacing),
 987             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
 988             &rack_limit_time_with_srtt, 0,
 989             "Do we limit pacing time based on srtt");
 990         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 991             SYSCTL_CHILDREN(rack_pacing),
 992             OID_AUTO, "init_win", CTLFLAG_RW,
 993             &rack_default_init_window, 0,
 994             "Do we have a rack initial window 0 = system default");
 995         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 996             SYSCTL_CHILDREN(rack_pacing),
 997             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
 998             &rack_per_of_gp_ss, 250,
 999             "If non zero, what percentage of goodput to pace at in slow start");
1000         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1001             SYSCTL_CHILDREN(rack_pacing),
1002             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1003             &rack_per_of_gp_ca, 150,
1004             "If non zero, what percentage of goodput to pace at in congestion avoidance");
1005         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1006             SYSCTL_CHILDREN(rack_pacing),
1007             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1008             &rack_per_of_gp_rec, 200,
1009             "If non zero, what percentage of goodput to pace at in recovery");
1010         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1011             SYSCTL_CHILDREN(rack_pacing),
1012             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1013             &rack_hptsi_segments, 40,
1014             "What size is the max for TSO segments in pacing and burst mitigation");
1015         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1016             SYSCTL_CHILDREN(rack_pacing),
1017             OID_AUTO, "burst_reduces", CTLFLAG_RW,
1018             &rack_slot_reduction, 4,
1019             "When doing only burst mitigation what is the reduce divisor");
1020         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1021             SYSCTL_CHILDREN(rack_sysctl_root),
1022             OID_AUTO, "use_pacing", CTLFLAG_RW,
1023             &rack_pace_every_seg, 0,
1024             "If set we use pacing, if clear we use only the original burst mitigation");
1025         SYSCTL_ADD_U64(&rack_sysctl_ctx,
1026             SYSCTL_CHILDREN(rack_pacing),
1027             OID_AUTO, "rate_cap", CTLFLAG_RW,
1028             &rack_bw_rate_cap, 0,
1029             "If set we apply this value to the absolute rate cap used by pacing");
1030         SYSCTL_ADD_U8(&rack_sysctl_ctx,
1031             SYSCTL_CHILDREN(rack_sysctl_root),
1032             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1033             &rack_req_measurements, 1,
1034             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1035         /* Hardware pacing */
1036         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1037             SYSCTL_CHILDREN(rack_sysctl_root),
1038             OID_AUTO,
1039             "hdwr_pacing",
1040             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1041             "Pacing related Controls");
1042         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1043             SYSCTL_CHILDREN(rack_hw_pacing),
1044             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1045             &rack_hw_rwnd_factor, 2,
1046             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1047         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1048             SYSCTL_CHILDREN(rack_hw_pacing),
1049             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1050             &rack_enobuf_hw_boost_mult, 2,
1051             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1052         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1053             SYSCTL_CHILDREN(rack_hw_pacing),
1054             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1055             &rack_enobuf_hw_max, 2,
1056             "What is the max boost the pacing time if we see a ENOBUFS?");
1057         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1058             SYSCTL_CHILDREN(rack_hw_pacing),
1059             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1060             &rack_enobuf_hw_min, 2,
1061             "What is the min boost the pacing time if we see a ENOBUFS?");
1062         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1063             SYSCTL_CHILDREN(rack_hw_pacing),
1064             OID_AUTO, "enable", CTLFLAG_RW,
1065             &rack_enable_hw_pacing, 0,
1066             "Should RACK attempt to use hw pacing?");
1067         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1068             SYSCTL_CHILDREN(rack_hw_pacing),
1069             OID_AUTO, "rate_cap", CTLFLAG_RW,
1070             &rack_hw_rate_caps, 1,
1071             "Does the highest hardware pacing rate cap the rate we will send at??");
1072         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1073             SYSCTL_CHILDREN(rack_hw_pacing),
1074             OID_AUTO, "rate_min", CTLFLAG_RW,
1075             &rack_hw_rate_min, 0,
1076             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1077         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1078             SYSCTL_CHILDREN(rack_hw_pacing),
1079             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1080             &rack_hw_rate_to_low, 0,
1081             "If we fall below this rate, dis-engage hw pacing?");
1082         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1083             SYSCTL_CHILDREN(rack_hw_pacing),
1084             OID_AUTO, "up_only", CTLFLAG_RW,
1085             &rack_hw_up_only, 1,
1086             "Do we allow hw pacing to lower the rate selected?");
1087         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1088             SYSCTL_CHILDREN(rack_hw_pacing),
1089             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1090             &rack_hw_pace_extra_slots, 2,
1091             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1092         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1093             SYSCTL_CHILDREN(rack_sysctl_root),
1094             OID_AUTO,
1095             "timely",
1096             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1097             "Rack Timely RTT Controls");
1098         /* Timely based GP dynmics */
1099         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1100             SYSCTL_CHILDREN(rack_timely),
1101             OID_AUTO, "upper", CTLFLAG_RW,
1102             &rack_gp_per_bw_mul_up, 2,
1103             "Rack timely upper range for equal b/w (in percentage)");
1104         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1105             SYSCTL_CHILDREN(rack_timely),
1106             OID_AUTO, "lower", CTLFLAG_RW,
1107             &rack_gp_per_bw_mul_down, 4,
1108             "Rack timely lower range for equal b/w (in percentage)");
1109         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1110             SYSCTL_CHILDREN(rack_timely),
1111             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1112             &rack_gp_rtt_maxmul, 3,
1113             "Rack timely multipler of lowest rtt for rtt_max");
1114         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1115             SYSCTL_CHILDREN(rack_timely),
1116             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1117             &rack_gp_rtt_mindiv, 4,
1118             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1119         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1120             SYSCTL_CHILDREN(rack_timely),
1121             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1122             &rack_gp_rtt_minmul, 1,
1123             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1124         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1125             SYSCTL_CHILDREN(rack_timely),
1126             OID_AUTO, "decrease", CTLFLAG_RW,
1127             &rack_gp_decrease_per, 20,
1128             "Rack timely decrease percentage of our GP multiplication factor");
1129         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1130             SYSCTL_CHILDREN(rack_timely),
1131             OID_AUTO, "increase", CTLFLAG_RW,
1132             &rack_gp_increase_per, 2,
1133             "Rack timely increase perentage of our GP multiplication factor");
1134         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1135             SYSCTL_CHILDREN(rack_timely),
1136             OID_AUTO, "lowerbound", CTLFLAG_RW,
1137             &rack_per_lower_bound, 50,
1138             "Rack timely lowest percentage we allow GP multiplier to fall to");
1139         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1140             SYSCTL_CHILDREN(rack_timely),
1141             OID_AUTO, "upperboundss", CTLFLAG_RW,
1142             &rack_per_upper_bound_ss, 0,
1143             "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1144         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1145             SYSCTL_CHILDREN(rack_timely),
1146             OID_AUTO, "upperboundca", CTLFLAG_RW,
1147             &rack_per_upper_bound_ca, 0,
1148             "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1149         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1150             SYSCTL_CHILDREN(rack_timely),
1151             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1152             &rack_do_dyn_mul, 0,
1153             "Rack timely do we enable dynmaic timely goodput by default");
1154         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1155             SYSCTL_CHILDREN(rack_timely),
1156             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1157             &rack_gp_no_rec_chg, 1,
1158             "Rack timely do we prohibit the recovery multiplier from being lowered");
1159         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1160             SYSCTL_CHILDREN(rack_timely),
1161             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1162             &rack_timely_dec_clear, 6,
1163             "Rack timely what threshold do we count to before another boost during b/w decent");
1164         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1165             SYSCTL_CHILDREN(rack_timely),
1166             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1167             &rack_timely_max_push_rise, 3,
1168             "Rack timely how many times do we push up with b/w increase");
1169         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1170             SYSCTL_CHILDREN(rack_timely),
1171             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1172             &rack_timely_max_push_drop, 3,
1173             "Rack timely how many times do we push back on b/w decent");
1174         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1175             SYSCTL_CHILDREN(rack_timely),
1176             OID_AUTO, "min_segs", CTLFLAG_RW,
1177             &rack_timely_min_segs, 4,
1178             "Rack timely when setting the cwnd what is the min num segments");
1179         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1180             SYSCTL_CHILDREN(rack_timely),
1181             OID_AUTO, "noback_max", CTLFLAG_RW,
1182             &rack_use_max_for_nobackoff, 0,
1183             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1184         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1185             SYSCTL_CHILDREN(rack_timely),
1186             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1187             &rack_timely_int_timely_only, 0,
1188             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1189         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1190             SYSCTL_CHILDREN(rack_timely),
1191             OID_AUTO, "nonstop", CTLFLAG_RW,
1192             &rack_timely_no_stopping, 0,
1193             "Rack timely don't stop increase");
1194         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1195             SYSCTL_CHILDREN(rack_timely),
1196             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1197             &rack_down_raise_thresh, 100,
1198             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1199         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1200             SYSCTL_CHILDREN(rack_timely),
1201             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1202             &rack_req_segs, 1,
1203             "Bottom dragging if not these many segments outstanding and room");
1204
1205         /* TLP and Rack related parameters */
1206         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1207             SYSCTL_CHILDREN(rack_sysctl_root),
1208             OID_AUTO,
1209             "tlp",
1210             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1211             "TLP and Rack related Controls");
1212         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1213             SYSCTL_CHILDREN(rack_tlp),
1214             OID_AUTO, "use_rrr", CTLFLAG_RW,
1215             &use_rack_rr, 1,
1216             "Do we use Rack Rapid Recovery");
1217         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1218             SYSCTL_CHILDREN(rack_tlp),
1219             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1220             &rack_max_abc_post_recovery, 2,
1221             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1222         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1223             SYSCTL_CHILDREN(rack_tlp),
1224             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1225             &rack_non_rxt_use_cr, 0,
1226             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1227         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1228             SYSCTL_CHILDREN(rack_tlp),
1229             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1230             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1231             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1232         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1233             SYSCTL_CHILDREN(rack_tlp),
1234             OID_AUTO, "limit", CTLFLAG_RW,
1235             &rack_tlp_limit, 2,
1236             "How many TLP's can be sent without sending new data");
1237         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1238             SYSCTL_CHILDREN(rack_tlp),
1239             OID_AUTO, "use_greater", CTLFLAG_RW,
1240             &rack_tlp_use_greater, 1,
1241             "Should we use the rack_rtt time if its greater than srtt");
1242         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1243             SYSCTL_CHILDREN(rack_tlp),
1244             OID_AUTO, "tlpminto", CTLFLAG_RW,
1245             &rack_tlp_min, 10000,
1246             "TLP minimum timeout per the specification (in microseconds)");
1247         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1248             SYSCTL_CHILDREN(rack_tlp),
1249             OID_AUTO, "send_oldest", CTLFLAG_RW,
1250             &rack_always_send_oldest, 0,
1251             "Should we always send the oldest TLP and RACK-TLP");
1252         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1253             SYSCTL_CHILDREN(rack_tlp),
1254             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1255             &rack_limited_retran, 0,
1256             "How many times can a rack timeout drive out sends");
1257         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1258             SYSCTL_CHILDREN(rack_tlp),
1259             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1260             &rack_lower_cwnd_at_tlp, 0,
1261             "When a TLP completes a retran should we enter recovery");
1262         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1263             SYSCTL_CHILDREN(rack_tlp),
1264             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1265             &rack_reorder_thresh, 2,
1266             "What factor for rack will be added when seeing reordering (shift right)");
1267         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1268             SYSCTL_CHILDREN(rack_tlp),
1269             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1270             &rack_tlp_thresh, 1,
1271             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1272         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1273             SYSCTL_CHILDREN(rack_tlp),
1274             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1275             &rack_reorder_fade, 60000000,
1276             "Does reorder detection fade, if so how many microseconds (0 means never)");
1277         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1278             SYSCTL_CHILDREN(rack_tlp),
1279             OID_AUTO, "pktdelay", CTLFLAG_RW,
1280             &rack_pkt_delay, 1000,
1281             "Extra RACK time (in microseconds) besides reordering thresh");
1282
1283         /* Timer related controls */
1284         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1285             SYSCTL_CHILDREN(rack_sysctl_root),
1286             OID_AUTO,
1287             "timers",
1288             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1289             "Timer related controls");
1290         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1291             SYSCTL_CHILDREN(rack_timers),
1292             OID_AUTO, "persmin", CTLFLAG_RW,
1293             &rack_persist_min, 250000,
1294             "What is the minimum time in microseconds between persists");
1295         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1296             SYSCTL_CHILDREN(rack_timers),
1297             OID_AUTO, "persmax", CTLFLAG_RW,
1298             &rack_persist_max, 2000000,
1299             "What is the largest delay in microseconds between persists");
1300         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1301             SYSCTL_CHILDREN(rack_timers),
1302             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1303             &rack_delayed_ack_time, 40000,
1304             "Delayed ack time (40ms in microseconds)");
1305         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1306             SYSCTL_CHILDREN(rack_timers),
1307             OID_AUTO, "minrto", CTLFLAG_RW,
1308             &rack_rto_min, 30000,
1309             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1310         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1311             SYSCTL_CHILDREN(rack_timers),
1312             OID_AUTO, "maxrto", CTLFLAG_RW,
1313             &rack_rto_max, 4000000,
1314             "Maxiumum RTO in microseconds -- should be at least as large as min_rto");
1315         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1316             SYSCTL_CHILDREN(rack_timers),
1317             OID_AUTO, "minto", CTLFLAG_RW,
1318             &rack_min_to, 1000,
1319             "Minimum rack timeout in microseconds");
1320         /* Measure controls */
1321         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1322             SYSCTL_CHILDREN(rack_sysctl_root),
1323             OID_AUTO,
1324             "measure",
1325             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1326             "Measure related controls");
1327         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1328             SYSCTL_CHILDREN(rack_measure),
1329             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1330             &rack_wma_divisor, 8,
1331             "When doing b/w calculation what is the  divisor for the WMA");
1332         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1333             SYSCTL_CHILDREN(rack_measure),
1334             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1335             &rack_cwnd_block_ends_measure, 0,
1336             "Does a cwnd just-return end the measurement window (app limited)");
1337         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1338             SYSCTL_CHILDREN(rack_measure),
1339             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1340             &rack_rwnd_block_ends_measure, 0,
1341             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1342         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1343             SYSCTL_CHILDREN(rack_measure),
1344             OID_AUTO, "min_target", CTLFLAG_RW,
1345             &rack_def_data_window, 20,
1346             "What is the minimum target window (in mss) for a GP measurements");
1347         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1348             SYSCTL_CHILDREN(rack_measure),
1349             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1350             &rack_goal_bdp, 2,
1351             "What is the goal BDP to measure");
1352         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1353             SYSCTL_CHILDREN(rack_measure),
1354             OID_AUTO, "min_srtts", CTLFLAG_RW,
1355             &rack_min_srtts, 1,
1356             "What is the goal BDP to measure");
1357         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1358             SYSCTL_CHILDREN(rack_measure),
1359             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1360             &rack_min_measure_usec, 0,
1361             "What is the Minimum time time for a measurement if 0, this is off");
1362         /* Misc rack controls */
1363         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1364             SYSCTL_CHILDREN(rack_sysctl_root),
1365             OID_AUTO,
1366             "misc",
1367             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1368             "Misc related controls");
1369 #ifdef TCP_ACCOUNTING
1370         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1371             SYSCTL_CHILDREN(rack_misc),
1372             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1373             &rack_tcp_accounting, 0,
1374             "Should we turn on TCP accounting for all rack sessions?");
1375 #endif
1376         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1377             SYSCTL_CHILDREN(rack_misc),
1378             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1379             &rack_prr_addbackmax, 2,
1380             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1381         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1382             SYSCTL_CHILDREN(rack_misc),
1383             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1384             &rack_stats_gets_ms_rtt, 1,
1385             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1386         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1387             SYSCTL_CHILDREN(rack_misc),
1388             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1389             &rack_client_low_buf, 0,
1390             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1391         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1392             SYSCTL_CHILDREN(rack_misc),
1393             OID_AUTO, "defprofile", CTLFLAG_RW,
1394             &rack_def_profile, 0,
1395             "Should RACK use a default profile (0=no, num == profile num)?");
1396         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1397             SYSCTL_CHILDREN(rack_misc),
1398             OID_AUTO, "cmpack", CTLFLAG_RW,
1399             &rack_use_cmp_acks, 1,
1400             "Should RACK have LRO send compressed acks");
1401         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1402             SYSCTL_CHILDREN(rack_misc),
1403             OID_AUTO, "fsb", CTLFLAG_RW,
1404             &rack_use_fsb, 1,
1405             "Should RACK use the fast send block?");
1406         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1407             SYSCTL_CHILDREN(rack_misc),
1408             OID_AUTO, "rfo", CTLFLAG_RW,
1409             &rack_use_rfo, 1,
1410             "Should RACK use rack_fast_output()?");
1411         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1412             SYSCTL_CHILDREN(rack_misc),
1413             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1414             &rack_use_rsm_rfo, 1,
1415             "Should RACK use rack_fast_rsm_output()?");
1416         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1417             SYSCTL_CHILDREN(rack_misc),
1418             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1419             &rack_enable_shared_cwnd, 1,
1420             "Should RACK try to use the shared cwnd on connections where allowed");
1421         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1422             SYSCTL_CHILDREN(rack_misc),
1423             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1424             &rack_limits_scwnd, 1,
1425             "Should RACK place low end time limits on the shared cwnd feature");
1426         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1427             SYSCTL_CHILDREN(rack_misc),
1428             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1429             &rack_enable_mqueue_for_nonpaced, 0,
1430             "Should RACK use mbuf queuing for non-paced connections");
1431         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1432             SYSCTL_CHILDREN(rack_misc),
1433             OID_AUTO, "iMac_dack", CTLFLAG_RW,
1434             &rack_use_imac_dack, 0,
1435             "Should RACK try to emulate iMac delayed ack");
1436         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1437             SYSCTL_CHILDREN(rack_misc),
1438             OID_AUTO, "no_prr", CTLFLAG_RW,
1439             &rack_disable_prr, 0,
1440             "Should RACK not use prr and only pace (must have pacing on)");
1441         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1442             SYSCTL_CHILDREN(rack_misc),
1443             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1444             &rack_verbose_logging, 0,
1445             "Should RACK black box logging be verbose");
1446         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1447             SYSCTL_CHILDREN(rack_misc),
1448             OID_AUTO, "data_after_close", CTLFLAG_RW,
1449             &rack_ignore_data_after_close, 1,
1450             "Do we hold off sending a RST until all pending data is ack'd");
1451         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1452             SYSCTL_CHILDREN(rack_misc),
1453             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1454             &rack_sack_not_required, 1,
1455             "Do we allow rack to run on connections not supporting SACK");
1456         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1457             SYSCTL_CHILDREN(rack_misc),
1458             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1459             &rack_send_a_lot_in_prr, 1,
1460             "Send a lot in prr");
1461         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1462             SYSCTL_CHILDREN(rack_misc),
1463             OID_AUTO, "autoscale", CTLFLAG_RW,
1464             &rack_autosndbuf_inc, 20,
1465             "What percentage should rack scale up its snd buffer by?");
1466         /* Sack Attacker detection stuff */
1467         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1468             SYSCTL_CHILDREN(rack_attack),
1469             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1470             &rack_highest_sack_thresh_seen, 0,
1471             "Highest sack to ack ratio seen");
1472         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1473             SYSCTL_CHILDREN(rack_attack),
1474             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1475             &rack_highest_move_thresh_seen, 0,
1476             "Highest move to non-move ratio seen");
1477         rack_ack_total = counter_u64_alloc(M_WAITOK);
1478         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1479             SYSCTL_CHILDREN(rack_attack),
1480             OID_AUTO, "acktotal", CTLFLAG_RD,
1481             &rack_ack_total,
1482             "Total number of Ack's");
1483         rack_express_sack = counter_u64_alloc(M_WAITOK);
1484         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1485             SYSCTL_CHILDREN(rack_attack),
1486             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1487             &rack_express_sack,
1488             "Total expresss number of Sack's");
1489         rack_sack_total = counter_u64_alloc(M_WAITOK);
1490         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1491             SYSCTL_CHILDREN(rack_attack),
1492             OID_AUTO, "sacktotal", CTLFLAG_RD,
1493             &rack_sack_total,
1494             "Total number of SACKs");
1495         rack_move_none = counter_u64_alloc(M_WAITOK);
1496         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1497             SYSCTL_CHILDREN(rack_attack),
1498             OID_AUTO, "move_none", CTLFLAG_RD,
1499             &rack_move_none,
1500             "Total number of SACK index reuse of postions under threshold");
1501         rack_move_some = counter_u64_alloc(M_WAITOK);
1502         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1503             SYSCTL_CHILDREN(rack_attack),
1504             OID_AUTO, "move_some", CTLFLAG_RD,
1505             &rack_move_some,
1506             "Total number of SACK index reuse of postions over threshold");
1507         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1508         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1509             SYSCTL_CHILDREN(rack_attack),
1510             OID_AUTO, "attacks", CTLFLAG_RD,
1511             &rack_sack_attacks_detected,
1512             "Total number of SACK attackers that had sack disabled");
1513         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1514         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1515             SYSCTL_CHILDREN(rack_attack),
1516             OID_AUTO, "reversed", CTLFLAG_RD,
1517             &rack_sack_attacks_reversed,
1518             "Total number of SACK attackers that were later determined false positive");
1519         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1520         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1521             SYSCTL_CHILDREN(rack_attack),
1522             OID_AUTO, "nextmerge", CTLFLAG_RD,
1523             &rack_sack_used_next_merge,
1524             "Total number of times we used the next merge");
1525         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1526         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1527             SYSCTL_CHILDREN(rack_attack),
1528             OID_AUTO, "prevmerge", CTLFLAG_RD,
1529             &rack_sack_used_prev_merge,
1530             "Total number of times we used the prev merge");
1531         /* Counters */
1532         rack_fto_send = counter_u64_alloc(M_WAITOK);
1533         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1534             SYSCTL_CHILDREN(rack_counters),
1535             OID_AUTO, "fto_send", CTLFLAG_RD,
1536             &rack_fto_send, "Total number of rack_fast_output sends");
1537         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1538         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1539             SYSCTL_CHILDREN(rack_counters),
1540             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1541             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1542         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1543         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1544             SYSCTL_CHILDREN(rack_counters),
1545             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1546             &rack_nfto_resend, "Total number of rack_output retransmissions");
1547         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1548         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1549             SYSCTL_CHILDREN(rack_counters),
1550             OID_AUTO, "nfto_send", CTLFLAG_RD,
1551             &rack_non_fto_send, "Total number of rack_output first sends");
1552         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1553         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1554             SYSCTL_CHILDREN(rack_counters),
1555             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1556             &rack_extended_rfo, "Total number of times we extended rfo");
1557
1558         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1559         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1560             SYSCTL_CHILDREN(rack_counters),
1561             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1562             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1563         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1564
1565         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1566             SYSCTL_CHILDREN(rack_counters),
1567             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1568             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1569
1570
1571
1572         rack_badfr = counter_u64_alloc(M_WAITOK);
1573         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1574             SYSCTL_CHILDREN(rack_counters),
1575             OID_AUTO, "badfr", CTLFLAG_RD,
1576             &rack_badfr, "Total number of bad FRs");
1577         rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1578         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1579             SYSCTL_CHILDREN(rack_counters),
1580             OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1581             &rack_badfr_bytes, "Total number of bad FRs");
1582         rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1583         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1584             SYSCTL_CHILDREN(rack_counters),
1585             OID_AUTO, "prrsndret", CTLFLAG_RD,
1586             &rack_rtm_prr_retran,
1587             "Total number of prr based retransmits");
1588         rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1589         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1590             SYSCTL_CHILDREN(rack_counters),
1591             OID_AUTO, "prrsndnew", CTLFLAG_RD,
1592             &rack_rtm_prr_newdata,
1593             "Total number of prr based new transmits");
1594         rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1595         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1596             SYSCTL_CHILDREN(rack_counters),
1597             OID_AUTO, "tsnf", CTLFLAG_RD,
1598             &rack_timestamp_mismatch,
1599             "Total number of timestamps that we could not find the reported ts");
1600         rack_find_high = counter_u64_alloc(M_WAITOK);
1601         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1602             SYSCTL_CHILDREN(rack_counters),
1603             OID_AUTO, "findhigh", CTLFLAG_RD,
1604             &rack_find_high,
1605             "Total number of FIN causing find-high");
1606         rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1607         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1608             SYSCTL_CHILDREN(rack_counters),
1609             OID_AUTO, "reordering", CTLFLAG_RD,
1610             &rack_reorder_seen,
1611             "Total number of times we added delay due to reordering");
1612         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1613         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1614             SYSCTL_CHILDREN(rack_counters),
1615             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1616             &rack_tlp_tot,
1617             "Total number of tail loss probe expirations");
1618         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1619         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1620             SYSCTL_CHILDREN(rack_counters),
1621             OID_AUTO, "tlp_new", CTLFLAG_RD,
1622             &rack_tlp_newdata,
1623             "Total number of tail loss probe sending new data");
1624         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1625         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1626             SYSCTL_CHILDREN(rack_counters),
1627             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1628             &rack_tlp_retran,
1629             "Total number of tail loss probe sending retransmitted data");
1630         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1631         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1632             SYSCTL_CHILDREN(rack_counters),
1633             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1634             &rack_tlp_retran_bytes,
1635             "Total bytes of tail loss probe sending retransmitted data");
1636         rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1637         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1638             SYSCTL_CHILDREN(rack_counters),
1639             OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1640             &rack_tlp_retran_fail,
1641             "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1642         rack_to_tot = counter_u64_alloc(M_WAITOK);
1643         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1644             SYSCTL_CHILDREN(rack_counters),
1645             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1646             &rack_to_tot,
1647             "Total number of times the rack to expired");
1648         rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1649         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1650             SYSCTL_CHILDREN(rack_counters),
1651             OID_AUTO, "arm_rack", CTLFLAG_RD,
1652             &rack_to_arm_rack,
1653             "Total number of times the rack timer armed");
1654         rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1655         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1656             SYSCTL_CHILDREN(rack_counters),
1657             OID_AUTO, "arm_tlp", CTLFLAG_RD,
1658             &rack_to_arm_tlp,
1659             "Total number of times the tlp timer armed");
1660         rack_calc_zero = counter_u64_alloc(M_WAITOK);
1661         rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1662         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1663             SYSCTL_CHILDREN(rack_counters),
1664             OID_AUTO, "calc_zero", CTLFLAG_RD,
1665             &rack_calc_zero,
1666             "Total number of times pacing time worked out to zero");
1667         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1668             SYSCTL_CHILDREN(rack_counters),
1669             OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1670             &rack_calc_nonzero,
1671             "Total number of times pacing time worked out to non-zero");
1672         rack_paced_segments = counter_u64_alloc(M_WAITOK);
1673         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674             SYSCTL_CHILDREN(rack_counters),
1675             OID_AUTO, "paced", CTLFLAG_RD,
1676             &rack_paced_segments,
1677             "Total number of times a segment send caused hptsi");
1678         rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1679         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680             SYSCTL_CHILDREN(rack_counters),
1681             OID_AUTO, "unpaced", CTLFLAG_RD,
1682             &rack_unpaced_segments,
1683             "Total number of times a segment did not cause hptsi");
1684         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1685         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1686             SYSCTL_CHILDREN(rack_counters),
1687             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1688             &rack_saw_enobuf,
1689             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1690         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1691         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1692             SYSCTL_CHILDREN(rack_counters),
1693             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1694             &rack_saw_enobuf_hw,
1695             "Total number of times a send returned enobuf for hdwr paced connections");
1696         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1697         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1698             SYSCTL_CHILDREN(rack_counters),
1699             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1700             &rack_saw_enetunreach,
1701             "Total number of times a send received a enetunreachable");
1702         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1703         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1704             SYSCTL_CHILDREN(rack_counters),
1705             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1706             &rack_hot_alloc,
1707             "Total allocations from the top of our list");
1708         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1709         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1710             SYSCTL_CHILDREN(rack_counters),
1711             OID_AUTO, "allocs", CTLFLAG_RD,
1712             &rack_to_alloc,
1713             "Total allocations of tracking structures");
1714         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1715         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1716             SYSCTL_CHILDREN(rack_counters),
1717             OID_AUTO, "allochard", CTLFLAG_RD,
1718             &rack_to_alloc_hard,
1719             "Total allocations done with sleeping the hard way");
1720         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1721         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1722             SYSCTL_CHILDREN(rack_counters),
1723             OID_AUTO, "allocemerg", CTLFLAG_RD,
1724             &rack_to_alloc_emerg,
1725             "Total allocations done from emergency cache");
1726         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1727         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1728             SYSCTL_CHILDREN(rack_counters),
1729             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1730             &rack_to_alloc_limited,
1731             "Total allocations dropped due to limit");
1732         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1733         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1734             SYSCTL_CHILDREN(rack_counters),
1735             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1736             &rack_alloc_limited_conns,
1737             "Connections with allocations dropped due to limit");
1738         rack_split_limited = counter_u64_alloc(M_WAITOK);
1739         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1740             SYSCTL_CHILDREN(rack_counters),
1741             OID_AUTO, "split_limited", CTLFLAG_RD,
1742             &rack_split_limited,
1743             "Split allocations dropped due to limit");
1744
1745         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
1746                 char name[32];
1747                 sprintf(name, "cmp_ack_cnt_%d", i);
1748                 rack_proc_comp_ack[i] = counter_u64_alloc(M_WAITOK);
1749                 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750                                        SYSCTL_CHILDREN(rack_counters),
1751                                        OID_AUTO, name, CTLFLAG_RD,
1752                                        &rack_proc_comp_ack[i],
1753                                        "Number of compressed acks we processed");
1754         }
1755         rack_large_ackcmp = counter_u64_alloc(M_WAITOK);
1756         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1757             SYSCTL_CHILDREN(rack_counters),
1758             OID_AUTO, "cmp_large_mbufs", CTLFLAG_RD,
1759             &rack_large_ackcmp,
1760             "Number of TCP connections with large mbuf's for compressed acks");
1761         rack_small_ackcmp = counter_u64_alloc(M_WAITOK);
1762         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1763             SYSCTL_CHILDREN(rack_counters),
1764             OID_AUTO, "cmp_small_mbufs", CTLFLAG_RD,
1765             &rack_small_ackcmp,
1766             "Number of TCP connections with small mbuf's for compressed acks");
1767 #ifdef INVARIANTS
1768         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1769         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1770             SYSCTL_CHILDREN(rack_counters),
1771             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1772             &rack_adjust_map_bw,
1773             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1774 #endif
1775         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1776         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1777             SYSCTL_CHILDREN(rack_counters),
1778             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1779             &rack_multi_single_eq,
1780             "Number of compressed acks total represented");
1781         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1782         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1783             SYSCTL_CHILDREN(rack_counters),
1784             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1785             &rack_proc_non_comp_ack,
1786             "Number of non compresseds acks that we processed");
1787
1788
1789         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1790         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1791             SYSCTL_CHILDREN(rack_counters),
1792             OID_AUTO, "sack_long", CTLFLAG_RD,
1793             &rack_sack_proc_all,
1794             "Total times we had to walk whole list for sack processing");
1795         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1796         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1797             SYSCTL_CHILDREN(rack_counters),
1798             OID_AUTO, "sack_restart", CTLFLAG_RD,
1799             &rack_sack_proc_restart,
1800             "Total times we had to walk whole list due to a restart");
1801         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1802         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1803             SYSCTL_CHILDREN(rack_counters),
1804             OID_AUTO, "sack_short", CTLFLAG_RD,
1805             &rack_sack_proc_short,
1806             "Total times we took shortcut for sack processing");
1807         rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1808         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1809             SYSCTL_CHILDREN(rack_counters),
1810             OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1811             &rack_enter_tlp_calc,
1812             "Total times we called calc-tlp");
1813         rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1814         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1815             SYSCTL_CHILDREN(rack_counters),
1816             OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1817             &rack_used_tlpmethod,
1818             "Total number of runt sacks");
1819         rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1820         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1821             SYSCTL_CHILDREN(rack_counters),
1822             OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1823             &rack_used_tlpmethod2,
1824             "Total number of times we hit TLP method 2");
1825         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1826         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1827             SYSCTL_CHILDREN(rack_attack),
1828             OID_AUTO, "skipacked", CTLFLAG_RD,
1829             &rack_sack_skipped_acked,
1830             "Total number of times we skipped previously sacked");
1831         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1832         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1833             SYSCTL_CHILDREN(rack_attack),
1834             OID_AUTO, "ofsplit", CTLFLAG_RD,
1835             &rack_sack_splits,
1836             "Total number of times we did the old fashion tree split");
1837         rack_progress_drops = counter_u64_alloc(M_WAITOK);
1838         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1839             SYSCTL_CHILDREN(rack_counters),
1840             OID_AUTO, "prog_drops", CTLFLAG_RD,
1841             &rack_progress_drops,
1842             "Total number of progress drops");
1843         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1844         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1845             SYSCTL_CHILDREN(rack_counters),
1846             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1847             &rack_input_idle_reduces,
1848             "Total number of idle reductions on input");
1849         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1850         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1851             SYSCTL_CHILDREN(rack_counters),
1852             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1853             &rack_collapsed_win,
1854             "Total number of collapsed windows");
1855         rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1856         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1857             SYSCTL_CHILDREN(rack_counters),
1858             OID_AUTO, "tlp_nada", CTLFLAG_RD,
1859             &rack_tlp_does_nada,
1860             "Total number of nada tlp calls");
1861         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1862         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1863             SYSCTL_CHILDREN(rack_counters),
1864             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1865             &rack_try_scwnd,
1866             "Total number of scwnd attempts");
1867
1868         rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1869         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1870             SYSCTL_CHILDREN(rack_counters),
1871             OID_AUTO, "timer_hole", CTLFLAG_RD,
1872             &rack_per_timer_hole,
1873             "Total persists start in timer hole");
1874
1875         rack_sbsndptr_wrong = counter_u64_alloc(M_WAITOK);
1876         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1877             SYSCTL_CHILDREN(rack_counters),
1878             OID_AUTO, "sndptr_wrong", CTLFLAG_RD,
1879             &rack_sbsndptr_wrong, "Total number of times the saved sbsndptr was incorret");
1880         rack_sbsndptr_right = counter_u64_alloc(M_WAITOK);
1881         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1882             SYSCTL_CHILDREN(rack_counters),
1883             OID_AUTO, "sndptr_right", CTLFLAG_RD,
1884             &rack_sbsndptr_right, "Total number of times the saved sbsndptr was corret");
1885
1886         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1887         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1888             OID_AUTO, "outsize", CTLFLAG_RD,
1889             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1890         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1891         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1892             OID_AUTO, "opts", CTLFLAG_RD,
1893             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1894         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1895             SYSCTL_CHILDREN(rack_sysctl_root),
1896             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1897             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1898 }
1899
1900 static __inline int
1901 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1902 {
1903         if (SEQ_GEQ(b->r_start, a->r_start) &&
1904             SEQ_LT(b->r_start, a->r_end)) {
1905                 /*
1906                  * The entry b is within the
1907                  * block a. i.e.:
1908                  * a --   |-------------|
1909                  * b --   |----|
1910                  * <or>
1911                  * b --       |------|
1912                  * <or>
1913                  * b --       |-----------|
1914                  */
1915                 return (0);
1916         } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1917                 /*
1918                  * b falls as either the next
1919                  * sequence block after a so a
1920                  * is said to be smaller than b.
1921                  * i.e:
1922                  * a --   |------|
1923                  * b --          |--------|
1924                  * or
1925                  * b --              |-----|
1926                  */
1927                 return (1);
1928         }
1929         /*
1930          * Whats left is where a is
1931          * larger than b. i.e:
1932          * a --         |-------|
1933          * b --  |---|
1934          * or even possibly
1935          * b --   |--------------|
1936          */
1937         return (-1);
1938 }
1939
1940 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1941 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1942
1943 static uint32_t
1944 rc_init_window(struct tcp_rack *rack)
1945 {
1946         uint32_t win;
1947
1948         if (rack->rc_init_win == 0) {
1949                 /*
1950                  * Nothing set by the user, use the system stack
1951                  * default.
1952                  */
1953                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1954         }
1955         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1956         return (win);
1957 }
1958
1959 static uint64_t
1960 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1961 {
1962         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1963                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1964         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1965                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1966         else
1967                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1968 }
1969
1970 static uint64_t
1971 rack_get_bw(struct tcp_rack *rack)
1972 {
1973         if (rack->use_fixed_rate) {
1974                 /* Return the fixed pacing rate */
1975                 return (rack_get_fixed_pacing_bw(rack));
1976         }
1977         if (rack->r_ctl.gp_bw == 0) {
1978                 /*
1979                  * We have yet no b/w measurement,
1980                  * if we have a user set initial bw
1981                  * return it. If we don't have that and
1982                  * we have an srtt, use the tcp IW (10) to
1983                  * calculate a fictional b/w over the SRTT
1984                  * which is more or less a guess. Note
1985                  * we don't use our IW from rack on purpose
1986                  * so if we have like IW=30, we are not
1987                  * calculating a "huge" b/w.
1988                  */
1989                 uint64_t bw, srtt;
1990                 if (rack->r_ctl.init_rate)
1991                         return (rack->r_ctl.init_rate);
1992
1993                 /* Has the user set a max peak rate? */
1994 #ifdef NETFLIX_PEAKRATE
1995                 if (rack->rc_tp->t_maxpeakrate)
1996                         return (rack->rc_tp->t_maxpeakrate);
1997 #endif
1998                 /* Ok lets come up with the IW guess, if we have a srtt */
1999                 if (rack->rc_tp->t_srtt == 0) {
2000                         /*
2001                          * Go with old pacing method
2002                          * i.e. burst mitigation only.
2003                          */
2004                         return (0);
2005                 }
2006                 /* Ok lets get the initial TCP win (not racks) */
2007                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2008                 srtt = (uint64_t)rack->rc_tp->t_srtt;
2009                 bw *= (uint64_t)USECS_IN_SECOND;
2010                 bw /= srtt;
2011                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2012                         bw = rack->r_ctl.bw_rate_cap;
2013                 return (bw);
2014         } else {
2015                 uint64_t bw;
2016
2017                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2018                         /* Averaging is done, we can return the value */
2019                         bw = rack->r_ctl.gp_bw;
2020                 } else {
2021                         /* Still doing initial average must calculate */
2022                         bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
2023                 }
2024 #ifdef NETFLIX_PEAKRATE
2025                 if ((rack->rc_tp->t_maxpeakrate) &&
2026                     (bw > rack->rc_tp->t_maxpeakrate)) {
2027                         /* The user has set a peak rate to pace at
2028                          * don't allow us to pace faster than that.
2029                          */
2030                         return (rack->rc_tp->t_maxpeakrate);
2031                 }
2032 #endif
2033                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2034                         bw = rack->r_ctl.bw_rate_cap;
2035                 return (bw);
2036         }
2037 }
2038
2039 static uint16_t
2040 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2041 {
2042         if (rack->use_fixed_rate) {
2043                 return (100);
2044         } else if (rack->in_probe_rtt && (rsm == NULL))
2045                 return (rack->r_ctl.rack_per_of_gp_probertt);
2046         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2047                   rack->r_ctl.rack_per_of_gp_rec)) {
2048                 if (rsm) {
2049                         /* a retransmission always use the recovery rate */
2050                         return (rack->r_ctl.rack_per_of_gp_rec);
2051                 } else if (rack->rack_rec_nonrxt_use_cr) {
2052                         /* Directed to use the configured rate */
2053                         goto configured_rate;
2054                 } else if (rack->rack_no_prr &&
2055                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2056                         /* No PRR, lets just use the b/w estimate only */
2057                         return (100);
2058                 } else {
2059                         /*
2060                          * Here we may have a non-retransmit but we
2061                          * have no overrides, so just use the recovery
2062                          * rate (prr is in effect).
2063                          */
2064                         return (rack->r_ctl.rack_per_of_gp_rec);
2065                 }
2066         }
2067 configured_rate:
2068         /* For the configured rate we look at our cwnd vs the ssthresh */
2069         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2070                 return (rack->r_ctl.rack_per_of_gp_ss);
2071         else
2072                 return (rack->r_ctl.rack_per_of_gp_ca);
2073 }
2074
2075 static void
2076 rack_log_hdwr_pacing(struct tcp_rack *rack,
2077                      uint64_t rate, uint64_t hw_rate, int line,
2078                      int error, uint16_t mod)
2079 {
2080         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2081                 union tcp_log_stackspecific log;
2082                 struct timeval tv;
2083                 const struct ifnet *ifp;
2084
2085                 memset(&log, 0, sizeof(log));
2086                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2087                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2088                 if (rack->r_ctl.crte) {
2089                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2090                 } else if (rack->rc_inp->inp_route.ro_nh &&
2091                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2092                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2093                 } else
2094                         ifp = NULL;
2095                 if (ifp) {
2096                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2097                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2098                 }
2099                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2100                 log.u_bbr.bw_inuse = rate;
2101                 log.u_bbr.flex5 = line;
2102                 log.u_bbr.flex6 = error;
2103                 log.u_bbr.flex7 = mod;
2104                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2105                 log.u_bbr.flex8 = rack->use_fixed_rate;
2106                 log.u_bbr.flex8 <<= 1;
2107                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2108                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2109                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2110                 if (rack->r_ctl.crte)
2111                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2112                 else
2113                         log.u_bbr.cur_del_rate = 0;
2114                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2115                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2116                     &rack->rc_inp->inp_socket->so_rcv,
2117                     &rack->rc_inp->inp_socket->so_snd,
2118                     BBR_LOG_HDWR_PACE, 0,
2119                     0, &log, false, &tv);
2120         }
2121 }
2122
2123 static uint64_t
2124 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2125 {
2126         /*
2127          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2128          */
2129         uint64_t bw_est, high_rate;
2130         uint64_t gain;
2131
2132         gain = (uint64_t)rack_get_output_gain(rack, rsm);
2133         bw_est = bw * gain;
2134         bw_est /= (uint64_t)100;
2135         /* Never fall below the minimum (def 64kbps) */
2136         if (bw_est < RACK_MIN_BW)
2137                 bw_est = RACK_MIN_BW;
2138         if (rack->r_rack_hw_rate_caps) {
2139                 /* Rate caps are in place */
2140                 if (rack->r_ctl.crte != NULL) {
2141                         /* We have a hdwr rate already */
2142                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2143                         if (bw_est >= high_rate) {
2144                                 /* We are capping bw at the highest rate table entry */
2145                                 rack_log_hdwr_pacing(rack,
2146                                                      bw_est, high_rate, __LINE__,
2147                                                      0, 3);
2148                                 bw_est = high_rate;
2149                                 if (capped)
2150                                         *capped = 1;
2151                         }
2152                 } else if ((rack->rack_hdrw_pacing == 0) &&
2153                            (rack->rack_hdw_pace_ena) &&
2154                            (rack->rack_attempt_hdwr_pace == 0) &&
2155                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2156                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2157                         /*
2158                          * Special case, we have not yet attempted hardware
2159                          * pacing, and yet we may, when we do, find out if we are
2160                          * above the highest rate. We need to know the maxbw for the interface
2161                          * in question (if it supports ratelimiting). We get back
2162                          * a 0, if the interface is not found in the RL lists.
2163                          */
2164                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2165                         if (high_rate) {
2166                                 /* Yep, we have a rate is it above this rate? */
2167                                 if (bw_est > high_rate) {
2168                                         bw_est = high_rate;
2169                                         if (capped)
2170                                                 *capped = 1;
2171                                 }
2172                         }
2173                 }
2174         }
2175         return (bw_est);
2176 }
2177
2178 static void
2179 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2180 {
2181         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2182                 union tcp_log_stackspecific log;
2183                 struct timeval tv;
2184
2185                 if ((mod != 1) && (rack_verbose_logging == 0)) {
2186                         /*
2187                          * We get 3 values currently for mod
2188                          * 1 - We are retransmitting and this tells the reason.
2189                          * 2 - We are clearing a dup-ack count.
2190                          * 3 - We are incrementing a dup-ack count.
2191                          *
2192                          * The clear/increment are only logged
2193                          * if you have BBverbose on.
2194                          */
2195                         return;
2196                 }
2197                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2198                 log.u_bbr.flex1 = tsused;
2199                 log.u_bbr.flex2 = thresh;
2200                 log.u_bbr.flex3 = rsm->r_flags;
2201                 log.u_bbr.flex4 = rsm->r_dupack;
2202                 log.u_bbr.flex5 = rsm->r_start;
2203                 log.u_bbr.flex6 = rsm->r_end;
2204                 log.u_bbr.flex8 = mod;
2205                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2206                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2207                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2208                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2209                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2210                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2211                 log.u_bbr.pacing_gain = rack->r_must_retran;
2212                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2213                     &rack->rc_inp->inp_socket->so_rcv,
2214                     &rack->rc_inp->inp_socket->so_snd,
2215                     BBR_LOG_SETTINGS_CHG, 0,
2216                     0, &log, false, &tv);
2217         }
2218 }
2219
2220 static void
2221 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2222 {
2223         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2224                 union tcp_log_stackspecific log;
2225                 struct timeval tv;
2226
2227                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2228                 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2229                 log.u_bbr.flex2 = to;
2230                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2231                 log.u_bbr.flex4 = slot;
2232                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2233                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2234                 log.u_bbr.flex7 = rack->rc_in_persist;
2235                 log.u_bbr.flex8 = which;
2236                 if (rack->rack_no_prr)
2237                         log.u_bbr.pkts_out = 0;
2238                 else
2239                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2240                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2241                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2242                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2243                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2244                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2245                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2246                 log.u_bbr.pacing_gain = rack->r_must_retran;
2247                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2248                 log.u_bbr.lost = rack_rto_min;
2249                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2250                     &rack->rc_inp->inp_socket->so_rcv,
2251                     &rack->rc_inp->inp_socket->so_snd,
2252                     BBR_LOG_TIMERSTAR, 0,
2253                     0, &log, false, &tv);
2254         }
2255 }
2256
2257 static void
2258 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2259 {
2260         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2261                 union tcp_log_stackspecific log;
2262                 struct timeval tv;
2263
2264                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2265                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2266                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2267                 log.u_bbr.flex8 = to_num;
2268                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2269                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2270                 if (rsm == NULL)
2271                         log.u_bbr.flex3 = 0;
2272                 else
2273                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2274                 if (rack->rack_no_prr)
2275                         log.u_bbr.flex5 = 0;
2276                 else
2277                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2278                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2279                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2280                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2281                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2282                 log.u_bbr.pacing_gain = rack->r_must_retran;
2283                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2284                     &rack->rc_inp->inp_socket->so_rcv,
2285                     &rack->rc_inp->inp_socket->so_snd,
2286                     BBR_LOG_RTO, 0,
2287                     0, &log, false, &tv);
2288         }
2289 }
2290
2291 static void
2292 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2293                  struct rack_sendmap *prev,
2294                  struct rack_sendmap *rsm,
2295                  struct rack_sendmap *next,
2296                  int flag, uint32_t th_ack, int line)
2297 {
2298         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2299                 union tcp_log_stackspecific log;
2300                 struct timeval tv;
2301
2302                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2303                 log.u_bbr.flex8 = flag;
2304                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2305                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2306                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2307                 log.u_bbr.delRate = (uint64_t)rsm;
2308                 log.u_bbr.rttProp = (uint64_t)next;
2309                 log.u_bbr.flex7 = 0;
2310                 if (prev) {
2311                         log.u_bbr.flex1 = prev->r_start;
2312                         log.u_bbr.flex2 = prev->r_end;
2313                         log.u_bbr.flex7 |= 0x4;
2314                 }
2315                 if (rsm) {
2316                         log.u_bbr.flex3 = rsm->r_start;
2317                         log.u_bbr.flex4 = rsm->r_end;
2318                         log.u_bbr.flex7 |= 0x2;
2319                 }
2320                 if (next) {
2321                         log.u_bbr.flex5 = next->r_start;
2322                         log.u_bbr.flex6 = next->r_end;
2323                         log.u_bbr.flex7 |= 0x1;
2324                 }
2325                 log.u_bbr.applimited = line;
2326                 log.u_bbr.pkts_out = th_ack;
2327                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2328                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2329                 if (rack->rack_no_prr)
2330                         log.u_bbr.lost = 0;
2331                 else
2332                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2333                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2334                     &rack->rc_inp->inp_socket->so_rcv,
2335                     &rack->rc_inp->inp_socket->so_snd,
2336                     TCP_LOG_MAPCHG, 0,
2337                     0, &log, false, &tv);
2338         }
2339 }
2340
2341 static void
2342 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2343                  struct rack_sendmap *rsm, int conf)
2344 {
2345         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2346                 union tcp_log_stackspecific log;
2347                 struct timeval tv;
2348                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2349                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2350                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2351                 log.u_bbr.flex1 = t;
2352                 log.u_bbr.flex2 = len;
2353                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2354                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2355                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2356                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2357                 log.u_bbr.flex7 = conf;
2358                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2359                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2360                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2361                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2362                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2363                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2364                 if (rsm) {
2365                         log.u_bbr.pkt_epoch = rsm->r_start;
2366                         log.u_bbr.lost = rsm->r_end;
2367                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2368                         log.u_bbr.pacing_gain = rsm->r_flags;
2369                 } else {
2370                         /* Its a SYN */
2371                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2372                         log.u_bbr.lost = 0;
2373                         log.u_bbr.cwnd_gain = 0;
2374                         log.u_bbr.pacing_gain = 0;
2375                 }
2376                 /* Write out general bits of interest rrs here */
2377                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2378                 log.u_bbr.use_lt_bw <<= 1;
2379                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2380                 log.u_bbr.use_lt_bw <<= 1;
2381                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2382                 log.u_bbr.use_lt_bw <<= 1;
2383                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2384                 log.u_bbr.use_lt_bw <<= 1;
2385                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2386                 log.u_bbr.use_lt_bw <<= 1;
2387                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2388                 log.u_bbr.use_lt_bw <<= 1;
2389                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2390                 log.u_bbr.use_lt_bw <<= 1;
2391                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2392                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2393                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2394                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2395                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2396                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2397                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2398                 log.u_bbr.bw_inuse <<= 32;
2399                 if (rsm)
2400                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2401                 TCP_LOG_EVENTP(tp, NULL,
2402                     &rack->rc_inp->inp_socket->so_rcv,
2403                     &rack->rc_inp->inp_socket->so_snd,
2404                     BBR_LOG_BBRRTT, 0,
2405                     0, &log, false, &tv);
2406
2407
2408         }
2409 }
2410
2411 static void
2412 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2413 {
2414         /*
2415          * Log the rtt sample we are
2416          * applying to the srtt algorithm in
2417          * useconds.
2418          */
2419         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2420                 union tcp_log_stackspecific log;
2421                 struct timeval tv;
2422
2423                 /* Convert our ms to a microsecond */
2424                 memset(&log, 0, sizeof(log));
2425                 log.u_bbr.flex1 = rtt;
2426                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2427                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2428                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2429                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2430                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2431                 log.u_bbr.flex7 = 1;
2432                 log.u_bbr.flex8 = rack->sack_attack_disable;
2433                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2434                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2435                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2436                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2437                 log.u_bbr.pacing_gain = rack->r_must_retran;
2438                 /*
2439                  * We capture in delRate the upper 32 bits as
2440                  * the confidence level we had declared, and the
2441                  * lower 32 bits as the actual RTT using the arrival
2442                  * timestamp.
2443                  */
2444                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2445                 log.u_bbr.delRate <<= 32;
2446                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2447                 /* Lets capture all the things that make up t_rtxcur */
2448                 log.u_bbr.applimited = rack_rto_min;
2449                 log.u_bbr.epoch = rack_rto_max;
2450                 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2451                 log.u_bbr.lost = rack_rto_min;
2452                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2453                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2454                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2455                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2456                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2457                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2458                     &rack->rc_inp->inp_socket->so_rcv,
2459                     &rack->rc_inp->inp_socket->so_snd,
2460                     TCP_LOG_RTT, 0,
2461                     0, &log, false, &tv);
2462         }
2463 }
2464
2465 static void
2466 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2467 {
2468         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2469                 union tcp_log_stackspecific log;
2470                 struct timeval tv;
2471
2472                 /* Convert our ms to a microsecond */
2473                 memset(&log, 0, sizeof(log));
2474                 log.u_bbr.flex1 = rtt;
2475                 log.u_bbr.flex2 = send_time;
2476                 log.u_bbr.flex3 = ack_time;
2477                 log.u_bbr.flex4 = where;
2478                 log.u_bbr.flex7 = 2;
2479                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2480                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2481                     &rack->rc_inp->inp_socket->so_rcv,
2482                     &rack->rc_inp->inp_socket->so_snd,
2483                     TCP_LOG_RTT, 0,
2484                     0, &log, false, &tv);
2485         }
2486 }
2487
2488
2489
2490 static inline void
2491 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2492 {
2493         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2494                 union tcp_log_stackspecific log;
2495                 struct timeval tv;
2496
2497                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2498                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2499                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2500                 log.u_bbr.flex1 = line;
2501                 log.u_bbr.flex2 = tick;
2502                 log.u_bbr.flex3 = tp->t_maxunacktime;
2503                 log.u_bbr.flex4 = tp->t_acktime;
2504                 log.u_bbr.flex8 = event;
2505                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2506                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2507                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2508                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2509                 log.u_bbr.pacing_gain = rack->r_must_retran;
2510                 TCP_LOG_EVENTP(tp, NULL,
2511                     &rack->rc_inp->inp_socket->so_rcv,
2512                     &rack->rc_inp->inp_socket->so_snd,
2513                     BBR_LOG_PROGRESS, 0,
2514                     0, &log, false, &tv);
2515         }
2516 }
2517
2518 static void
2519 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2520 {
2521         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2522                 union tcp_log_stackspecific log;
2523
2524                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2525                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2526                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2527                 log.u_bbr.flex1 = slot;
2528                 if (rack->rack_no_prr)
2529                         log.u_bbr.flex2 = 0;
2530                 else
2531                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2532                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2533                 log.u_bbr.flex8 = rack->rc_in_persist;
2534                 log.u_bbr.timeStamp = cts;
2535                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2536                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2537                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2538                 log.u_bbr.pacing_gain = rack->r_must_retran;
2539                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2540                     &rack->rc_inp->inp_socket->so_rcv,
2541                     &rack->rc_inp->inp_socket->so_snd,
2542                     BBR_LOG_BBRSND, 0,
2543                     0, &log, false, tv);
2544         }
2545 }
2546
2547 static void
2548 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2549 {
2550         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2551                 union tcp_log_stackspecific log;
2552                 struct timeval tv;
2553
2554                 memset(&log, 0, sizeof(log));
2555                 log.u_bbr.flex1 = did_out;
2556                 log.u_bbr.flex2 = nxt_pkt;
2557                 log.u_bbr.flex3 = way_out;
2558                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2559                 if (rack->rack_no_prr)
2560                         log.u_bbr.flex5 = 0;
2561                 else
2562                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2563                 log.u_bbr.flex6 = nsegs;
2564                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2565                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2566                 log.u_bbr.flex7 <<= 1;
2567                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2568                 log.u_bbr.flex7 <<= 1;
2569                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2570                 log.u_bbr.flex8 = rack->rc_in_persist;
2571                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2572                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2573                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2574                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2575                 log.u_bbr.use_lt_bw <<= 1;
2576                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2577                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2578                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2579                 log.u_bbr.pacing_gain = rack->r_must_retran;
2580                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2581                     &rack->rc_inp->inp_socket->so_rcv,
2582                     &rack->rc_inp->inp_socket->so_snd,
2583                     BBR_LOG_DOSEG_DONE, 0,
2584                     0, &log, false, &tv);
2585         }
2586 }
2587
2588 static void
2589 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2590 {
2591         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2592                 union tcp_log_stackspecific log;
2593                 struct timeval tv;
2594                 uint32_t cts;
2595
2596                 memset(&log, 0, sizeof(log));
2597                 cts = tcp_get_usecs(&tv);
2598                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2599                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2600                 log.u_bbr.flex4 = arg1;
2601                 log.u_bbr.flex5 = arg2;
2602                 log.u_bbr.flex6 = arg3;
2603                 log.u_bbr.flex8 = frm;
2604                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2605                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2606                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2607                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2608                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2609                 log.u_bbr.pacing_gain = rack->r_must_retran;
2610                 TCP_LOG_EVENTP(tp, NULL,
2611                     &tp->t_inpcb->inp_socket->so_rcv,
2612                     &tp->t_inpcb->inp_socket->so_snd,
2613                     TCP_HDWR_PACE_SIZE, 0,
2614                     0, &log, false, &tv);
2615         }
2616 }
2617
2618 static void
2619 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2620                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2621 {
2622         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2623                 union tcp_log_stackspecific log;
2624                 struct timeval tv;
2625
2626                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2627                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2628                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2629                 log.u_bbr.flex1 = slot;
2630                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2631                 log.u_bbr.flex4 = reason;
2632                 if (rack->rack_no_prr)
2633                         log.u_bbr.flex5 = 0;
2634                 else
2635                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2636                 log.u_bbr.flex7 = hpts_calling;
2637                 log.u_bbr.flex8 = rack->rc_in_persist;
2638                 log.u_bbr.lt_epoch = cwnd_to_use;
2639                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2640                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2641                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2642                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2643                 log.u_bbr.pacing_gain = rack->r_must_retran;
2644                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2645                     &rack->rc_inp->inp_socket->so_rcv,
2646                     &rack->rc_inp->inp_socket->so_snd,
2647                     BBR_LOG_JUSTRET, 0,
2648                     tlen, &log, false, &tv);
2649         }
2650 }
2651
2652 static void
2653 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2654                    struct timeval *tv, uint32_t flags_on_entry)
2655 {
2656         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2657                 union tcp_log_stackspecific log;
2658
2659                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2660                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2661                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2662                 log.u_bbr.flex1 = line;
2663                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2664                 log.u_bbr.flex3 = flags_on_entry;
2665                 log.u_bbr.flex4 = us_cts;
2666                 if (rack->rack_no_prr)
2667                         log.u_bbr.flex5 = 0;
2668                 else
2669                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2670                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2671                 log.u_bbr.flex7 = hpts_removed;
2672                 log.u_bbr.flex8 = 1;
2673                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2674                 log.u_bbr.timeStamp = us_cts;
2675                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2676                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2677                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2678                 log.u_bbr.pacing_gain = rack->r_must_retran;
2679                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2680                     &rack->rc_inp->inp_socket->so_rcv,
2681                     &rack->rc_inp->inp_socket->so_snd,
2682                     BBR_LOG_TIMERCANC, 0,
2683                     0, &log, false, tv);
2684         }
2685 }
2686
2687 static void
2688 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2689                           uint32_t flex1, uint32_t flex2,
2690                           uint32_t flex3, uint32_t flex4,
2691                           uint32_t flex5, uint32_t flex6,
2692                           uint16_t flex7, uint8_t mod)
2693 {
2694         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2695                 union tcp_log_stackspecific log;
2696                 struct timeval tv;
2697
2698                 if (mod == 1) {
2699                         /* No you can't use 1, its for the real to cancel */
2700                         return;
2701                 }
2702                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2703                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2704                 log.u_bbr.flex1 = flex1;
2705                 log.u_bbr.flex2 = flex2;
2706                 log.u_bbr.flex3 = flex3;
2707                 log.u_bbr.flex4 = flex4;
2708                 log.u_bbr.flex5 = flex5;
2709                 log.u_bbr.flex6 = flex6;
2710                 log.u_bbr.flex7 = flex7;
2711                 log.u_bbr.flex8 = mod;
2712                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2713                     &rack->rc_inp->inp_socket->so_rcv,
2714                     &rack->rc_inp->inp_socket->so_snd,
2715                     BBR_LOG_TIMERCANC, 0,
2716                     0, &log, false, &tv);
2717         }
2718 }
2719
2720 static void
2721 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2722 {
2723         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2724                 union tcp_log_stackspecific log;
2725                 struct timeval tv;
2726
2727                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2728                 log.u_bbr.flex1 = timers;
2729                 log.u_bbr.flex2 = ret;
2730                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2731                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2732                 log.u_bbr.flex5 = cts;
2733                 if (rack->rack_no_prr)
2734                         log.u_bbr.flex6 = 0;
2735                 else
2736                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2737                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2738                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2739                 log.u_bbr.pacing_gain = rack->r_must_retran;
2740                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2741                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2742                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2743                     &rack->rc_inp->inp_socket->so_rcv,
2744                     &rack->rc_inp->inp_socket->so_snd,
2745                     BBR_LOG_TO_PROCESS, 0,
2746                     0, &log, false, &tv);
2747         }
2748 }
2749
2750 static void
2751 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2752 {
2753         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2754                 union tcp_log_stackspecific log;
2755                 struct timeval tv;
2756
2757                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2758                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2759                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2760                 if (rack->rack_no_prr)
2761                         log.u_bbr.flex3 = 0;
2762                 else
2763                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2764                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2765                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2766                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2767                 log.u_bbr.flex8 = frm;
2768                 log.u_bbr.pkts_out = orig_cwnd;
2769                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2770                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2771                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2772                 log.u_bbr.use_lt_bw <<= 1;
2773                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2774                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2775                     &rack->rc_inp->inp_socket->so_rcv,
2776                     &rack->rc_inp->inp_socket->so_snd,
2777                     BBR_LOG_BBRUPD, 0,
2778                     0, &log, false, &tv);
2779         }
2780 }
2781
2782 #ifdef NETFLIX_EXP_DETECTION
2783 static void
2784 rack_log_sad(struct tcp_rack *rack, int event)
2785 {
2786         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2787                 union tcp_log_stackspecific log;
2788                 struct timeval tv;
2789
2790                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2791                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2792                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2793                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2794                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2795                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2796                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2797                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2798                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2799                 log.u_bbr.lt_epoch |= rack->do_detection;
2800                 log.u_bbr.applimited = tcp_map_minimum;
2801                 log.u_bbr.flex7 = rack->sack_attack_disable;
2802                 log.u_bbr.flex8 = event;
2803                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2804                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2805                 log.u_bbr.delivered = tcp_sad_decay_val;
2806                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2807                     &rack->rc_inp->inp_socket->so_rcv,
2808                     &rack->rc_inp->inp_socket->so_snd,
2809                     TCP_SAD_DETECTION, 0,
2810                     0, &log, false, &tv);
2811         }
2812 }
2813 #endif
2814
2815 static void
2816 rack_counter_destroy(void)
2817 {
2818         int i;
2819
2820         counter_u64_free(rack_fto_send);
2821         counter_u64_free(rack_fto_rsm_send);
2822         counter_u64_free(rack_nfto_resend);
2823         counter_u64_free(rack_hw_pace_init_fail);
2824         counter_u64_free(rack_hw_pace_lost);
2825         counter_u64_free(rack_non_fto_send);
2826         counter_u64_free(rack_extended_rfo);
2827         counter_u64_free(rack_ack_total);
2828         counter_u64_free(rack_express_sack);
2829         counter_u64_free(rack_sack_total);
2830         counter_u64_free(rack_move_none);
2831         counter_u64_free(rack_move_some);
2832         counter_u64_free(rack_sack_attacks_detected);
2833         counter_u64_free(rack_sack_attacks_reversed);
2834         counter_u64_free(rack_sack_used_next_merge);
2835         counter_u64_free(rack_sack_used_prev_merge);
2836         counter_u64_free(rack_badfr);
2837         counter_u64_free(rack_badfr_bytes);
2838         counter_u64_free(rack_rtm_prr_retran);
2839         counter_u64_free(rack_rtm_prr_newdata);
2840         counter_u64_free(rack_timestamp_mismatch);
2841         counter_u64_free(rack_find_high);
2842         counter_u64_free(rack_reorder_seen);
2843         counter_u64_free(rack_tlp_tot);
2844         counter_u64_free(rack_tlp_newdata);
2845         counter_u64_free(rack_tlp_retran);
2846         counter_u64_free(rack_tlp_retran_bytes);
2847         counter_u64_free(rack_tlp_retran_fail);
2848         counter_u64_free(rack_to_tot);
2849         counter_u64_free(rack_to_arm_rack);
2850         counter_u64_free(rack_to_arm_tlp);
2851         counter_u64_free(rack_calc_zero);
2852         counter_u64_free(rack_calc_nonzero);
2853         counter_u64_free(rack_paced_segments);
2854         counter_u64_free(rack_unpaced_segments);
2855         counter_u64_free(rack_saw_enobuf);
2856         counter_u64_free(rack_saw_enobuf_hw);
2857         counter_u64_free(rack_saw_enetunreach);
2858         counter_u64_free(rack_hot_alloc);
2859         counter_u64_free(rack_to_alloc);
2860         counter_u64_free(rack_to_alloc_hard);
2861         counter_u64_free(rack_to_alloc_emerg);
2862         counter_u64_free(rack_to_alloc_limited);
2863         counter_u64_free(rack_alloc_limited_conns);
2864         counter_u64_free(rack_split_limited);
2865         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
2866                 counter_u64_free(rack_proc_comp_ack[i]);
2867         }
2868         counter_u64_free(rack_multi_single_eq);
2869         counter_u64_free(rack_proc_non_comp_ack);
2870         counter_u64_free(rack_sack_proc_all);
2871         counter_u64_free(rack_sack_proc_restart);
2872         counter_u64_free(rack_sack_proc_short);
2873         counter_u64_free(rack_enter_tlp_calc);
2874         counter_u64_free(rack_used_tlpmethod);
2875         counter_u64_free(rack_used_tlpmethod2);
2876         counter_u64_free(rack_sack_skipped_acked);
2877         counter_u64_free(rack_sack_splits);
2878         counter_u64_free(rack_progress_drops);
2879         counter_u64_free(rack_input_idle_reduces);
2880         counter_u64_free(rack_collapsed_win);
2881         counter_u64_free(rack_tlp_does_nada);
2882         counter_u64_free(rack_try_scwnd);
2883         counter_u64_free(rack_per_timer_hole);
2884         counter_u64_free(rack_large_ackcmp);
2885         counter_u64_free(rack_small_ackcmp);
2886 #ifdef INVARIANTS
2887         counter_u64_free(rack_adjust_map_bw);
2888 #endif
2889         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2890         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2891 }
2892
2893 static struct rack_sendmap *
2894 rack_alloc(struct tcp_rack *rack)
2895 {
2896         struct rack_sendmap *rsm;
2897
2898         /*
2899          * First get the top of the list it in
2900          * theory is the "hottest" rsm we have,
2901          * possibly just freed by ack processing.
2902          */
2903         if (rack->rc_free_cnt > rack_free_cache) {
2904                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2905                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2906                 counter_u64_add(rack_hot_alloc, 1);
2907                 rack->rc_free_cnt--;
2908                 return (rsm);
2909         }
2910         /*
2911          * Once we get under our free cache we probably
2912          * no longer have a "hot" one available. Lets
2913          * get one from UMA.
2914          */
2915         rsm = uma_zalloc(rack_zone, M_NOWAIT);
2916         if (rsm) {
2917                 rack->r_ctl.rc_num_maps_alloced++;
2918                 counter_u64_add(rack_to_alloc, 1);
2919                 return (rsm);
2920         }
2921         /*
2922          * Dig in to our aux rsm's (the last two) since
2923          * UMA failed to get us one.
2924          */
2925         if (rack->rc_free_cnt) {
2926                 counter_u64_add(rack_to_alloc_emerg, 1);
2927                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2928                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2929                 rack->rc_free_cnt--;
2930                 return (rsm);
2931         }
2932         return (NULL);
2933 }
2934
2935 static struct rack_sendmap *
2936 rack_alloc_full_limit(struct tcp_rack *rack)
2937 {
2938         if ((V_tcp_map_entries_limit > 0) &&
2939             (rack->do_detection == 0) &&
2940             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2941                 counter_u64_add(rack_to_alloc_limited, 1);
2942                 if (!rack->alloc_limit_reported) {
2943                         rack->alloc_limit_reported = 1;
2944                         counter_u64_add(rack_alloc_limited_conns, 1);
2945                 }
2946                 return (NULL);
2947         }
2948         return (rack_alloc(rack));
2949 }
2950
2951 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2952 static struct rack_sendmap *
2953 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2954 {
2955         struct rack_sendmap *rsm;
2956
2957         if (limit_type) {
2958                 /* currently there is only one limit type */
2959                 if (V_tcp_map_split_limit > 0 &&
2960                     (rack->do_detection == 0) &&
2961                     rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2962                         counter_u64_add(rack_split_limited, 1);
2963                         if (!rack->alloc_limit_reported) {
2964                                 rack->alloc_limit_reported = 1;
2965                                 counter_u64_add(rack_alloc_limited_conns, 1);
2966                         }
2967                         return (NULL);
2968                 }
2969         }
2970
2971         /* allocate and mark in the limit type, if set */
2972         rsm = rack_alloc(rack);
2973         if (rsm != NULL && limit_type) {
2974                 rsm->r_limit_type = limit_type;
2975                 rack->r_ctl.rc_num_split_allocs++;
2976         }
2977         return (rsm);
2978 }
2979
2980 static void
2981 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2982 {
2983         if (rsm->r_flags & RACK_APP_LIMITED) {
2984                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2985                         rack->r_ctl.rc_app_limited_cnt--;
2986                 }
2987         }
2988         if (rsm->r_limit_type) {
2989                 /* currently there is only one limit type */
2990                 rack->r_ctl.rc_num_split_allocs--;
2991         }
2992         if (rsm == rack->r_ctl.rc_first_appl) {
2993                 if (rack->r_ctl.rc_app_limited_cnt == 0)
2994                         rack->r_ctl.rc_first_appl = NULL;
2995                 else {
2996                         /* Follow the next one out */
2997                         struct rack_sendmap fe;
2998
2999                         fe.r_start = rsm->r_nseq_appl;
3000                         rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3001                 }
3002         }
3003         if (rsm == rack->r_ctl.rc_resend)
3004                 rack->r_ctl.rc_resend = NULL;
3005         if (rsm == rack->r_ctl.rc_rsm_at_retran)
3006                 rack->r_ctl.rc_rsm_at_retran = NULL;
3007         if (rsm == rack->r_ctl.rc_end_appl)
3008                 rack->r_ctl.rc_end_appl = NULL;
3009         if (rack->r_ctl.rc_tlpsend == rsm)
3010                 rack->r_ctl.rc_tlpsend = NULL;
3011         if (rack->r_ctl.rc_sacklast == rsm)
3012                 rack->r_ctl.rc_sacklast = NULL;
3013         memset(rsm, 0, sizeof(struct rack_sendmap));
3014         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3015         rack->rc_free_cnt++;
3016 }
3017
3018 static void
3019 rack_free_trim(struct tcp_rack *rack)
3020 {
3021         struct rack_sendmap *rsm;
3022
3023         /*
3024          * Free up all the tail entries until
3025          * we get our list down to the limit.
3026          */
3027         while (rack->rc_free_cnt > rack_free_cache) {
3028                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3029                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3030                 rack->rc_free_cnt--;
3031                 uma_zfree(rack_zone, rsm);
3032         }
3033 }
3034
3035
3036 static uint32_t
3037 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3038 {
3039         uint64_t srtt, bw, len, tim;
3040         uint32_t segsiz, def_len, minl;
3041
3042         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3043         def_len = rack_def_data_window * segsiz;
3044         if (rack->rc_gp_filled == 0) {
3045                 /*
3046                  * We have no measurement (IW is in flight?) so
3047                  * we can only guess using our data_window sysctl
3048                  * value (usually 100MSS).
3049                  */
3050                 return (def_len);
3051         }
3052         /*
3053          * Now we have a number of factors to consider.
3054          *
3055          * 1) We have a desired BDP which is usually
3056          *    at least 2.
3057          * 2) We have a minimum number of rtt's usually 1 SRTT
3058          *    but we allow it too to be more.
3059          * 3) We want to make sure a measurement last N useconds (if
3060          *    we have set rack_min_measure_usec.
3061          *
3062          * We handle the first concern here by trying to create a data
3063          * window of max(rack_def_data_window, DesiredBDP). The
3064          * second concern we handle in not letting the measurement
3065          * window end normally until at least the required SRTT's
3066          * have gone by which is done further below in
3067          * rack_enough_for_measurement(). Finally the third concern
3068          * we also handle here by calculating how long that time
3069          * would take at the current BW and then return the
3070          * max of our first calculation and that length. Note
3071          * that if rack_min_measure_usec is 0, we don't deal
3072          * with concern 3. Also for both Concern 1 and 3 an
3073          * application limited period could end the measurement
3074          * earlier.
3075          *
3076          * So lets calculate the BDP with the "known" b/w using
3077          * the SRTT has our rtt and then multiply it by the
3078          * goal.
3079          */
3080         bw = rack_get_bw(rack);
3081         srtt = (uint64_t)tp->t_srtt;
3082         len = bw * srtt;
3083         len /= (uint64_t)HPTS_USEC_IN_SEC;
3084         len *= max(1, rack_goal_bdp);
3085         /* Now we need to round up to the nearest MSS */
3086         len = roundup(len, segsiz);
3087         if (rack_min_measure_usec) {
3088                 /* Now calculate our min length for this b/w */
3089                 tim = rack_min_measure_usec;
3090                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3091                 if (minl == 0)
3092                         minl = 1;
3093                 minl = roundup(minl, segsiz);
3094                 if (len < minl)
3095                         len = minl;
3096         }
3097         /*
3098          * Now if we have a very small window we want
3099          * to attempt to get the window that is
3100          * as small as possible. This happens on
3101          * low b/w connections and we don't want to
3102          * span huge numbers of rtt's between measurements.
3103          *
3104          * We basically include 2 over our "MIN window" so
3105          * that the measurement can be shortened (possibly) by
3106          * an ack'ed packet.
3107          */
3108         if (len < def_len)
3109                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3110         else
3111                 return (max((uint32_t)len, def_len));
3112
3113 }
3114
3115 static int
3116 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack)
3117 {
3118         uint32_t tim, srtts, segsiz;
3119
3120         /*
3121          * Has enough time passed for the GP measurement to be valid?
3122          */
3123         if ((tp->snd_max == tp->snd_una) ||
3124             (th_ack == tp->snd_max)){
3125                 /* All is acked */
3126                 return (1);
3127         }
3128         if (SEQ_LT(th_ack, tp->gput_seq)) {
3129                 /* Not enough bytes yet */
3130                 return (0);
3131         }
3132         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3133         if (SEQ_LT(th_ack, tp->gput_ack) &&
3134             ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3135                 /* Not enough bytes yet */
3136                 return (0);
3137         }
3138         if (rack->r_ctl.rc_first_appl &&
3139             (rack->r_ctl.rc_first_appl->r_start == th_ack)) {
3140                 /*
3141                  * We are up to the app limited point
3142                  * we have to measure irrespective of the time..
3143                  */
3144                 return (1);
3145         }
3146         /* Now what about time? */
3147         srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3148         tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3149         if (tim >= srtts) {
3150                 return (1);
3151         }
3152         /* Nope not even a full SRTT has passed */
3153         return (0);
3154 }
3155
3156 static void
3157 rack_log_timely(struct tcp_rack *rack,
3158                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3159                 uint64_t up_bnd, int line, uint8_t method)
3160 {
3161         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3162                 union tcp_log_stackspecific log;
3163                 struct timeval tv;
3164
3165                 memset(&log, 0, sizeof(log));
3166                 log.u_bbr.flex1 = logged;
3167                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3168                 log.u_bbr.flex2 <<= 4;
3169                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3170                 log.u_bbr.flex2 <<= 4;
3171                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3172                 log.u_bbr.flex2 <<= 4;
3173                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3174                 log.u_bbr.flex3 = rack->rc_gp_incr;
3175                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3176                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3177                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3178                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3179                 log.u_bbr.flex8 = method;
3180                 log.u_bbr.cur_del_rate = cur_bw;
3181                 log.u_bbr.delRate = low_bnd;
3182                 log.u_bbr.bw_inuse = up_bnd;
3183                 log.u_bbr.rttProp = rack_get_bw(rack);
3184                 log.u_bbr.pkt_epoch = line;
3185                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3186                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3187                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3188                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3189                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3190                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3191                 log.u_bbr.cwnd_gain <<= 1;
3192                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3193                 log.u_bbr.cwnd_gain <<= 1;
3194                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3195                 log.u_bbr.cwnd_gain <<= 1;
3196                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3197                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3198                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3199                     &rack->rc_inp->inp_socket->so_rcv,
3200                     &rack->rc_inp->inp_socket->so_snd,
3201                     TCP_TIMELY_WORK, 0,
3202                     0, &log, false, &tv);
3203         }
3204 }
3205
3206 static int
3207 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3208 {
3209         /*
3210          * Before we increase we need to know if
3211          * the estimate just made was less than
3212          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3213          *
3214          * If we already are pacing at a fast enough
3215          * rate to push us faster there is no sense of
3216          * increasing.
3217          *
3218          * We first caculate our actual pacing rate (ss or ca multipler
3219          * times our cur_bw).
3220          *
3221          * Then we take the last measured rate and multipy by our
3222          * maximum pacing overage to give us a max allowable rate.
3223          *
3224          * If our act_rate is smaller than our max_allowable rate
3225          * then we should increase. Else we should hold steady.
3226          *
3227          */
3228         uint64_t act_rate, max_allow_rate;
3229
3230         if (rack_timely_no_stopping)
3231                 return (1);
3232
3233         if ((cur_bw == 0) || (last_bw_est == 0)) {
3234                 /*
3235                  * Initial startup case or
3236                  * everything is acked case.
3237                  */
3238                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3239                                 __LINE__, 9);
3240                 return (1);
3241         }
3242         if (mult <= 100) {
3243                 /*
3244                  * We can always pace at or slightly above our rate.
3245                  */
3246                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3247                                 __LINE__, 9);
3248                 return (1);
3249         }
3250         act_rate = cur_bw * (uint64_t)mult;
3251         act_rate /= 100;
3252         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3253         max_allow_rate /= 100;
3254         if (act_rate < max_allow_rate) {
3255                 /*
3256                  * Here the rate we are actually pacing at
3257                  * is smaller than 10% above our last measurement.
3258                  * This means we are pacing below what we would
3259                  * like to try to achieve (plus some wiggle room).
3260                  */
3261                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3262                                 __LINE__, 9);
3263                 return (1);
3264         } else {
3265                 /*
3266                  * Here we are already pacing at least rack_max_per_above(10%)
3267                  * what we are getting back. This indicates most likely
3268                  * that we are being limited (cwnd/rwnd/app) and can't
3269                  * get any more b/w. There is no sense of trying to
3270                  * raise up the pacing rate its not speeding us up
3271                  * and we already are pacing faster than we are getting.
3272                  */
3273                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3274                                 __LINE__, 8);
3275                 return (0);
3276         }
3277 }
3278
3279 static void
3280 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3281 {
3282         /*
3283          * When we drag bottom, we want to assure
3284          * that no multiplier is below 1.0, if so
3285          * we want to restore it to at least that.
3286          */
3287         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3288                 /* This is unlikely we usually do not touch recovery */
3289                 rack->r_ctl.rack_per_of_gp_rec = 100;
3290         }
3291         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3292                 rack->r_ctl.rack_per_of_gp_ca = 100;
3293         }
3294         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3295                 rack->r_ctl.rack_per_of_gp_ss = 100;
3296         }
3297 }
3298
3299 static void
3300 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3301 {
3302         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3303                 rack->r_ctl.rack_per_of_gp_ca = 100;
3304         }
3305         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3306                 rack->r_ctl.rack_per_of_gp_ss = 100;
3307         }
3308 }
3309
3310 static void
3311 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3312 {
3313         int32_t  calc, logged, plus;
3314
3315         logged = 0;
3316
3317         if (override) {
3318                 /*
3319                  * override is passed when we are
3320                  * loosing b/w and making one last
3321                  * gasp at trying to not loose out
3322                  * to a new-reno flow.
3323                  */
3324                 goto extra_boost;
3325         }
3326         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3327         if (rack->rc_gp_incr &&
3328             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3329                 /*
3330                  * Reset and get 5 strokes more before the boost. Note
3331                  * that the count is 0 based so we have to add one.
3332                  */
3333 extra_boost:
3334                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3335                 rack->rc_gp_timely_inc_cnt = 0;
3336         } else
3337                 plus = (uint32_t)rack_gp_increase_per;
3338         /* Must be at least 1% increase for true timely increases */
3339         if ((plus < 1) &&
3340             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3341                 plus = 1;
3342         if (rack->rc_gp_saw_rec &&
3343             (rack->rc_gp_no_rec_chg == 0) &&
3344             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3345                                   rack->r_ctl.rack_per_of_gp_rec)) {
3346                 /* We have been in recovery ding it too */
3347                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3348                 if (calc > 0xffff)
3349                         calc = 0xffff;
3350                 logged |= 1;
3351                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3352                 if (rack_per_upper_bound_ss &&
3353                     (rack->rc_dragged_bottom == 0) &&
3354                     (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3355                         rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3356         }
3357         if (rack->rc_gp_saw_ca &&
3358             (rack->rc_gp_saw_ss == 0) &&
3359             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3360                                   rack->r_ctl.rack_per_of_gp_ca)) {
3361                 /* In CA */
3362                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3363                 if (calc > 0xffff)
3364                         calc = 0xffff;
3365                 logged |= 2;
3366                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3367                 if (rack_per_upper_bound_ca &&
3368                     (rack->rc_dragged_bottom == 0) &&
3369                     (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3370                         rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3371         }
3372         if (rack->rc_gp_saw_ss &&
3373             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3374                                   rack->r_ctl.rack_per_of_gp_ss)) {
3375                 /* In SS */
3376                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3377                 if (calc > 0xffff)
3378                         calc = 0xffff;
3379                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3380                 if (rack_per_upper_bound_ss &&
3381                     (rack->rc_dragged_bottom == 0) &&
3382                     (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3383                         rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3384                 logged |= 4;
3385         }
3386         if (logged &&
3387             (rack->rc_gp_incr == 0)){
3388                 /* Go into increment mode */
3389                 rack->rc_gp_incr = 1;
3390                 rack->rc_gp_timely_inc_cnt = 0;
3391         }
3392         if (rack->rc_gp_incr &&
3393             logged &&
3394             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3395                 rack->rc_gp_timely_inc_cnt++;
3396         }
3397         rack_log_timely(rack,  logged, plus, 0, 0,
3398                         __LINE__, 1);
3399 }
3400
3401 static uint32_t
3402 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3403 {
3404         /*
3405          * norm_grad = rtt_diff / minrtt;
3406          * new_per = curper * (1 - B * norm_grad)
3407          *
3408          * B = rack_gp_decrease_per (default 10%)
3409          * rtt_dif = input var current rtt-diff
3410          * curper = input var current percentage
3411          * minrtt = from rack filter
3412          *
3413          */
3414         uint64_t perf;
3415
3416         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3417                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3418                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3419                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3420                      (uint64_t)1000000)) /
3421                 (uint64_t)1000000);
3422         if (perf > curper) {
3423                 /* TSNH */
3424                 perf = curper - 1;
3425         }
3426         return ((uint32_t)perf);
3427 }
3428
3429 static uint32_t
3430 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3431 {
3432         /*
3433          *                                   highrttthresh
3434          * result = curper * (1 - (B * ( 1 -  ------          ))
3435          *                                     gp_srtt
3436          *
3437          * B = rack_gp_decrease_per (default 10%)
3438          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3439          */
3440         uint64_t perf;
3441         uint32_t highrttthresh;
3442
3443         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3444
3445         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3446                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3447                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3448                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3449         return (perf);
3450 }
3451
3452 static void
3453 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3454 {
3455         uint64_t logvar, logvar2, logvar3;
3456         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3457
3458         if (rack->rc_gp_incr) {
3459                 /* Turn off increment counting */
3460                 rack->rc_gp_incr = 0;
3461                 rack->rc_gp_timely_inc_cnt = 0;
3462         }
3463         ss_red = ca_red = rec_red = 0;
3464         logged = 0;
3465         /* Calculate the reduction value */
3466         if (rtt_diff < 0) {
3467                 rtt_diff *= -1;
3468         }
3469         /* Must be at least 1% reduction */
3470         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3471                 /* We have been in recovery ding it too */
3472                 if (timely_says == 2) {
3473                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3474                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3475                         if (alt < new_per)
3476                                 val = alt;
3477                         else
3478                                 val = new_per;
3479                 } else
3480                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3481                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3482                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3483                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3484                 } else {
3485                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3486                         rec_red = 0;
3487                 }
3488                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3489                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3490                 logged |= 1;
3491         }
3492         if (rack->rc_gp_saw_ss) {
3493                 /* Sent in SS */
3494                 if (timely_says == 2) {
3495                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3496                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3497                         if (alt < new_per)
3498                                 val = alt;
3499                         else
3500                                 val = new_per;
3501                 } else
3502                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3503                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3504                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3505                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3506                 } else {
3507                         ss_red = new_per;
3508                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3509                         logvar = new_per;
3510                         logvar <<= 32;
3511                         logvar |= alt;
3512                         logvar2 = (uint32_t)rtt;
3513                         logvar2 <<= 32;
3514                         logvar2 |= (uint32_t)rtt_diff;
3515                         logvar3 = rack_gp_rtt_maxmul;
3516                         logvar3 <<= 32;
3517                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3518                         rack_log_timely(rack, timely_says,
3519                                         logvar2, logvar3,
3520                                         logvar, __LINE__, 10);
3521                 }
3522                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3523                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3524                 logged |= 4;
3525         } else if (rack->rc_gp_saw_ca) {
3526                 /* Sent in CA */
3527                 if (timely_says == 2) {
3528                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3529                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3530                         if (alt < new_per)
3531                                 val = alt;
3532                         else
3533                                 val = new_per;
3534                 } else
3535                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3536                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3537                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3538                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3539                 } else {
3540                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3541                         ca_red = 0;
3542                         logvar = new_per;
3543                         logvar <<= 32;
3544                         logvar |= alt;
3545                         logvar2 = (uint32_t)rtt;
3546                         logvar2 <<= 32;
3547                         logvar2 |= (uint32_t)rtt_diff;
3548                         logvar3 = rack_gp_rtt_maxmul;
3549                         logvar3 <<= 32;
3550                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3551                         rack_log_timely(rack, timely_says,
3552                                         logvar2, logvar3,
3553                                         logvar, __LINE__, 10);
3554                 }
3555                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3556                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3557                 logged |= 2;
3558         }
3559         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3560                 rack->rc_gp_timely_dec_cnt++;
3561                 if (rack_timely_dec_clear &&
3562                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3563                         rack->rc_gp_timely_dec_cnt = 0;
3564         }
3565         logvar = ss_red;
3566         logvar <<= 32;
3567         logvar |= ca_red;
3568         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3569                         __LINE__, 2);
3570 }
3571
3572 static void
3573 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3574                      uint32_t rtt, uint32_t line, uint8_t reas)
3575 {
3576         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3577                 union tcp_log_stackspecific log;
3578                 struct timeval tv;
3579
3580                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3581                 log.u_bbr.flex1 = line;
3582                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3583                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3584                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3585                 log.u_bbr.flex5 = rtt;
3586                 log.u_bbr.flex6 = rack->rc_highly_buffered;
3587                 log.u_bbr.flex6 <<= 1;
3588                 log.u_bbr.flex6 |= rack->forced_ack;
3589                 log.u_bbr.flex6 <<= 1;
3590                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3591                 log.u_bbr.flex6 <<= 1;
3592                 log.u_bbr.flex6 |= rack->in_probe_rtt;
3593                 log.u_bbr.flex6 <<= 1;
3594                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3595                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3596                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3597                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3598                 log.u_bbr.flex8 = reas;
3599                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3600                 log.u_bbr.delRate = rack_get_bw(rack);
3601                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3602                 log.u_bbr.cur_del_rate <<= 32;
3603                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3604                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3605                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3606                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3607                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3608                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3609                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3610                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3611                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3612                 log.u_bbr.rttProp = us_cts;
3613                 log.u_bbr.rttProp <<= 32;
3614                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3615                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3616                     &rack->rc_inp->inp_socket->so_rcv,
3617                     &rack->rc_inp->inp_socket->so_snd,
3618                     BBR_LOG_RTT_SHRINKS, 0,
3619                     0, &log, false, &rack->r_ctl.act_rcv_time);
3620         }
3621 }
3622
3623 static void
3624 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3625 {
3626         uint64_t bwdp;
3627
3628         bwdp = rack_get_bw(rack);
3629         bwdp *= (uint64_t)rtt;
3630         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3631         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3632         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3633                 /*
3634                  * A window protocol must be able to have 4 packets
3635                  * outstanding as the floor in order to function
3636                  * (especially considering delayed ack :D).
3637                  */
3638                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3639         }
3640 }
3641
3642 static void
3643 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3644 {
3645         /**
3646          * ProbeRTT is a bit different in rack_pacing than in
3647          * BBR. It is like BBR in that it uses the lowering of
3648          * the RTT as a signal that we saw something new and
3649          * counts from there for how long between. But it is
3650          * different in that its quite simple. It does not
3651          * play with the cwnd and wait until we get down
3652          * to N segments outstanding and hold that for
3653          * 200ms. Instead it just sets the pacing reduction
3654          * rate to a set percentage (70 by default) and hold
3655          * that for a number of recent GP Srtt's.
3656          */
3657         uint32_t segsiz;
3658
3659         if (rack->rc_gp_dyn_mul == 0)
3660                 return;
3661
3662         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3663                 /* We are idle */
3664                 return;
3665         }
3666         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3667             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3668                 /*
3669                  * Stop the goodput now, the idea here is
3670                  * that future measurements with in_probe_rtt
3671                  * won't register if they are not greater so
3672                  * we want to get what info (if any) is available
3673                  * now.
3674                  */
3675                 rack_do_goodput_measurement(rack->rc_tp, rack,
3676                                             rack->rc_tp->snd_una, __LINE__);
3677         }
3678         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3679         rack->r_ctl.rc_time_probertt_entered = us_cts;
3680         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3681                      rack->r_ctl.rc_pace_min_segs);
3682         rack->in_probe_rtt = 1;
3683         rack->measure_saw_probe_rtt = 1;
3684         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3685         rack->r_ctl.rc_time_probertt_starts = 0;
3686         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3687         if (rack_probertt_use_min_rtt_entry)
3688                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3689         else
3690                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3691         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3692                              __LINE__, RACK_RTTS_ENTERPROBE);
3693 }
3694
3695 static void
3696 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3697 {
3698         struct rack_sendmap *rsm;
3699         uint32_t segsiz;
3700
3701         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3702                      rack->r_ctl.rc_pace_min_segs);
3703         rack->in_probe_rtt = 0;
3704         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3705             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3706                 /*
3707                  * Stop the goodput now, the idea here is
3708                  * that future measurements with in_probe_rtt
3709                  * won't register if they are not greater so
3710                  * we want to get what info (if any) is available
3711                  * now.
3712                  */
3713                 rack_do_goodput_measurement(rack->rc_tp, rack,
3714                                             rack->rc_tp->snd_una, __LINE__);
3715         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3716                 /*
3717                  * We don't have enough data to make a measurement.
3718                  * So lets just stop and start here after exiting
3719                  * probe-rtt. We probably are not interested in
3720                  * the results anyway.
3721                  */
3722                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3723         }
3724         /*
3725          * Measurements through the current snd_max are going
3726          * to be limited by the slower pacing rate.
3727          *
3728          * We need to mark these as app-limited so we
3729          * don't collapse the b/w.
3730          */
3731         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3732         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3733                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3734                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3735                 else {
3736                         /*
3737                          * Go out to the end app limited and mark
3738                          * this new one as next and move the end_appl up
3739                          * to this guy.
3740                          */
3741                         if (rack->r_ctl.rc_end_appl)
3742                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3743                         rack->r_ctl.rc_end_appl = rsm;
3744                 }
3745                 rsm->r_flags |= RACK_APP_LIMITED;
3746                 rack->r_ctl.rc_app_limited_cnt++;
3747         }
3748         /*
3749          * Now, we need to examine our pacing rate multipliers.
3750          * If its under 100%, we need to kick it back up to
3751          * 100%. We also don't let it be over our "max" above
3752          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3753          * Note setting clamp_atexit_prtt to 0 has the effect
3754          * of setting CA/SS to 100% always at exit (which is
3755          * the default behavior).
3756          */
3757         if (rack_probertt_clear_is) {
3758                 rack->rc_gp_incr = 0;
3759                 rack->rc_gp_bwred = 0;
3760                 rack->rc_gp_timely_inc_cnt = 0;
3761                 rack->rc_gp_timely_dec_cnt = 0;
3762         }
3763         /* Do we do any clamping at exit? */
3764         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3765                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3766                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3767         }
3768         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3769                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3770                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3771         }
3772         /*
3773          * Lets set rtt_diff to 0, so that we will get a "boost"
3774          * after exiting.
3775          */
3776         rack->r_ctl.rc_rtt_diff = 0;
3777
3778         /* Clear all flags so we start fresh */
3779         rack->rc_tp->t_bytes_acked = 0;
3780         rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3781         /*
3782          * If configured to, set the cwnd and ssthresh to
3783          * our targets.
3784          */
3785         if (rack_probe_rtt_sets_cwnd) {
3786                 uint64_t ebdp;
3787                 uint32_t setto;
3788
3789                 /* Set ssthresh so we get into CA once we hit our target */
3790                 if (rack_probertt_use_min_rtt_exit == 1) {
3791                         /* Set to min rtt */
3792                         rack_set_prtt_target(rack, segsiz,
3793                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3794                 } else if (rack_probertt_use_min_rtt_exit == 2) {
3795                         /* Set to current gp rtt */
3796                         rack_set_prtt_target(rack, segsiz,
3797                                              rack->r_ctl.rc_gp_srtt);
3798                 } else if (rack_probertt_use_min_rtt_exit == 3) {
3799                         /* Set to entry gp rtt */
3800                         rack_set_prtt_target(rack, segsiz,
3801                                              rack->r_ctl.rc_entry_gp_rtt);
3802                 } else {
3803                         uint64_t sum;
3804                         uint32_t setval;
3805
3806                         sum = rack->r_ctl.rc_entry_gp_rtt;
3807                         sum *= 10;
3808                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3809                         if (sum >= 20) {
3810                                 /*
3811                                  * A highly buffered path needs
3812                                  * cwnd space for timely to work.
3813                                  * Lets set things up as if
3814                                  * we are heading back here again.
3815                                  */
3816                                 setval = rack->r_ctl.rc_entry_gp_rtt;
3817                         } else if (sum >= 15) {
3818                                 /*
3819                                  * Lets take the smaller of the
3820                                  * two since we are just somewhat
3821                                  * buffered.
3822                                  */
3823                                 setval = rack->r_ctl.rc_gp_srtt;
3824                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3825                                         setval = rack->r_ctl.rc_entry_gp_rtt;
3826                         } else {
3827                                 /*
3828                                  * Here we are not highly buffered
3829                                  * and should pick the min we can to
3830                                  * keep from causing loss.
3831                                  */
3832                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3833                         }
3834                         rack_set_prtt_target(rack, segsiz,
3835                                              setval);
3836                 }
3837                 if (rack_probe_rtt_sets_cwnd > 1) {
3838                         /* There is a percentage here to boost */
3839                         ebdp = rack->r_ctl.rc_target_probertt_flight;
3840                         ebdp *= rack_probe_rtt_sets_cwnd;
3841                         ebdp /= 100;
3842                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3843                 } else
3844                         setto = rack->r_ctl.rc_target_probertt_flight;
3845                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3846                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3847                         /* Enforce a min */
3848                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3849                 }
3850                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3851                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3852         }
3853         rack_log_rtt_shrinks(rack,  us_cts,
3854                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3855                              __LINE__, RACK_RTTS_EXITPROBE);
3856         /* Clear times last so log has all the info */
3857         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3858         rack->r_ctl.rc_time_probertt_entered = us_cts;
3859         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3860         rack->r_ctl.rc_time_of_last_probertt = us_cts;
3861 }
3862
3863 static void
3864 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3865 {
3866         /* Check in on probe-rtt */
3867         if (rack->rc_gp_filled == 0) {
3868                 /* We do not do p-rtt unless we have gp measurements */
3869                 return;
3870         }
3871         if (rack->in_probe_rtt) {
3872                 uint64_t no_overflow;
3873                 uint32_t endtime, must_stay;
3874
3875                 if (rack->r_ctl.rc_went_idle_time &&
3876                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3877                         /*
3878                          * We went idle during prtt, just exit now.
3879                          */
3880                         rack_exit_probertt(rack, us_cts);
3881                 } else if (rack_probe_rtt_safety_val &&
3882                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3883                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3884                         /*
3885                          * Probe RTT safety value triggered!
3886                          */
3887                         rack_log_rtt_shrinks(rack,  us_cts,
3888                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3889                                              __LINE__, RACK_RTTS_SAFETY);
3890                         rack_exit_probertt(rack, us_cts);
3891                 }
3892                 /* Calculate the max we will wait */
3893                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3894                 if (rack->rc_highly_buffered)
3895                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3896                 /* Calculate the min we must wait */
3897                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3898                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3899                     TSTMP_LT(us_cts, endtime)) {
3900                         uint32_t calc;
3901                         /* Do we lower more? */
3902 no_exit:
3903                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3904                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3905                         else
3906                                 calc = 0;
3907                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3908                         if (calc) {
3909                                 /* Maybe */
3910                                 calc *= rack_per_of_gp_probertt_reduce;
3911                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3912                                 /* Limit it too */
3913                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3914                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3915                         }
3916                         /* We must reach target or the time set */
3917                         return;
3918                 }
3919                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3920                         if ((TSTMP_LT(us_cts, must_stay) &&
3921                              rack->rc_highly_buffered) ||
3922                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3923                               rack->r_ctl.rc_target_probertt_flight)) {
3924                                 /* We are not past the must_stay time */
3925                                 goto no_exit;
3926                         }
3927                         rack_log_rtt_shrinks(rack,  us_cts,
3928                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3929                                              __LINE__, RACK_RTTS_REACHTARGET);
3930                         rack->r_ctl.rc_time_probertt_starts = us_cts;
3931                         if (rack->r_ctl.rc_time_probertt_starts == 0)
3932                                 rack->r_ctl.rc_time_probertt_starts = 1;
3933                         /* Restore back to our rate we want to pace at in prtt */
3934                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3935                 }
3936                 /*
3937                  * Setup our end time, some number of gp_srtts plus 200ms.
3938                  */
3939                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3940                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3941                 if (rack_probertt_gpsrtt_cnt_div)
3942                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3943                 else
3944                         endtime = 0;
3945                 endtime += rack_min_probertt_hold;
3946                 endtime += rack->r_ctl.rc_time_probertt_starts;
3947                 if (TSTMP_GEQ(us_cts,  endtime)) {
3948                         /* yes, exit probertt */
3949                         rack_exit_probertt(rack, us_cts);
3950                 }
3951
3952         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3953                 /* Go into probertt, its been too long since we went lower */
3954                 rack_enter_probertt(rack, us_cts);
3955         }
3956 }
3957
3958 static void
3959 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3960                        uint32_t rtt, int32_t rtt_diff)
3961 {
3962         uint64_t cur_bw, up_bnd, low_bnd, subfr;
3963         uint32_t losses;
3964
3965         if ((rack->rc_gp_dyn_mul == 0) ||
3966             (rack->use_fixed_rate) ||
3967             (rack->in_probe_rtt) ||
3968             (rack->rc_always_pace == 0)) {
3969                 /* No dynamic GP multipler in play */
3970                 return;
3971         }
3972         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3973         cur_bw = rack_get_bw(rack);
3974         /* Calculate our up and down range */
3975         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3976         up_bnd /= 100;
3977         up_bnd += rack->r_ctl.last_gp_comp_bw;
3978
3979         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3980         subfr /= 100;
3981         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3982         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3983                 /*
3984                  * This is the case where our RTT is above
3985                  * the max target and we have been configured
3986                  * to just do timely no bonus up stuff in that case.
3987                  *
3988                  * There are two configurations, set to 1, and we
3989                  * just do timely if we are over our max. If its
3990                  * set above 1 then we slam the multipliers down
3991                  * to 100 and then decrement per timely.
3992                  */
3993                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3994                                 __LINE__, 3);
3995                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3996                         rack_validate_multipliers_at_or_below_100(rack);
3997                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3998         } else if ((last_bw_est < low_bnd) && !losses) {
3999                 /*
4000                  * We are decreasing this is a bit complicated this
4001                  * means we are loosing ground. This could be
4002                  * because another flow entered and we are competing
4003                  * for b/w with it. This will push the RTT up which
4004                  * makes timely unusable unless we want to get shoved
4005                  * into a corner and just be backed off (the age
4006                  * old problem with delay based CC).
4007                  *
4008                  * On the other hand if it was a route change we
4009                  * would like to stay somewhat contained and not
4010                  * blow out the buffers.
4011                  */
4012                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4013                                 __LINE__, 3);
4014                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4015                 if (rack->rc_gp_bwred == 0) {
4016                         /* Go into reduction counting */
4017                         rack->rc_gp_bwred = 1;
4018                         rack->rc_gp_timely_dec_cnt = 0;
4019                 }
4020                 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
4021                     (timely_says == 0)) {
4022                         /*
4023                          * Push another time with a faster pacing
4024                          * to try to gain back (we include override to
4025                          * get a full raise factor).
4026                          */
4027                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4028                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4029                             (timely_says == 0) ||
4030                             (rack_down_raise_thresh == 0)) {
4031                                 /*
4032                                  * Do an override up in b/w if we were
4033                                  * below the threshold or if the threshold
4034                                  * is zero we always do the raise.
4035                                  */
4036                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4037                         } else {
4038                                 /* Log it stays the same */
4039                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4040                                                 __LINE__, 11);
4041                         }
4042                         rack->rc_gp_timely_dec_cnt++;
4043                         /* We are not incrementing really no-count */
4044                         rack->rc_gp_incr = 0;
4045                         rack->rc_gp_timely_inc_cnt = 0;
4046                 } else {
4047                         /*
4048                          * Lets just use the RTT
4049                          * information and give up
4050                          * pushing.
4051                          */
4052                         goto use_timely;
4053                 }
4054         } else if ((timely_says != 2) &&
4055                     !losses &&
4056                     (last_bw_est > up_bnd)) {
4057                 /*
4058                  * We are increasing b/w lets keep going, updating
4059                  * our b/w and ignoring any timely input, unless
4060                  * of course we are at our max raise (if there is one).
4061                  */
4062
4063                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4064                                 __LINE__, 3);
4065                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4066                 if (rack->rc_gp_saw_ss &&
4067                     rack_per_upper_bound_ss &&
4068                      (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
4069                             /*
4070                              * In cases where we can't go higher
4071                              * we should just use timely.
4072                              */
4073                             goto use_timely;
4074                 }
4075                 if (rack->rc_gp_saw_ca &&
4076                     rack_per_upper_bound_ca &&
4077                     (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4078                             /*
4079                              * In cases where we can't go higher
4080                              * we should just use timely.
4081                              */
4082                             goto use_timely;
4083                 }
4084                 rack->rc_gp_bwred = 0;
4085                 rack->rc_gp_timely_dec_cnt = 0;
4086                 /* You get a set number of pushes if timely is trying to reduce */
4087                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4088                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4089                 } else {
4090                         /* Log it stays the same */
4091                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4092                             __LINE__, 12);
4093                 }
4094                 return;
4095         } else {
4096                 /*
4097                  * We are staying between the lower and upper range bounds
4098                  * so use timely to decide.
4099                  */
4100                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4101                                 __LINE__, 3);
4102 use_timely:
4103                 if (timely_says) {
4104                         rack->rc_gp_incr = 0;
4105                         rack->rc_gp_timely_inc_cnt = 0;
4106                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4107                             !losses &&
4108                             (last_bw_est < low_bnd)) {
4109                                 /* We are loosing ground */
4110                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4111                                 rack->rc_gp_timely_dec_cnt++;
4112                                 /* We are not incrementing really no-count */
4113                                 rack->rc_gp_incr = 0;
4114                                 rack->rc_gp_timely_inc_cnt = 0;
4115                         } else
4116                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4117                 } else {
4118                         rack->rc_gp_bwred = 0;
4119                         rack->rc_gp_timely_dec_cnt = 0;
4120                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4121                 }
4122         }
4123 }
4124
4125 static int32_t
4126 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4127 {
4128         int32_t timely_says;
4129         uint64_t log_mult, log_rtt_a_diff;
4130
4131         log_rtt_a_diff = rtt;
4132         log_rtt_a_diff <<= 32;
4133         log_rtt_a_diff |= (uint32_t)rtt_diff;
4134         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4135                     rack_gp_rtt_maxmul)) {
4136                 /* Reduce the b/w multipler */
4137                 timely_says = 2;
4138                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4139                 log_mult <<= 32;
4140                 log_mult |= prev_rtt;
4141                 rack_log_timely(rack,  timely_says, log_mult,
4142                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4143                                 log_rtt_a_diff, __LINE__, 4);
4144         } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4145                            ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4146                             max(rack_gp_rtt_mindiv , 1)))) {
4147                 /* Increase the b/w multipler */
4148                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4149                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4150                          max(rack_gp_rtt_mindiv , 1));
4151                 log_mult <<= 32;
4152                 log_mult |= prev_rtt;
4153                 timely_says = 0;
4154                 rack_log_timely(rack,  timely_says, log_mult ,
4155                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4156                                 log_rtt_a_diff, __LINE__, 5);
4157         } else {
4158                 /*
4159                  * Use a gradient to find it the timely gradient
4160                  * is:
4161                  * grad = rc_rtt_diff / min_rtt;
4162                  *
4163                  * anything below or equal to 0 will be
4164                  * a increase indication. Anything above
4165                  * zero is a decrease. Note we take care
4166                  * of the actual gradient calculation
4167                  * in the reduction (its not needed for
4168                  * increase).
4169                  */
4170                 log_mult = prev_rtt;
4171                 if (rtt_diff <= 0) {
4172                         /*
4173                          * Rttdiff is less than zero, increase the
4174                          * b/w multipler (its 0 or negative)
4175                          */
4176                         timely_says = 0;
4177                         rack_log_timely(rack,  timely_says, log_mult,
4178                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4179                 } else {
4180                         /* Reduce the b/w multipler */
4181                         timely_says = 1;
4182                         rack_log_timely(rack,  timely_says, log_mult,
4183                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4184                 }
4185         }
4186         return (timely_says);
4187 }
4188
4189 static void
4190 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4191                             tcp_seq th_ack, int line)
4192 {
4193         uint64_t tim, bytes_ps, ltim, stim, utim;
4194         uint32_t segsiz, bytes, reqbytes, us_cts;
4195         int32_t gput, new_rtt_diff, timely_says;
4196         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4197         int did_add = 0;
4198
4199         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4200         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4201         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4202                 tim = us_cts - tp->gput_ts;
4203         else
4204                 tim = 0;
4205
4206         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4207                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4208         else
4209                 stim = 0;
4210         /*
4211          * Use the larger of the send time or ack time. This prevents us
4212          * from being influenced by ack artifacts to come up with too
4213          * high of measurement. Note that since we are spanning over many more
4214          * bytes in most of our measurements hopefully that is less likely to
4215          * occur.
4216          */
4217         if (tim > stim)
4218                 utim = max(tim, 1);
4219         else
4220                 utim = max(stim, 1);
4221         /* Lets get a msec time ltim too for the old stuff */
4222         ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4223         gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4224         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4225         if ((tim == 0) && (stim == 0)) {
4226                 /*
4227                  * Invalid measurement time, maybe
4228                  * all on one ack/one send?
4229                  */
4230                 bytes = 0;
4231                 bytes_ps = 0;
4232                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4233                                            0, 0, 0, 10, __LINE__, NULL);
4234                 goto skip_measurement;
4235         }
4236         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4237                 /* We never made a us_rtt measurement? */
4238                 bytes = 0;
4239                 bytes_ps = 0;
4240                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4241                                            0, 0, 0, 10, __LINE__, NULL);
4242                 goto skip_measurement;
4243         }
4244         /*
4245          * Calculate the maximum possible b/w this connection
4246          * could have. We base our calculation on the lowest
4247          * rtt we have seen during the measurement and the
4248          * largest rwnd the client has given us in that time. This
4249          * forms a BDP that is the maximum that we could ever
4250          * get to the client. Anything larger is not valid.
4251          *
4252          * I originally had code here that rejected measurements
4253          * where the time was less than 1/2 the latest us_rtt.
4254          * But after thinking on that I realized its wrong since
4255          * say you had a 150Mbps or even 1Gbps link, and you
4256          * were a long way away.. example I am in Europe (100ms rtt)
4257          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4258          * bytes my time would be 1.2ms, and yet my rtt would say
4259          * the measurement was invalid the time was < 50ms. The
4260          * same thing is true for 150Mb (8ms of time).
4261          *
4262          * A better way I realized is to look at what the maximum
4263          * the connection could possibly do. This is gated on
4264          * the lowest RTT we have seen and the highest rwnd.
4265          * We should in theory never exceed that, if we are
4266          * then something on the path is storing up packets
4267          * and then feeding them all at once to our endpoint
4268          * messing up our measurement.
4269          */
4270         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4271         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4272         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4273         if (SEQ_LT(th_ack, tp->gput_seq)) {
4274                 /* No measurement can be made */
4275                 bytes = 0;
4276                 bytes_ps = 0;
4277                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4278                                            0, 0, 0, 10, __LINE__, NULL);
4279                 goto skip_measurement;
4280         } else
4281                 bytes = (th_ack - tp->gput_seq);
4282         bytes_ps = (uint64_t)bytes;
4283         /*
4284          * Don't measure a b/w for pacing unless we have gotten at least
4285          * an initial windows worth of data in this measurement interval.
4286          *
4287          * Small numbers of bytes get badly influenced by delayed ack and
4288          * other artifacts. Note we take the initial window or our
4289          * defined minimum GP (defaulting to 10 which hopefully is the
4290          * IW).
4291          */
4292         if (rack->rc_gp_filled == 0) {
4293                 /*
4294                  * The initial estimate is special. We
4295                  * have blasted out an IW worth of packets
4296                  * without a real valid ack ts results. We
4297                  * then setup the app_limited_needs_set flag,
4298                  * this should get the first ack in (probably 2
4299                  * MSS worth) to be recorded as the timestamp.
4300                  * We thus allow a smaller number of bytes i.e.
4301                  * IW - 2MSS.
4302                  */
4303                 reqbytes -= (2 * segsiz);
4304                 /* Also lets fill previous for our first measurement to be neutral */
4305                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4306         }
4307         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4308                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4309                                            rack->r_ctl.rc_app_limited_cnt,
4310                                            0, 0, 10, __LINE__, NULL);
4311                 goto skip_measurement;
4312         }
4313         /*
4314          * We now need to calculate the Timely like status so
4315          * we can update (possibly) the b/w multipliers.
4316          */
4317         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4318         if (rack->rc_gp_filled == 0) {
4319                 /* No previous reading */
4320                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4321         } else {
4322                 if (rack->measure_saw_probe_rtt == 0) {
4323                         /*
4324                          * We don't want a probertt to be counted
4325                          * since it will be negative incorrectly. We
4326                          * expect to be reducing the RTT when we
4327                          * pace at a slower rate.
4328                          */
4329                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4330                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4331                 }
4332         }
4333         timely_says = rack_make_timely_judgement(rack,
4334                 rack->r_ctl.rc_gp_srtt,
4335                 rack->r_ctl.rc_rtt_diff,
4336                 rack->r_ctl.rc_prev_gp_srtt
4337                 );
4338         bytes_ps *= HPTS_USEC_IN_SEC;
4339         bytes_ps /= utim;
4340         if (bytes_ps > rack->r_ctl.last_max_bw) {
4341                 /*
4342                  * Something is on path playing
4343                  * since this b/w is not possible based
4344                  * on our BDP (highest rwnd and lowest rtt
4345                  * we saw in the measurement window).
4346                  *
4347                  * Another option here would be to
4348                  * instead skip the measurement.
4349                  */
4350                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4351                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4352                                            11, __LINE__, NULL);
4353                 bytes_ps = rack->r_ctl.last_max_bw;
4354         }
4355         /* We store gp for b/w in bytes per second */
4356         if (rack->rc_gp_filled == 0) {
4357                 /* Initial measurment */
4358                 if (bytes_ps) {
4359                         rack->r_ctl.gp_bw = bytes_ps;
4360                         rack->rc_gp_filled = 1;
4361                         rack->r_ctl.num_measurements = 1;
4362                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4363                 } else {
4364                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4365                                                    rack->r_ctl.rc_app_limited_cnt,
4366                                                    0, 0, 10, __LINE__, NULL);
4367                 }
4368                 if (rack->rc_inp->inp_in_hpts &&
4369                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4370                         /*
4371                          * Ok we can't trust the pacer in this case
4372                          * where we transition from un-paced to paced.
4373                          * Or for that matter when the burst mitigation
4374                          * was making a wild guess and got it wrong.
4375                          * Stop the pacer and clear up all the aggregate
4376                          * delays etc.
4377                          */
4378                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
4379                         rack->r_ctl.rc_hpts_flags = 0;
4380                         rack->r_ctl.rc_last_output_to = 0;
4381                 }
4382                 did_add = 2;
4383         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4384                 /* Still a small number run an average */
4385                 rack->r_ctl.gp_bw += bytes_ps;
4386                 addpart = rack->r_ctl.num_measurements;
4387                 rack->r_ctl.num_measurements++;
4388                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4389                         /* We have collected enought to move forward */
4390                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4391                 }
4392                 did_add = 3;
4393         } else {
4394                 /*
4395                  * We want to take 1/wma of the goodput and add in to 7/8th
4396                  * of the old value weighted by the srtt. So if your measurement
4397                  * period is say 2 SRTT's long you would get 1/4 as the
4398                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4399                  *
4400                  * But we must be careful not to take too much i.e. if the
4401                  * srtt is say 20ms and the measurement is taken over
4402                  * 400ms our weight would be 400/20 i.e. 20. On the
4403                  * other hand if we get a measurement over 1ms with a
4404                  * 10ms rtt we only want to take a much smaller portion.
4405                  */
4406                 if (rack->r_ctl.num_measurements < 0xff) {
4407                         rack->r_ctl.num_measurements++;
4408                 }
4409                 srtt = (uint64_t)tp->t_srtt;
4410                 if (srtt == 0) {
4411                         /*
4412                          * Strange why did t_srtt go back to zero?
4413                          */
4414                         if (rack->r_ctl.rc_rack_min_rtt)
4415                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4416                         else
4417                                 srtt = HPTS_USEC_IN_MSEC;
4418                 }
4419                 /*
4420                  * XXXrrs: Note for reviewers, in playing with
4421                  * dynamic pacing I discovered this GP calculation
4422                  * as done originally leads to some undesired results.
4423                  * Basically you can get longer measurements contributing
4424                  * too much to the WMA. Thus I changed it if you are doing
4425                  * dynamic adjustments to only do the aportioned adjustment
4426                  * if we have a very small (time wise) measurement. Longer
4427                  * measurements just get there weight (defaulting to 1/8)
4428                  * add to the WMA. We may want to think about changing
4429                  * this to always do that for both sides i.e. dynamic
4430                  * and non-dynamic... but considering lots of folks
4431                  * were playing with this I did not want to change the
4432                  * calculation per.se. without your thoughts.. Lawerence?
4433                  * Peter??
4434                  */
4435                 if (rack->rc_gp_dyn_mul == 0) {
4436                         subpart = rack->r_ctl.gp_bw * utim;
4437                         subpart /= (srtt * 8);
4438                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4439                                 /*
4440                                  * The b/w update takes no more
4441                                  * away then 1/2 our running total
4442                                  * so factor it in.
4443                                  */
4444                                 addpart = bytes_ps * utim;
4445                                 addpart /= (srtt * 8);
4446                         } else {
4447                                 /*
4448                                  * Don't allow a single measurement
4449                                  * to account for more than 1/2 of the
4450                                  * WMA. This could happen on a retransmission
4451                                  * where utim becomes huge compared to
4452                                  * srtt (multiple retransmissions when using
4453                                  * the sending rate which factors in all the
4454                                  * transmissions from the first one).
4455                                  */
4456                                 subpart = rack->r_ctl.gp_bw / 2;
4457                                 addpart = bytes_ps / 2;
4458                         }
4459                         resid_bw = rack->r_ctl.gp_bw - subpart;
4460                         rack->r_ctl.gp_bw = resid_bw + addpart;
4461                         did_add = 1;
4462                 } else {
4463                         if ((utim / srtt) <= 1) {
4464                                 /*
4465                                  * The b/w update was over a small period
4466                                  * of time. The idea here is to prevent a small
4467                                  * measurement time period from counting
4468                                  * too much. So we scale it based on the
4469                                  * time so it attributes less than 1/rack_wma_divisor
4470                                  * of its measurement.
4471                                  */
4472                                 subpart = rack->r_ctl.gp_bw * utim;
4473                                 subpart /= (srtt * rack_wma_divisor);
4474                                 addpart = bytes_ps * utim;
4475                                 addpart /= (srtt * rack_wma_divisor);
4476                         } else {
4477                                 /*
4478                                  * The scaled measurement was long
4479                                  * enough so lets just add in the
4480                                  * portion of the measurment i.e. 1/rack_wma_divisor
4481                                  */
4482                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4483                                 addpart = bytes_ps / rack_wma_divisor;
4484                         }
4485                         if ((rack->measure_saw_probe_rtt == 0) ||
4486                             (bytes_ps > rack->r_ctl.gp_bw)) {
4487                                 /*
4488                                  * For probe-rtt we only add it in
4489                                  * if its larger, all others we just
4490                                  * add in.
4491                                  */
4492                                 did_add = 1;
4493                                 resid_bw = rack->r_ctl.gp_bw - subpart;
4494                                 rack->r_ctl.gp_bw = resid_bw + addpart;
4495                         }
4496                 }
4497         }
4498         if ((rack->gp_ready == 0) &&
4499             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4500                 /* We have enough measurements now */
4501                 rack->gp_ready = 1;
4502                 rack_set_cc_pacing(rack);
4503                 if (rack->defer_options)
4504                         rack_apply_deferred_options(rack);
4505         }
4506         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4507                                    rack_get_bw(rack), 22, did_add, NULL);
4508         /* We do not update any multipliers if we are in or have seen a probe-rtt */
4509         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4510                 rack_update_multiplier(rack, timely_says, bytes_ps,
4511                                        rack->r_ctl.rc_gp_srtt,
4512                                        rack->r_ctl.rc_rtt_diff);
4513         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4514                                    rack_get_bw(rack), 3, line, NULL);
4515         /* reset the gp srtt and setup the new prev */
4516         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4517         /* Record the lost count for the next measurement */
4518         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4519         /*
4520          * We restart our diffs based on the gpsrtt in the
4521          * measurement window.
4522          */
4523         rack->rc_gp_rtt_set = 0;
4524         rack->rc_gp_saw_rec = 0;
4525         rack->rc_gp_saw_ca = 0;
4526         rack->rc_gp_saw_ss = 0;
4527         rack->rc_dragged_bottom = 0;
4528 skip_measurement:
4529
4530 #ifdef STATS
4531         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4532                                  gput);
4533         /*
4534          * XXXLAS: This is a temporary hack, and should be
4535          * chained off VOI_TCP_GPUT when stats(9) grows an
4536          * API to deal with chained VOIs.
4537          */
4538         if (tp->t_stats_gput_prev > 0)
4539                 stats_voi_update_abs_s32(tp->t_stats,
4540                                          VOI_TCP_GPUT_ND,
4541                                          ((gput - tp->t_stats_gput_prev) * 100) /
4542                                          tp->t_stats_gput_prev);
4543 #endif
4544         tp->t_flags &= ~TF_GPUTINPROG;
4545         tp->t_stats_gput_prev = gput;
4546         /*
4547          * Now are we app limited now and there is space from where we
4548          * were to where we want to go?
4549          *
4550          * We don't do the other case i.e. non-applimited here since
4551          * the next send will trigger us picking up the missing data.
4552          */
4553         if (rack->r_ctl.rc_first_appl &&
4554             TCPS_HAVEESTABLISHED(tp->t_state) &&
4555             rack->r_ctl.rc_app_limited_cnt &&
4556             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4557             ((rack->r_ctl.rc_first_appl->r_start - th_ack) >
4558              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4559                 /*
4560                  * Yep there is enough outstanding to make a measurement here.
4561                  */
4562                 struct rack_sendmap *rsm, fe;
4563
4564                 tp->t_flags |= TF_GPUTINPROG;
4565                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4566                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4567                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4568                 rack->app_limited_needs_set = 0;
4569                 tp->gput_seq = th_ack;
4570                 if (rack->in_probe_rtt)
4571                         rack->measure_saw_probe_rtt = 1;
4572                 else if ((rack->measure_saw_probe_rtt) &&
4573                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4574                         rack->measure_saw_probe_rtt = 0;
4575                 if ((rack->r_ctl.rc_first_appl->r_start - th_ack) >= rack_get_measure_window(tp, rack)) {
4576                         /* There is a full window to gain info from */
4577                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4578                 } else {
4579                         /* We can only measure up to the applimited point */
4580                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_start - th_ack);
4581                 }
4582                 /*
4583                  * Now we need to find the timestamp of the send at tp->gput_seq
4584                  * for the send based measurement.
4585                  */
4586                 fe.r_start = tp->gput_seq;
4587                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4588                 if (rsm) {
4589                         /* Ok send-based limit is set */
4590                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4591                                 /*
4592                                  * Move back to include the earlier part
4593                                  * so our ack time lines up right (this may
4594                                  * make an overlapping measurement but thats
4595                                  * ok).
4596                                  */
4597                                 tp->gput_seq = rsm->r_start;
4598                         }
4599                         if (rsm->r_flags & RACK_ACKED)
4600                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4601                         else
4602                                 rack->app_limited_needs_set = 1;
4603                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4604                 } else {
4605                         /*
4606                          * If we don't find the rsm due to some
4607                          * send-limit set the current time, which
4608                          * basically disables the send-limit.
4609                          */
4610                         struct timeval tv;
4611
4612                         microuptime(&tv);
4613                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4614                 }
4615                 rack_log_pacing_delay_calc(rack,
4616                                            tp->gput_seq,
4617                                            tp->gput_ack,
4618                                            (uint64_t)rsm,
4619                                            tp->gput_ts,
4620                                            rack->r_ctl.rc_app_limited_cnt,
4621                                            9,
4622                                            __LINE__, NULL);
4623         }
4624 }
4625
4626 /*
4627  * CC wrapper hook functions
4628  */
4629 static void
4630 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4631     uint16_t type, int32_t recovery)
4632 {
4633         uint32_t prior_cwnd, acked;
4634         struct tcp_log_buffer *lgb = NULL;
4635         uint8_t labc_to_use;
4636
4637         INP_WLOCK_ASSERT(tp->t_inpcb);
4638         tp->ccv->nsegs = nsegs;
4639         acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4640         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4641                 uint32_t max;
4642
4643                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4644                 if (tp->ccv->bytes_this_ack > max) {
4645                         tp->ccv->bytes_this_ack = max;
4646                 }
4647         }
4648 #ifdef STATS
4649         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4650             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4651 #endif
4652         if ((tp->t_flags & TF_GPUTINPROG) &&
4653             rack_enough_for_measurement(tp, rack, th_ack)) {
4654                 /* Measure the Goodput */
4655                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__);
4656 #ifdef NETFLIX_PEAKRATE
4657                 if ((type == CC_ACK) &&
4658                     (tp->t_maxpeakrate)) {
4659                         /*
4660                          * We update t_peakrate_thr. This gives us roughly
4661                          * one update per round trip time. Note
4662                          * it will only be used if pace_always is off i.e
4663                          * we don't do this for paced flows.
4664                          */
4665                         rack_update_peakrate_thr(tp);
4666                 }
4667 #endif
4668         }
4669         /* Which way our we limited, if not cwnd limited no advance in CA */
4670         if (tp->snd_cwnd <= tp->snd_wnd)
4671                 tp->ccv->flags |= CCF_CWND_LIMITED;
4672         else
4673                 tp->ccv->flags &= ~CCF_CWND_LIMITED;
4674         if (tp->snd_cwnd > tp->snd_ssthresh) {
4675                 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4676                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4677                 /* For the setting of a window past use the actual scwnd we are using */
4678                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4679                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4680                         tp->ccv->flags |= CCF_ABC_SENTAWND;
4681                 }
4682         } else {
4683                 tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4684                 tp->t_bytes_acked = 0;
4685         }
4686         prior_cwnd = tp->snd_cwnd;
4687         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4688             (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4689                 labc_to_use = rack->rc_labc;
4690         else
4691                 labc_to_use = rack_max_abc_post_recovery;
4692         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4693                 union tcp_log_stackspecific log;
4694                 struct timeval tv;
4695
4696                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4697                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4698                 log.u_bbr.flex1 = th_ack;
4699                 log.u_bbr.flex2 = tp->ccv->flags;
4700                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4701                 log.u_bbr.flex4 = tp->ccv->nsegs;
4702                 log.u_bbr.flex5 = labc_to_use;
4703                 log.u_bbr.flex6 = prior_cwnd;
4704                 log.u_bbr.flex7 = V_tcp_do_newsack;
4705                 log.u_bbr.flex8 = 1;
4706                 lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4707                                      0, &log, false, NULL, NULL, 0, &tv);
4708         }
4709         if (CC_ALGO(tp)->ack_received != NULL) {
4710                 /* XXXLAS: Find a way to live without this */
4711                 tp->ccv->curack = th_ack;
4712                 tp->ccv->labc = labc_to_use;
4713                 tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4714                 CC_ALGO(tp)->ack_received(tp->ccv, type);
4715         }
4716         if (lgb) {
4717                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4718         }
4719         if (rack->r_must_retran) {
4720                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4721                         /*
4722                          * We now are beyond the rxt point so lets disable
4723                          * the flag.
4724                          */
4725                         rack->r_ctl.rc_out_at_rto = 0;
4726                         rack->r_must_retran = 0;
4727                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4728                         /*
4729                          * Only decrement the rc_out_at_rto if the cwnd advances
4730                          * at least a whole segment. Otherwise next time the peer
4731                          * acks, we won't be able to send this generaly happens
4732                          * when we are in Congestion Avoidance.
4733                          */
4734                         if (acked <= rack->r_ctl.rc_out_at_rto){
4735                                 rack->r_ctl.rc_out_at_rto -= acked;
4736                         } else {
4737                                 rack->r_ctl.rc_out_at_rto = 0;
4738                         }
4739                 }
4740         }
4741 #ifdef STATS
4742         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4743 #endif
4744         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4745                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4746         }
4747 #ifdef NETFLIX_PEAKRATE
4748         /* we enforce max peak rate if it is set and we are not pacing */
4749         if ((rack->rc_always_pace == 0) &&
4750             tp->t_peakrate_thr &&
4751             (tp->snd_cwnd > tp->t_peakrate_thr)) {
4752                 tp->snd_cwnd = tp->t_peakrate_thr;
4753         }
4754 #endif
4755 }
4756
4757 static void
4758 tcp_rack_partialack(struct tcpcb *tp)
4759 {
4760         struct tcp_rack *rack;
4761
4762         rack = (struct tcp_rack *)tp->t_fb_ptr;
4763         INP_WLOCK_ASSERT(tp->t_inpcb);
4764         /*
4765          * If we are doing PRR and have enough
4766          * room to send <or> we are pacing and prr
4767          * is disabled we will want to see if we
4768          * can send data (by setting r_wanted_output to
4769          * true).
4770          */
4771         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4772             rack->rack_no_prr)
4773                 rack->r_wanted_output = 1;
4774 }
4775
4776 static void
4777 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4778 {
4779         struct tcp_rack *rack;
4780         uint32_t orig_cwnd;
4781
4782         orig_cwnd = tp->snd_cwnd;
4783         INP_WLOCK_ASSERT(tp->t_inpcb);
4784         rack = (struct tcp_rack *)tp->t_fb_ptr;
4785         /* only alert CC if we alerted when we entered */
4786         if (CC_ALGO(tp)->post_recovery != NULL) {
4787                 tp->ccv->curack = th_ack;
4788                 CC_ALGO(tp)->post_recovery(tp->ccv);
4789                 if (tp->snd_cwnd < tp->snd_ssthresh) {
4790                         /*
4791                          * Rack has burst control and pacing
4792                          * so lets not set this any lower than
4793                          * snd_ssthresh per RFC-6582 (option 2).
4794                          */
4795                         tp->snd_cwnd = tp->snd_ssthresh;
4796                 }
4797         }
4798         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4799                 union tcp_log_stackspecific log;
4800                 struct timeval tv;
4801
4802                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4803                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4804                 log.u_bbr.flex1 = th_ack;
4805                 log.u_bbr.flex2 = tp->ccv->flags;
4806                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4807                 log.u_bbr.flex4 = tp->ccv->nsegs;
4808                 log.u_bbr.flex5 = V_tcp_abc_l_var;
4809                 log.u_bbr.flex6 = orig_cwnd;
4810                 log.u_bbr.flex7 = V_tcp_do_newsack;
4811                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4812                 log.u_bbr.flex8 = 2;
4813                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4814                                0, &log, false, NULL, NULL, 0, &tv);
4815         }
4816         if ((rack->rack_no_prr == 0) &&
4817             (rack->no_prr_addback == 0) &&
4818             (rack->r_ctl.rc_prr_sndcnt > 0)) {
4819                 /*
4820                  * Suck the next prr cnt back into cwnd, but
4821                  * only do that if we are not application limited.
4822                  */
4823                 if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4824                         /*
4825                          * We are allowed to add back to the cwnd the amount we did
4826                          * not get out if:
4827                          * a) no_prr_addback is off.
4828                          * b) we are not app limited
4829                          * c) we are doing prr
4830                          * <and>
4831                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4832                          */
4833                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4834                                             rack->r_ctl.rc_prr_sndcnt);
4835                 }
4836                 rack->r_ctl.rc_prr_sndcnt = 0;
4837                 rack_log_to_prr(rack, 1, 0);
4838         }
4839         rack_log_to_prr(rack, 14, orig_cwnd);
4840         tp->snd_recover = tp->snd_una;
4841         EXIT_RECOVERY(tp->t_flags);
4842 }
4843
4844 static void
4845 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack)
4846 {
4847         struct tcp_rack *rack;
4848         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4849
4850         INP_WLOCK_ASSERT(tp->t_inpcb);
4851 #ifdef STATS
4852         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4853 #endif
4854         if (IN_RECOVERY(tp->t_flags) == 0) {
4855                 in_rec_at_entry = 0;
4856                 ssthresh_enter = tp->snd_ssthresh;
4857                 cwnd_enter = tp->snd_cwnd;
4858         } else
4859                 in_rec_at_entry = 1;
4860         rack = (struct tcp_rack *)tp->t_fb_ptr;
4861         switch (type) {
4862         case CC_NDUPACK:
4863                 tp->t_flags &= ~TF_WASFRECOVERY;
4864                 tp->t_flags &= ~TF_WASCRECOVERY;
4865                 if (!IN_FASTRECOVERY(tp->t_flags)) {
4866                         rack->r_ctl.rc_prr_delivered = 0;
4867                         rack->r_ctl.rc_prr_out = 0;
4868                         if (rack->rack_no_prr == 0) {
4869                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4870                                 rack_log_to_prr(rack, 2, in_rec_at_entry);
4871                         }
4872                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4873                         tp->snd_recover = tp->snd_max;
4874                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4875                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4876                 }
4877                 break;
4878         case CC_ECN:
4879                 if (!IN_CONGRECOVERY(tp->t_flags) ||
4880                     /*
4881                      * Allow ECN reaction on ACK to CWR, if
4882                      * that data segment was also CE marked.
4883                      */
4884                     SEQ_GEQ(ack, tp->snd_recover)) {
4885                         EXIT_CONGRECOVERY(tp->t_flags);
4886                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4887                         tp->snd_recover = tp->snd_max + 1;
4888                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4889                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4890                 }
4891                 break;
4892         case CC_RTO:
4893                 tp->t_dupacks = 0;
4894                 tp->t_bytes_acked = 0;
4895                 EXIT_RECOVERY(tp->t_flags);
4896                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4897                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4898                 orig_cwnd = tp->snd_cwnd;
4899                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4900                 rack_log_to_prr(rack, 16, orig_cwnd);
4901                 if (tp->t_flags2 & TF2_ECN_PERMIT)
4902                         tp->t_flags2 |= TF2_ECN_SND_CWR;
4903                 break;
4904         case CC_RTO_ERR:
4905                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4906                 /* RTO was unnecessary, so reset everything. */
4907                 tp->snd_cwnd = tp->snd_cwnd_prev;
4908                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4909                 tp->snd_recover = tp->snd_recover_prev;
4910                 if (tp->t_flags & TF_WASFRECOVERY) {
4911                         ENTER_FASTRECOVERY(tp->t_flags);
4912                         tp->t_flags &= ~TF_WASFRECOVERY;
4913                 }
4914                 if (tp->t_flags & TF_WASCRECOVERY) {
4915                         ENTER_CONGRECOVERY(tp->t_flags);
4916                         tp->t_flags &= ~TF_WASCRECOVERY;
4917                 }
4918                 tp->snd_nxt = tp->snd_max;
4919                 tp->t_badrxtwin = 0;
4920                 break;
4921         }
4922         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4923             (type != CC_RTO)){
4924                 tp->ccv->curack = ack;
4925                 CC_ALGO(tp)->cong_signal(tp->ccv, type);
4926         }
4927         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4928                 rack_log_to_prr(rack, 15, cwnd_enter);
4929                 rack->r_ctl.dsack_byte_cnt = 0;
4930                 rack->r_ctl.retran_during_recovery = 0;
4931                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4932                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4933                 rack->r_ent_rec_ns = 1;
4934         }
4935 }
4936
4937 static inline void
4938 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4939 {
4940         uint32_t i_cwnd;
4941
4942         INP_WLOCK_ASSERT(tp->t_inpcb);
4943
4944 #ifdef NETFLIX_STATS
4945         KMOD_TCPSTAT_INC(tcps_idle_restarts);
4946         if (tp->t_state == TCPS_ESTABLISHED)
4947                 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4948 #endif
4949         if (CC_ALGO(tp)->after_idle != NULL)
4950                 CC_ALGO(tp)->after_idle(tp->ccv);
4951
4952         if (tp->snd_cwnd == 1)
4953                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
4954         else
4955                 i_cwnd = rc_init_window(rack);
4956
4957         /*
4958          * Being idle is no differnt than the initial window. If the cc
4959          * clamps it down below the initial window raise it to the initial
4960          * window.
4961          */
4962         if (tp->snd_cwnd < i_cwnd) {
4963                 tp->snd_cwnd = i_cwnd;
4964         }
4965 }
4966
4967 /*
4968  * Indicate whether this ack should be delayed.  We can delay the ack if
4969  * following conditions are met:
4970  *      - There is no delayed ack timer in progress.
4971  *      - Our last ack wasn't a 0-sized window. We never want to delay
4972  *        the ack that opens up a 0-sized window.
4973  *      - LRO wasn't used for this segment. We make sure by checking that the
4974  *        segment size is not larger than the MSS.
4975  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4976  *        connection.
4977  */
4978 #define DELAY_ACK(tp, tlen)                      \
4979         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4980         ((tp->t_flags & TF_DELACK) == 0) &&      \
4981         (tlen <= tp->t_maxseg) &&                \
4982         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4983
4984 static struct rack_sendmap *
4985 rack_find_lowest_rsm(struct tcp_rack *rack)
4986 {
4987         struct rack_sendmap *rsm;
4988
4989         /*
4990          * Walk the time-order transmitted list looking for an rsm that is
4991          * not acked. This will be the one that was sent the longest time
4992          * ago that is still outstanding.
4993          */
4994         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4995                 if (rsm->r_flags & RACK_ACKED) {
4996                         continue;
4997                 }
4998                 goto finish;
4999         }
5000 finish:
5001         return (rsm);
5002 }
5003
5004 static struct rack_sendmap *
5005 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5006 {
5007         struct rack_sendmap *prsm;
5008
5009         /*
5010          * Walk the sequence order list backward until we hit and arrive at
5011          * the highest seq not acked. In theory when this is called it
5012          * should be the last segment (which it was not).
5013          */
5014         counter_u64_add(rack_find_high, 1);
5015         prsm = rsm;
5016         RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
5017                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5018                         continue;
5019                 }
5020                 return (prsm);
5021         }
5022         return (NULL);
5023 }
5024
5025 static uint32_t
5026 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
5027 {
5028         int32_t lro;
5029         uint32_t thresh;
5030
5031         /*
5032          * lro is the flag we use to determine if we have seen reordering.
5033          * If it gets set we have seen reordering. The reorder logic either
5034          * works in one of two ways:
5035          *
5036          * If reorder-fade is configured, then we track the last time we saw
5037          * re-ordering occur. If we reach the point where enough time as
5038          * passed we no longer consider reordering has occuring.
5039          *
5040          * Or if reorder-face is 0, then once we see reordering we consider
5041          * the connection to alway be subject to reordering and just set lro
5042          * to 1.
5043          *
5044          * In the end if lro is non-zero we add the extra time for
5045          * reordering in.
5046          */
5047         if (srtt == 0)
5048                 srtt = 1;
5049         if (rack->r_ctl.rc_reorder_ts) {
5050                 if (rack->r_ctl.rc_reorder_fade) {
5051                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5052                                 lro = cts - rack->r_ctl.rc_reorder_ts;
5053                                 if (lro == 0) {
5054                                         /*
5055                                          * No time as passed since the last
5056                                          * reorder, mark it as reordering.
5057                                          */
5058                                         lro = 1;
5059                                 }
5060                         } else {
5061                                 /* Negative time? */
5062                                 lro = 0;
5063                         }
5064                         if (lro > rack->r_ctl.rc_reorder_fade) {
5065                                 /* Turn off reordering seen too */
5066                                 rack->r_ctl.rc_reorder_ts = 0;
5067                                 lro = 0;
5068                         }
5069                 } else {
5070                         /* Reodering does not fade */
5071                         lro = 1;
5072                 }
5073         } else {
5074                 lro = 0;
5075         }
5076         thresh = srtt + rack->r_ctl.rc_pkt_delay;
5077         if (lro) {
5078                 /* It must be set, if not you get 1/4 rtt */
5079                 if (rack->r_ctl.rc_reorder_shift)
5080                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5081                 else
5082                         thresh += (srtt >> 2);
5083         } else {
5084                 thresh += 1;
5085         }
5086         /* We don't let the rack timeout be above a RTO */
5087         if (thresh > rack->rc_tp->t_rxtcur) {
5088                 thresh = rack->rc_tp->t_rxtcur;
5089         }
5090         /* And we don't want it above the RTO max either */
5091         if (thresh > rack_rto_max) {
5092                 thresh = rack_rto_max;
5093         }
5094         return (thresh);
5095 }
5096
5097 static uint32_t
5098 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5099                      struct rack_sendmap *rsm, uint32_t srtt)
5100 {
5101         struct rack_sendmap *prsm;
5102         uint32_t thresh, len;
5103         int segsiz;
5104
5105         if (srtt == 0)
5106                 srtt = 1;
5107         if (rack->r_ctl.rc_tlp_threshold)
5108                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5109         else
5110                 thresh = (srtt * 2);
5111
5112         /* Get the previous sent packet, if any */
5113         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5114         counter_u64_add(rack_enter_tlp_calc, 1);
5115         len = rsm->r_end - rsm->r_start;
5116         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5117                 /* Exactly like the ID */
5118                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5119                         uint32_t alt_thresh;
5120                         /*
5121                          * Compensate for delayed-ack with the d-ack time.
5122                          */
5123                         counter_u64_add(rack_used_tlpmethod, 1);
5124                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5125                         if (alt_thresh > thresh)
5126                                 thresh = alt_thresh;
5127                 }
5128         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5129                 /* 2.1 behavior */
5130                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5131                 if (prsm && (len <= segsiz)) {
5132                         /*
5133                          * Two packets outstanding, thresh should be (2*srtt) +
5134                          * possible inter-packet delay (if any).
5135                          */
5136                         uint32_t inter_gap = 0;
5137                         int idx, nidx;
5138
5139                         counter_u64_add(rack_used_tlpmethod, 1);
5140                         idx = rsm->r_rtr_cnt - 1;
5141                         nidx = prsm->r_rtr_cnt - 1;
5142                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5143                                 /* Yes it was sent later (or at the same time) */
5144                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5145                         }
5146                         thresh += inter_gap;
5147                 } else if (len <= segsiz) {
5148                         /*
5149                          * Possibly compensate for delayed-ack.
5150                          */
5151                         uint32_t alt_thresh;
5152
5153                         counter_u64_add(rack_used_tlpmethod2, 1);
5154                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5155                         if (alt_thresh > thresh)
5156                                 thresh = alt_thresh;
5157                 }
5158         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5159                 /* 2.2 behavior */
5160                 if (len <= segsiz) {
5161                         uint32_t alt_thresh;
5162                         /*
5163                          * Compensate for delayed-ack with the d-ack time.
5164                          */
5165                         counter_u64_add(rack_used_tlpmethod, 1);
5166                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5167                         if (alt_thresh > thresh)
5168                                 thresh = alt_thresh;
5169                 }
5170         }
5171         /* Not above an RTO */
5172         if (thresh > tp->t_rxtcur) {
5173                 thresh = tp->t_rxtcur;
5174         }
5175         /* Not above a RTO max */
5176         if (thresh > rack_rto_max) {
5177                 thresh = rack_rto_max;
5178         }
5179         /* Apply user supplied min TLP */
5180         if (thresh < rack_tlp_min) {
5181                 thresh = rack_tlp_min;
5182         }
5183         return (thresh);
5184 }
5185
5186 static uint32_t
5187 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5188 {
5189         /*
5190          * We want the rack_rtt which is the
5191          * last rtt we measured. However if that
5192          * does not exist we fallback to the srtt (which
5193          * we probably will never do) and then as a last
5194          * resort we use RACK_INITIAL_RTO if no srtt is
5195          * yet set.
5196          */
5197         if (rack->rc_rack_rtt)
5198                 return (rack->rc_rack_rtt);
5199         else if (tp->t_srtt == 0)
5200                 return (RACK_INITIAL_RTO);
5201         return (tp->t_srtt);
5202 }
5203
5204 static struct rack_sendmap *
5205 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5206 {
5207         /*
5208          * Check to see that we don't need to fall into recovery. We will
5209          * need to do so if our oldest transmit is past the time we should
5210          * have had an ack.
5211          */
5212         struct tcp_rack *rack;
5213         struct rack_sendmap *rsm;
5214         int32_t idx;
5215         uint32_t srtt, thresh;
5216
5217         rack = (struct tcp_rack *)tp->t_fb_ptr;
5218         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5219                 return (NULL);
5220         }
5221         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5222         if (rsm == NULL)
5223                 return (NULL);
5224
5225         if (rsm->r_flags & RACK_ACKED) {
5226                 rsm = rack_find_lowest_rsm(rack);
5227                 if (rsm == NULL)
5228                         return (NULL);
5229         }
5230         idx = rsm->r_rtr_cnt - 1;
5231         srtt = rack_grab_rtt(tp, rack);
5232         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5233         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5234                 return (NULL);
5235         }
5236         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5237                 return (NULL);
5238         }
5239         /* Ok if we reach here we are over-due and this guy can be sent */
5240         if (IN_RECOVERY(tp->t_flags) == 0) {
5241                 /*
5242                  * For the one that enters us into recovery record undo
5243                  * info.
5244                  */
5245                 rack->r_ctl.rc_rsm_start = rsm->r_start;
5246                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
5247                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
5248         }
5249         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
5250         return (rsm);
5251 }
5252
5253 static uint32_t
5254 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5255 {
5256         int32_t t;
5257         int32_t tt;
5258         uint32_t ret_val;
5259
5260         t = (tp->t_srtt + (tp->t_rttvar << 2));
5261         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5262             rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5263         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
5264                 tp->t_rxtshift++;
5265         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5266         ret_val = (uint32_t)tt;
5267         return (ret_val);
5268 }
5269
5270 static uint32_t
5271 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5272 {
5273         /*
5274          * Start the FR timer, we do this based on getting the first one in
5275          * the rc_tmap. Note that if its NULL we must stop the timer. in all
5276          * events we need to stop the running timer (if its running) before
5277          * starting the new one.
5278          */
5279         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5280         uint32_t srtt_cur;
5281         int32_t idx;
5282         int32_t is_tlp_timer = 0;
5283         struct rack_sendmap *rsm;
5284
5285         if (rack->t_timers_stopped) {
5286                 /* All timers have been stopped none are to run */
5287                 return (0);
5288         }
5289         if (rack->rc_in_persist) {
5290                 /* We can't start any timer in persists */
5291                 return (rack_get_persists_timer_val(tp, rack));
5292         }
5293         rack->rc_on_min_to = 0;
5294         if ((tp->t_state < TCPS_ESTABLISHED) ||
5295             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5296                 goto activate_rxt;
5297         }
5298         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5299         if ((rsm == NULL) || sup_rack) {
5300                 /* Nothing on the send map or no rack */
5301 activate_rxt:
5302                 time_since_sent = 0;
5303                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5304                 if (rsm) {
5305                         /*
5306                          * Should we discount the RTX timer any?
5307                          *
5308                          * We want to discount it the smallest amount.
5309                          * If a timer (Rack/TLP or RXT) has gone off more
5310                          * recently thats the discount we want to use (now - timer time).
5311                          * If the retransmit of the oldest packet was more recent then
5312                          * we want to use that (now - oldest-packet-last_transmit_time).
5313                          *
5314                          */
5315                         idx = rsm->r_rtr_cnt - 1;
5316                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5317                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5318                         else
5319                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5320                         if (TSTMP_GT(cts, tstmp_touse))
5321                             time_since_sent = cts - tstmp_touse;
5322                 }
5323                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5324                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5325                         to = tp->t_rxtcur;
5326                         if (to > time_since_sent)
5327                                 to -= time_since_sent;
5328                         else
5329                                 to = rack->r_ctl.rc_min_to;
5330                         if (to == 0)
5331                                 to = 1;
5332                         /* Special case for KEEPINIT */
5333                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5334                             (TP_KEEPINIT(tp) != 0) &&
5335                             rsm) {
5336                                 /*
5337                                  * We have to put a ceiling on the rxt timer
5338                                  * of the keep-init timeout.
5339                                  */
5340                                 uint32_t max_time, red;
5341
5342                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5343                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5344                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5345                                         if (red < max_time)
5346                                                 max_time -= red;
5347                                         else
5348                                                 max_time = 1;
5349                                 }
5350                                 /* Reduce timeout to the keep value if needed */
5351                                 if (max_time < to)
5352                                         to = max_time;
5353                         }
5354                         return (to);
5355                 }
5356                 return (0);
5357         }
5358         if (rsm->r_flags & RACK_ACKED) {
5359                 rsm = rack_find_lowest_rsm(rack);
5360                 if (rsm == NULL) {
5361                         /* No lowest? */
5362                         goto activate_rxt;
5363                 }
5364         }
5365         if (rack->sack_attack_disable) {
5366                 /*
5367                  * We don't want to do
5368                  * any TLP's if you are an attacker.
5369                  * Though if you are doing what
5370                  * is expected you may still have
5371                  * SACK-PASSED marks.
5372                  */
5373                 goto activate_rxt;
5374         }
5375         /* Convert from ms to usecs */
5376         if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5377                 if ((tp->t_flags & TF_SENTFIN) &&
5378                     ((tp->snd_max - tp->snd_una) == 1) &&
5379                     (rsm->r_flags & RACK_HAS_FIN)) {
5380                         /*
5381                          * We don't start a rack timer if all we have is a
5382                          * FIN outstanding.
5383                          */
5384                         goto activate_rxt;
5385                 }
5386                 if ((rack->use_rack_rr == 0) &&
5387                     (IN_FASTRECOVERY(tp->t_flags)) &&
5388                     (rack->rack_no_prr == 0) &&
5389                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5390                         /*
5391                          * We are not cheating, in recovery  and
5392                          * not enough ack's to yet get our next
5393                          * retransmission out.
5394                          *
5395                          * Note that classified attackers do not
5396                          * get to use the rack-cheat.
5397                          */
5398                         goto activate_tlp;
5399                 }
5400                 srtt = rack_grab_rtt(tp, rack);
5401                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5402                 idx = rsm->r_rtr_cnt - 1;
5403                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5404                 if (SEQ_GEQ(exp, cts)) {
5405                         to = exp - cts;
5406                         if (to < rack->r_ctl.rc_min_to) {
5407                                 to = rack->r_ctl.rc_min_to;
5408                                 if (rack->r_rr_config == 3)
5409                                         rack->rc_on_min_to = 1;
5410                         }
5411                 } else {
5412                         to = rack->r_ctl.rc_min_to;
5413                         if (rack->r_rr_config == 3)
5414                                 rack->rc_on_min_to = 1;
5415                 }
5416         } else {
5417                 /* Ok we need to do a TLP not RACK */
5418 activate_tlp:
5419                 if ((rack->rc_tlp_in_progress != 0) &&
5420                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5421                         /*
5422                          * The previous send was a TLP and we have sent
5423                          * N TLP's without sending new data.
5424                          */
5425                         goto activate_rxt;
5426                 }
5427                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5428                 if (rsm == NULL) {
5429                         /* We found no rsm to TLP with. */
5430                         goto activate_rxt;
5431                 }
5432                 if (rsm->r_flags & RACK_HAS_FIN) {
5433                         /* If its a FIN we dont do TLP */
5434                         rsm = NULL;
5435                         goto activate_rxt;
5436                 }
5437                 idx = rsm->r_rtr_cnt - 1;
5438                 time_since_sent = 0;
5439                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5440                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5441                 else
5442                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5443                 if (TSTMP_GT(cts, tstmp_touse))
5444                     time_since_sent = cts - tstmp_touse;
5445                 is_tlp_timer = 1;
5446                 if (tp->t_srtt) {
5447                         if ((rack->rc_srtt_measure_made == 0) &&
5448                             (tp->t_srtt == 1)) {
5449                                 /*
5450                                  * If another stack as run and set srtt to 1,
5451                                  * then the srtt was 0, so lets use the initial.
5452                                  */
5453                                 srtt = RACK_INITIAL_RTO;
5454                         } else {
5455                                 srtt_cur = tp->t_srtt;
5456                                 srtt = srtt_cur;
5457                         }
5458                 } else
5459                         srtt = RACK_INITIAL_RTO;
5460                 /*
5461                  * If the SRTT is not keeping up and the
5462                  * rack RTT has spiked we want to use
5463                  * the last RTT not the smoothed one.
5464                  */
5465                 if (rack_tlp_use_greater &&
5466                     tp->t_srtt &&
5467                     (srtt < rack_grab_rtt(tp, rack))) {
5468                         srtt = rack_grab_rtt(tp, rack);
5469                 }
5470                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5471                 if (thresh > time_since_sent) {
5472                         to = thresh - time_since_sent;
5473                 } else {
5474                         to = rack->r_ctl.rc_min_to;
5475                         rack_log_alt_to_to_cancel(rack,
5476                                                   thresh,               /* flex1 */
5477                                                   time_since_sent,      /* flex2 */
5478                                                   tstmp_touse,          /* flex3 */
5479                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5480                                                   (uint32_t)rsm->r_tim_lastsent[idx],
5481                                                   srtt,
5482                                                   idx, 99);
5483                 }
5484                 if (to < rack_tlp_min) {
5485                         to = rack_tlp_min;
5486                 }
5487                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5488                         /*
5489                          * If the TLP time works out to larger than the max
5490                          * RTO lets not do TLP.. just RTO.
5491                          */
5492                         goto activate_rxt;
5493                 }
5494         }
5495         if (is_tlp_timer == 0) {
5496                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5497         } else {
5498                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5499         }
5500         if (to == 0)
5501                 to = 1;
5502         return (to);
5503 }
5504
5505 static void
5506 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5507 {
5508         if (rack->rc_in_persist == 0) {
5509                 if (tp->t_flags & TF_GPUTINPROG) {
5510                         /*
5511                          * Stop the goodput now, the calling of the
5512                          * measurement function clears the flag.
5513                          */
5514                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__);
5515                 }
5516 #ifdef NETFLIX_SHARED_CWND
5517                 if (rack->r_ctl.rc_scw) {
5518                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5519                         rack->rack_scwnd_is_idle = 1;
5520                 }
5521 #endif
5522                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5523                 if (rack->r_ctl.rc_went_idle_time == 0)
5524                         rack->r_ctl.rc_went_idle_time = 1;
5525                 rack_timer_cancel(tp, rack, cts, __LINE__);
5526                 tp->t_rxtshift = 0;
5527                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5528                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5529                 rack->rc_in_persist = 1;
5530         }
5531 }
5532
5533 static void
5534 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5535 {
5536         if (rack->rc_inp->inp_in_hpts) {
5537                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5538                 rack->r_ctl.rc_hpts_flags = 0;
5539         }
5540 #ifdef NETFLIX_SHARED_CWND
5541         if (rack->r_ctl.rc_scw) {
5542                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5543                 rack->rack_scwnd_is_idle = 0;
5544         }
5545 #endif
5546         if (rack->rc_gp_dyn_mul &&
5547             (rack->use_fixed_rate == 0) &&
5548             (rack->rc_always_pace)) {
5549                 /*
5550                  * Do we count this as if a probe-rtt just
5551                  * finished?
5552                  */
5553                 uint32_t time_idle, idle_min;
5554
5555                 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5556                 idle_min = rack_min_probertt_hold;
5557                 if (rack_probertt_gpsrtt_cnt_div) {
5558                         uint64_t extra;
5559                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5560                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5561                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5562                         idle_min += (uint32_t)extra;
5563                 }
5564                 if (time_idle >= idle_min) {
5565                         /* Yes, we count it as a probe-rtt. */
5566                         uint32_t us_cts;
5567
5568                         us_cts = tcp_get_usecs(NULL);
5569                         if (rack->in_probe_rtt == 0) {
5570                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5571                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5572                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5573                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5574                         } else {
5575                                 rack_exit_probertt(rack, us_cts);
5576                         }
5577                 }
5578         }
5579         rack->rc_in_persist = 0;
5580         rack->r_ctl.rc_went_idle_time = 0;
5581         tp->t_rxtshift = 0;
5582         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5583            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5584         rack->r_ctl.rc_agg_delayed = 0;
5585         rack->r_early = 0;
5586         rack->r_late = 0;
5587         rack->r_ctl.rc_agg_early = 0;
5588 }
5589
5590 static void
5591 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5592                    struct hpts_diag *diag, struct timeval *tv)
5593 {
5594         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5595                 union tcp_log_stackspecific log;
5596
5597                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5598                 log.u_bbr.flex1 = diag->p_nxt_slot;
5599                 log.u_bbr.flex2 = diag->p_cur_slot;
5600                 log.u_bbr.flex3 = diag->slot_req;
5601                 log.u_bbr.flex4 = diag->inp_hptsslot;
5602                 log.u_bbr.flex5 = diag->slot_remaining;
5603                 log.u_bbr.flex6 = diag->need_new_to;
5604                 log.u_bbr.flex7 = diag->p_hpts_active;
5605                 log.u_bbr.flex8 = diag->p_on_min_sleep;
5606                 /* Hijack other fields as needed */
5607                 log.u_bbr.epoch = diag->have_slept;
5608                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5609                 log.u_bbr.pkts_out = diag->co_ret;
5610                 log.u_bbr.applimited = diag->hpts_sleep_time;
5611                 log.u_bbr.delivered = diag->p_prev_slot;
5612                 log.u_bbr.inflight = diag->p_runningtick;
5613                 log.u_bbr.bw_inuse = diag->wheel_tick;
5614                 log.u_bbr.rttProp = diag->wheel_cts;
5615                 log.u_bbr.timeStamp = cts;
5616                 log.u_bbr.delRate = diag->maxticks;
5617                 log.u_bbr.cur_del_rate = diag->p_curtick;
5618                 log.u_bbr.cur_del_rate <<= 32;
5619                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5620                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5621                     &rack->rc_inp->inp_socket->so_rcv,
5622                     &rack->rc_inp->inp_socket->so_snd,
5623                     BBR_LOG_HPTSDIAG, 0,
5624                     0, &log, false, tv);
5625         }
5626
5627 }
5628
5629 static void
5630 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5631 {
5632         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5633                 union tcp_log_stackspecific log;
5634                 struct timeval tv;
5635
5636                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5637                 log.u_bbr.flex1 = sb->sb_flags;
5638                 log.u_bbr.flex2 = len;
5639                 log.u_bbr.flex3 = sb->sb_state;
5640                 log.u_bbr.flex8 = type;
5641                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5642                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5643                     &rack->rc_inp->inp_socket->so_rcv,
5644                     &rack->rc_inp->inp_socket->so_snd,
5645                     TCP_LOG_SB_WAKE, 0,
5646                     len, &log, false, &tv);
5647         }
5648 }
5649
5650 static void
5651 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5652       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5653 {
5654         struct hpts_diag diag;
5655         struct inpcb *inp;
5656         struct timeval tv;
5657         uint32_t delayed_ack = 0;
5658         uint32_t hpts_timeout;
5659         uint32_t entry_slot = slot;
5660         uint8_t stopped;
5661         uint32_t left = 0;
5662         uint32_t us_cts;
5663
5664         inp = tp->t_inpcb;
5665         if ((tp->t_state == TCPS_CLOSED) ||
5666             (tp->t_state == TCPS_LISTEN)) {
5667                 return;
5668         }
5669         if (inp->inp_in_hpts) {
5670                 /* Already on the pacer */
5671                 return;
5672         }
5673         stopped = rack->rc_tmr_stopped;
5674         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5675                 left = rack->r_ctl.rc_timer_exp - cts;
5676         }
5677         rack->r_ctl.rc_timer_exp = 0;
5678         rack->r_ctl.rc_hpts_flags = 0;
5679         us_cts = tcp_get_usecs(&tv);
5680         /* Now early/late accounting */
5681         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL);
5682         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5683                 /*
5684                  * We have a early carry over set,
5685                  * we can always add more time so we
5686                  * can always make this compensation.
5687                  *
5688                  * Note if ack's are allowed to wake us do not
5689                  * penalize the next timer for being awoke
5690                  * by an ack aka the rc_agg_early (non-paced mode).
5691                  */
5692                 slot += rack->r_ctl.rc_agg_early;
5693                 rack->r_early = 0;
5694                 rack->r_ctl.rc_agg_early = 0;
5695         }
5696         if (rack->r_late) {
5697                 /*
5698                  * This is harder, we can
5699                  * compensate some but it
5700                  * really depends on what
5701                  * the current pacing time is.
5702                  */
5703                 if (rack->r_ctl.rc_agg_delayed >= slot) {
5704                         /*
5705                          * We can't compensate for it all.
5706                          * And we have to have some time
5707                          * on the clock. We always have a min
5708                          * 10 slots (10 x 10 i.e. 100 usecs).
5709                          */
5710                         if (slot <= HPTS_TICKS_PER_USEC) {
5711                                 /* We gain delay */
5712                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_USEC - slot);
5713                                 slot = HPTS_TICKS_PER_USEC;
5714                         } else {
5715                                 /* We take off some */
5716                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_USEC);
5717                                 slot = HPTS_TICKS_PER_USEC;
5718                         }
5719                 } else {
5720                         slot -= rack->r_ctl.rc_agg_delayed;
5721                         rack->r_ctl.rc_agg_delayed = 0;
5722                         /* Make sure we have 100 useconds at minimum */
5723                         if (slot < HPTS_TICKS_PER_USEC) {
5724                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_USEC - slot;
5725                                 slot = HPTS_TICKS_PER_USEC;
5726                         }
5727                         if (rack->r_ctl.rc_agg_delayed == 0)
5728                                 rack->r_late = 0;
5729                 }
5730         }
5731         if (slot) {
5732                 /* We are pacing too */
5733                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5734         }
5735         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5736 #ifdef NETFLIX_EXP_DETECTION
5737         if (rack->sack_attack_disable &&
5738             (slot < tcp_sad_pacing_interval)) {
5739                 /*
5740                  * We have a potential attacker on
5741                  * the line. We have possibly some
5742                  * (or now) pacing time set. We want to
5743                  * slow down the processing of sacks by some
5744                  * amount (if it is an attacker). Set the default
5745                  * slot for attackers in place (unless the orginal
5746                  * interval is longer). Its stored in
5747                  * micro-seconds, so lets convert to msecs.
5748                  */
5749                 slot = tcp_sad_pacing_interval;
5750         }
5751 #endif
5752         if (tp->t_flags & TF_DELACK) {
5753                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5754                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5755         }
5756         if (delayed_ack && ((hpts_timeout == 0) ||
5757                             (delayed_ack < hpts_timeout)))
5758                 hpts_timeout = delayed_ack;
5759         else
5760                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5761         /*
5762          * If no timers are going to run and we will fall off the hptsi
5763          * wheel, we resort to a keep-alive timer if its configured.
5764          */
5765         if ((hpts_timeout == 0) &&
5766             (slot == 0)) {
5767                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5768                     (tp->t_state <= TCPS_CLOSING)) {
5769                         /*
5770                          * Ok we have no timer (persists, rack, tlp, rxt  or
5771                          * del-ack), we don't have segments being paced. So
5772                          * all that is left is the keepalive timer.
5773                          */
5774                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5775                                 /* Get the established keep-alive time */
5776                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5777                         } else {
5778                                 /*
5779                                  * Get the initial setup keep-alive time,
5780                                  * note that this is probably not going to
5781                                  * happen, since rack will be running a rxt timer
5782                                  * if a SYN of some sort is outstanding. It is
5783                                  * actually handled in rack_timeout_rxt().
5784                                  */
5785                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5786                         }
5787                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5788                         if (rack->in_probe_rtt) {
5789                                 /*
5790                                  * We want to instead not wake up a long time from
5791                                  * now but to wake up about the time we would
5792                                  * exit probe-rtt and initiate a keep-alive ack.
5793                                  * This will get us out of probe-rtt and update
5794                                  * our min-rtt.
5795                                  */
5796                                 hpts_timeout = rack_min_probertt_hold;
5797                         }
5798                 }
5799         }
5800         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5801             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5802                 /*
5803                  * RACK, TLP, persists and RXT timers all are restartable
5804                  * based on actions input .. i.e we received a packet (ack
5805                  * or sack) and that changes things (rw, or snd_una etc).
5806                  * Thus we can restart them with a new value. For
5807                  * keep-alive, delayed_ack we keep track of what was left
5808                  * and restart the timer with a smaller value.
5809                  */
5810                 if (left < hpts_timeout)
5811                         hpts_timeout = left;
5812         }
5813         if (hpts_timeout) {
5814                 /*
5815                  * Hack alert for now we can't time-out over 2,147,483
5816                  * seconds (a bit more than 596 hours), which is probably ok
5817                  * :).
5818                  */
5819                 if (hpts_timeout > 0x7ffffffe)
5820                         hpts_timeout = 0x7ffffffe;
5821                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5822         }
5823         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL);
5824         if ((rack->gp_ready == 0) &&
5825             (rack->use_fixed_rate == 0) &&
5826             (hpts_timeout < slot) &&
5827             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5828                 /*
5829                  * We have no good estimate yet for the
5830                  * old clunky burst mitigation or the
5831                  * real pacing. And the tlp or rxt is smaller
5832                  * than the pacing calculation. Lets not
5833                  * pace that long since we know the calculation
5834                  * so far is not accurate.
5835                  */
5836                 slot = hpts_timeout;
5837         }
5838         rack->r_ctl.last_pacing_time = slot;
5839         /**
5840          * Turn off all the flags for queuing by default. The
5841          * flags have important meanings to what happens when
5842          * LRO interacts with the transport. Most likely (by default now)
5843          * mbuf_queueing and ack compression are on. So the transport
5844          * has a couple of flags that control what happens (if those
5845          * are not on then these flags won't have any effect since it
5846          * won't go through the queuing LRO path).
5847          *
5848          * INP_MBUF_QUEUE_READY - This flags says that I am busy
5849          *                        pacing output, so don't disturb. But
5850          *                        it also means LRO can wake me if there
5851          *                        is a SACK arrival.
5852          *
5853          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5854          *                       with the above flag (QUEUE_READY) and
5855          *                       when present it says don't even wake me
5856          *                       if a SACK arrives.
5857          *
5858          * The idea behind these flags is that if we are pacing we
5859          * set the MBUF_QUEUE_READY and only get woken up if
5860          * a SACK arrives (which could change things) or if
5861          * our pacing timer expires. If, however, we have a rack
5862          * timer running, then we don't even want a sack to wake
5863          * us since the rack timer has to expire before we can send.
5864          *
5865          * Other cases should usually have none of the flags set
5866          * so LRO can call into us.
5867          */
5868         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5869         if (slot) {
5870                 rack->r_ctl.rc_last_output_to = us_cts + slot;
5871                 /*
5872                  * A pacing timer (slot) is being set, in
5873                  * such a case we cannot send (we are blocked by
5874                  * the timer). So lets tell LRO that it should not
5875                  * wake us unless there is a SACK. Note this only
5876                  * will be effective if mbuf queueing is on or
5877                  * compressed acks are being processed.
5878                  */
5879                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5880                 /*
5881                  * But wait if we have a Rack timer running
5882                  * even a SACK should not disturb us (with
5883                  * the exception of r_rr_config 3).
5884                  */
5885                 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5886                     (rack->r_rr_config != 3))
5887                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5888                 if (rack->rc_ack_can_sendout_data) {
5889                         /*
5890                          * Ahh but wait, this is that special case
5891                          * where the pacing timer can be disturbed
5892                          * backout the changes (used for non-paced
5893                          * burst limiting).
5894                          */
5895                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5896                 }
5897                 if ((rack->use_rack_rr) &&
5898                     (rack->r_rr_config < 2) &&
5899                     ((hpts_timeout) && (hpts_timeout < slot))) {
5900                         /*
5901                          * Arrange for the hpts to kick back in after the
5902                          * t-o if the t-o does not cause a send.
5903                          */
5904                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5905                                                    __LINE__, &diag);
5906                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5907                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5908                 } else {
5909                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
5910                                                    __LINE__, &diag);
5911                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5912                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5913                 }
5914         } else if (hpts_timeout) {
5915                 /*
5916                  * With respect to inp_flags2 here, lets let any new acks wake
5917                  * us up here. Since we are not pacing (no pacing timer), output
5918                  * can happen so we should let it. If its a Rack timer, then any inbound
5919                  * packet probably won't change the sending (we will be blocked)
5920                  * but it may change the prr stats so letting it in (the set defaults
5921                  * at the start of this block) are good enough.
5922                  */
5923                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5924                                            __LINE__, &diag);
5925                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5926                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5927         } else {
5928                 /* No timer starting */
5929 #ifdef INVARIANTS
5930                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5931                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5932                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5933                 }
5934 #endif
5935         }
5936         rack->rc_tmr_stopped = 0;
5937         if (slot)
5938                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5939 }
5940
5941 /*
5942  * RACK Timer, here we simply do logging and house keeping.
5943  * the normal rack_output() function will call the
5944  * appropriate thing to check if we need to do a RACK retransmit.
5945  * We return 1, saying don't proceed with rack_output only
5946  * when all timers have been stopped (destroyed PCB?).
5947  */
5948 static int
5949 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5950 {
5951         /*
5952          * This timer simply provides an internal trigger to send out data.
5953          * The check_recovery_mode call will see if there are needed
5954          * retransmissions, if so we will enter fast-recovery. The output
5955          * call may or may not do the same thing depending on sysctl
5956          * settings.
5957          */
5958         struct rack_sendmap *rsm;
5959
5960         if (tp->t_timers->tt_flags & TT_STOPPED) {
5961                 return (1);
5962         }
5963         counter_u64_add(rack_to_tot, 1);
5964         if (rack->r_state && (rack->r_state != tp->t_state))
5965                 rack_set_state(tp, rack);
5966         rack->rc_on_min_to = 0;
5967         rsm = rack_check_recovery_mode(tp, cts);
5968         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5969         if (rsm) {
5970                 rack->r_ctl.rc_resend = rsm;
5971                 rack->r_timer_override = 1;
5972                 if (rack->use_rack_rr) {
5973                         /*
5974                          * Don't accumulate extra pacing delay
5975                          * we are allowing the rack timer to
5976                          * over-ride pacing i.e. rrr takes precedence
5977                          * if the pacing interval is longer than the rrr
5978                          * time (in other words we get the min pacing
5979                          * time versus rrr pacing time).
5980                          */
5981                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5982                 }
5983         }
5984         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5985         if (rsm == NULL) {
5986                 /* restart a timer and return 1 */
5987                 rack_start_hpts_timer(rack, tp, cts,
5988                                       0, 0, 0);
5989                 return (1);
5990         }
5991         return (0);
5992 }
5993
5994 static void
5995 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5996 {
5997         if (rsm->m->m_len > rsm->orig_m_len) {
5998                 /*
5999                  * Mbuf grew, caused by sbcompress, our offset does
6000                  * not change.
6001                  */
6002                 rsm->orig_m_len = rsm->m->m_len;
6003         } else if (rsm->m->m_len < rsm->orig_m_len) {
6004                 /*
6005                  * Mbuf shrank, trimmed off the top by an ack, our
6006                  * offset changes.
6007                  */
6008                 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6009                 rsm->orig_m_len = rsm->m->m_len;
6010         }
6011 }
6012
6013 static void
6014 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6015 {
6016         struct mbuf *m;
6017         uint32_t soff;
6018
6019         if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
6020                 /* Fix up the orig_m_len and possibly the mbuf offset */
6021                 rack_adjust_orig_mlen(src_rsm);
6022         }
6023         m = src_rsm->m;
6024         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6025         while (soff >= m->m_len) {
6026                 /* Move out past this mbuf */
6027                 soff -= m->m_len;
6028                 m = m->m_next;
6029                 KASSERT((m != NULL),
6030                         ("rsm:%p nrsm:%p hit at soff:%u null m",
6031                          src_rsm, rsm, soff));
6032         }
6033         rsm->m = m;
6034         rsm->soff = soff;
6035         rsm->orig_m_len = m->m_len;
6036 }
6037
6038 static __inline void
6039 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
6040                struct rack_sendmap *rsm, uint32_t start)
6041 {
6042         int idx;
6043
6044         nrsm->r_start = start;
6045         nrsm->r_end = rsm->r_end;
6046         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6047         nrsm->r_flags = rsm->r_flags;
6048         nrsm->r_dupack = rsm->r_dupack;
6049         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6050         nrsm->r_rtr_bytes = 0;
6051         rsm->r_end = nrsm->r_start;
6052         nrsm->r_just_ret = rsm->r_just_ret;
6053         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6054                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6055         }
6056         /* Now if we have SYN flag we keep it on the left edge */
6057         if (nrsm->r_flags & RACK_HAS_SYN)
6058                 nrsm->r_flags &= ~RACK_HAS_SYN;
6059         /* Now if we have a FIN flag we keep it on the right edge */
6060         if (rsm->r_flags & RACK_HAS_FIN)
6061                 rsm->r_flags &= ~RACK_HAS_FIN;
6062         /* Push bit must go to the right edge as well */
6063         if (rsm->r_flags & RACK_HAD_PUSH)
6064                 rsm->r_flags &= ~RACK_HAD_PUSH;
6065         /* Clone over the state of the hw_tls flag */
6066         nrsm->r_hw_tls = rsm->r_hw_tls;
6067         /*
6068          * Now we need to find nrsm's new location in the mbuf chain
6069          * we basically calculate a new offset, which is soff +
6070          * how much is left in original rsm. Then we walk out the mbuf
6071          * chain to find the righ postion, it may be the same mbuf
6072          * or maybe not.
6073          */
6074         KASSERT(((rsm->m != NULL) ||
6075                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6076                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6077         if (rsm->m)
6078                 rack_setup_offset_for_rsm(rsm, nrsm);
6079 }
6080
6081 static struct rack_sendmap *
6082 rack_merge_rsm(struct tcp_rack *rack,
6083                struct rack_sendmap *l_rsm,
6084                struct rack_sendmap *r_rsm)
6085 {
6086         /*
6087          * We are merging two ack'd RSM's,
6088          * the l_rsm is on the left (lower seq
6089          * values) and the r_rsm is on the right
6090          * (higher seq value). The simplest way
6091          * to merge these is to move the right
6092          * one into the left. I don't think there
6093          * is any reason we need to try to find
6094          * the oldest (or last oldest retransmitted).
6095          */
6096         struct rack_sendmap *rm;
6097
6098         rack_log_map_chg(rack->rc_tp, rack, NULL,
6099                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6100         l_rsm->r_end = r_rsm->r_end;
6101         if (l_rsm->r_dupack < r_rsm->r_dupack)
6102                 l_rsm->r_dupack = r_rsm->r_dupack;
6103         if (r_rsm->r_rtr_bytes)
6104                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6105         if (r_rsm->r_in_tmap) {
6106                 /* This really should not happen */
6107                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6108                 r_rsm->r_in_tmap = 0;
6109         }
6110
6111         /* Now the flags */
6112         if (r_rsm->r_flags & RACK_HAS_FIN)
6113                 l_rsm->r_flags |= RACK_HAS_FIN;
6114         if (r_rsm->r_flags & RACK_TLP)
6115                 l_rsm->r_flags |= RACK_TLP;
6116         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6117                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6118         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6119             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6120                 /*
6121                  * If both are app-limited then let the
6122                  * free lower the count. If right is app
6123                  * limited and left is not, transfer.
6124                  */
6125                 l_rsm->r_flags |= RACK_APP_LIMITED;
6126                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
6127                 if (r_rsm == rack->r_ctl.rc_first_appl)
6128                         rack->r_ctl.rc_first_appl = l_rsm;
6129         }
6130         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6131 #ifdef INVARIANTS
6132         if (rm != r_rsm) {
6133                 panic("removing head in rack:%p rsm:%p rm:%p",
6134                       rack, r_rsm, rm);
6135         }
6136 #endif
6137         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6138                 /* Transfer the split limit to the map we free */
6139                 r_rsm->r_limit_type = l_rsm->r_limit_type;
6140                 l_rsm->r_limit_type = 0;
6141         }
6142         rack_free(rack, r_rsm);
6143         return (l_rsm);
6144 }
6145
6146 /*
6147  * TLP Timer, here we simply setup what segment we want to
6148  * have the TLP expire on, the normal rack_output() will then
6149  * send it out.
6150  *
6151  * We return 1, saying don't proceed with rack_output only
6152  * when all timers have been stopped (destroyed PCB?).
6153  */
6154 static int
6155 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6156 {
6157         /*
6158          * Tail Loss Probe.
6159          */
6160         struct rack_sendmap *rsm = NULL;
6161         struct rack_sendmap *insret;
6162         struct socket *so;
6163         uint32_t amm;
6164         uint32_t out, avail;
6165         int collapsed_win = 0;
6166
6167         if (tp->t_timers->tt_flags & TT_STOPPED) {
6168                 return (1);
6169         }
6170         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6171                 /* Its not time yet */
6172                 return (0);
6173         }
6174         if (ctf_progress_timeout_check(tp, true)) {
6175                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6176                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6177                 return (1);
6178         }
6179         /*
6180          * A TLP timer has expired. We have been idle for 2 rtts. So we now
6181          * need to figure out how to force a full MSS segment out.
6182          */
6183         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6184         rack->r_ctl.retran_during_recovery = 0;
6185         rack->r_ctl.dsack_byte_cnt = 0;
6186         counter_u64_add(rack_tlp_tot, 1);
6187         if (rack->r_state && (rack->r_state != tp->t_state))
6188                 rack_set_state(tp, rack);
6189         so = tp->t_inpcb->inp_socket;
6190         avail = sbavail(&so->so_snd);
6191         out = tp->snd_max - tp->snd_una;
6192         if (out > tp->snd_wnd) {
6193                 /* special case, we need a retransmission */
6194                 collapsed_win = 1;
6195                 goto need_retran;
6196         }
6197         /*
6198          * Check our send oldest always settings, and if
6199          * there is an oldest to send jump to the need_retran.
6200          */
6201         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6202                 goto need_retran;
6203
6204         if (avail > out) {
6205                 /* New data is available */
6206                 amm = avail - out;
6207                 if (amm > ctf_fixed_maxseg(tp)) {
6208                         amm = ctf_fixed_maxseg(tp);
6209                         if ((amm + out) > tp->snd_wnd) {
6210                                 /* We are rwnd limited */
6211                                 goto need_retran;
6212                         }
6213                 } else if (amm < ctf_fixed_maxseg(tp)) {
6214                         /* not enough to fill a MTU */
6215                         goto need_retran;
6216                 }
6217                 if (IN_FASTRECOVERY(tp->t_flags)) {
6218                         /* Unlikely */
6219                         if (rack->rack_no_prr == 0) {
6220                                 if (out + amm <= tp->snd_wnd) {
6221                                         rack->r_ctl.rc_prr_sndcnt = amm;
6222                                         rack_log_to_prr(rack, 4, 0);
6223                                 }
6224                         } else
6225                                 goto need_retran;
6226                 } else {
6227                         /* Set the send-new override */
6228                         if (out + amm <= tp->snd_wnd)
6229                                 rack->r_ctl.rc_tlp_new_data = amm;
6230                         else
6231                                 goto need_retran;
6232                 }
6233                 rack->r_ctl.rc_tlpsend = NULL;
6234                 counter_u64_add(rack_tlp_newdata, 1);
6235                 goto send;
6236         }
6237 need_retran:
6238         /*
6239          * Ok we need to arrange the last un-acked segment to be re-sent, or
6240          * optionally the first un-acked segment.
6241          */
6242         if (collapsed_win == 0) {
6243                 if (rack_always_send_oldest)
6244                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6245                 else {
6246                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6247                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6248                                 rsm = rack_find_high_nonack(rack, rsm);
6249                         }
6250                 }
6251                 if (rsm == NULL) {
6252                         counter_u64_add(rack_tlp_does_nada, 1);
6253 #ifdef TCP_BLACKBOX
6254                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6255 #endif
6256                         goto out;
6257                 }
6258         } else {
6259                 /*
6260                  * We must find the last segment
6261                  * that was acceptable by the client.
6262                  */
6263                 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6264                         if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6265                                 /* Found one */
6266                                 break;
6267                         }
6268                 }
6269                 if (rsm == NULL) {
6270                         /* None? if so send the first */
6271                         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6272                         if (rsm == NULL) {
6273                                 counter_u64_add(rack_tlp_does_nada, 1);
6274 #ifdef TCP_BLACKBOX
6275                                 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6276 #endif
6277                                 goto out;
6278                         }
6279                 }
6280         }
6281         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6282                 /*
6283                  * We need to split this the last segment in two.
6284                  */
6285                 struct rack_sendmap *nrsm;
6286
6287                 nrsm = rack_alloc_full_limit(rack);
6288                 if (nrsm == NULL) {
6289                         /*
6290                          * No memory to split, we will just exit and punt
6291                          * off to the RXT timer.
6292                          */
6293                         counter_u64_add(rack_tlp_does_nada, 1);
6294                         goto out;
6295                 }
6296                 rack_clone_rsm(rack, nrsm, rsm,
6297                                (rsm->r_end - ctf_fixed_maxseg(tp)));
6298                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6299                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6300 #ifdef INVARIANTS
6301                 if (insret != NULL) {
6302                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6303                               nrsm, insret, rack, rsm);
6304                 }
6305 #endif
6306                 if (rsm->r_in_tmap) {
6307                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6308                         nrsm->r_in_tmap = 1;
6309                 }
6310                 rsm->r_flags &= (~RACK_HAS_FIN);
6311                 rsm = nrsm;
6312         }
6313         rack->r_ctl.rc_tlpsend = rsm;
6314 send:
6315         rack->r_timer_override = 1;
6316         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6317         return (0);
6318 out:
6319         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6320         return (0);
6321 }
6322
6323 /*
6324  * Delayed ack Timer, here we simply need to setup the
6325  * ACK_NOW flag and remove the DELACK flag. From there
6326  * the output routine will send the ack out.
6327  *
6328  * We only return 1, saying don't proceed, if all timers
6329  * are stopped (destroyed PCB?).
6330  */
6331 static int
6332 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6333 {
6334         if (tp->t_timers->tt_flags & TT_STOPPED) {
6335                 return (1);
6336         }
6337         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6338         tp->t_flags &= ~TF_DELACK;
6339         tp->t_flags |= TF_ACKNOW;
6340         KMOD_TCPSTAT_INC(tcps_delack);
6341         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6342         return (0);
6343 }
6344
6345 /*
6346  * Persists timer, here we simply send the
6347  * same thing as a keepalive will.
6348  * the one byte send.
6349  *
6350  * We only return 1, saying don't proceed, if all timers
6351  * are stopped (destroyed PCB?).
6352  */
6353 static int
6354 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6355 {
6356         struct tcptemp *t_template;
6357         struct inpcb *inp;
6358         int32_t retval = 1;
6359
6360         inp = tp->t_inpcb;
6361
6362         if (tp->t_timers->tt_flags & TT_STOPPED) {
6363                 return (1);
6364         }
6365         if (rack->rc_in_persist == 0)
6366                 return (0);
6367         if (ctf_progress_timeout_check(tp, false)) {
6368                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6369                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6370                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6371                 return (1);
6372         }
6373         KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6374         /*
6375          * Persistence timer into zero window. Force a byte to be output, if
6376          * possible.
6377          */
6378         KMOD_TCPSTAT_INC(tcps_persisttimeo);
6379         /*
6380          * Hack: if the peer is dead/unreachable, we do not time out if the
6381          * window is closed.  After a full backoff, drop the connection if
6382          * the idle time (no responses to probes) reaches the maximum
6383          * backoff that we would use if retransmitting.
6384          */
6385         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6386             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6387              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6388                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6389                 retval = 1;
6390                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6391                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6392                 goto out;
6393         }
6394         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6395             tp->snd_una == tp->snd_max)
6396                 rack_exit_persist(tp, rack, cts);
6397         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6398         /*
6399          * If the user has closed the socket then drop a persisting
6400          * connection after a much reduced timeout.
6401          */
6402         if (tp->t_state > TCPS_CLOSE_WAIT &&
6403             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6404                 retval = 1;
6405                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6406                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6407                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6408                 goto out;
6409         }
6410         t_template = tcpip_maketemplate(rack->rc_inp);
6411         if (t_template) {
6412                 /* only set it if we were answered */
6413                 if (rack->forced_ack == 0) {
6414                         rack->forced_ack = 1;
6415                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6416                 }
6417                 tcp_respond(tp, t_template->tt_ipgen,
6418                             &t_template->tt_t, (struct mbuf *)NULL,
6419                             tp->rcv_nxt, tp->snd_una - 1, 0);
6420                 /* This sends an ack */
6421                 if (tp->t_flags & TF_DELACK)
6422                         tp->t_flags &= ~TF_DELACK;
6423                 free(t_template, M_TEMP);
6424         }
6425         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6426                 tp->t_rxtshift++;
6427 out:
6428         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6429         rack_start_hpts_timer(rack, tp, cts,
6430                               0, 0, 0);
6431         return (retval);
6432 }
6433
6434 /*
6435  * If a keepalive goes off, we had no other timers
6436  * happening. We always return 1 here since this
6437  * routine either drops the connection or sends
6438  * out a segment with respond.
6439  */
6440 static int
6441 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6442 {
6443         struct tcptemp *t_template;
6444         struct inpcb *inp;
6445
6446         if (tp->t_timers->tt_flags & TT_STOPPED) {
6447                 return (1);
6448         }
6449         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6450         inp = tp->t_inpcb;
6451         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6452         /*
6453          * Keep-alive timer went off; send something or drop connection if
6454          * idle for too long.
6455          */
6456         KMOD_TCPSTAT_INC(tcps_keeptimeo);
6457         if (tp->t_state < TCPS_ESTABLISHED)
6458                 goto dropit;
6459         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6460             tp->t_state <= TCPS_CLOSING) {
6461                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6462                         goto dropit;
6463                 /*
6464                  * Send a packet designed to force a response if the peer is
6465                  * up and reachable: either an ACK if the connection is
6466                  * still alive, or an RST if the peer has closed the
6467                  * connection due to timeout or reboot. Using sequence
6468                  * number tp->snd_una-1 causes the transmitted zero-length
6469                  * segment to lie outside the receive window; by the
6470                  * protocol spec, this requires the correspondent TCP to
6471                  * respond.
6472                  */
6473                 KMOD_TCPSTAT_INC(tcps_keepprobe);
6474                 t_template = tcpip_maketemplate(inp);
6475                 if (t_template) {
6476                         if (rack->forced_ack == 0) {
6477                                 rack->forced_ack = 1;
6478                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6479                         }
6480                         tcp_respond(tp, t_template->tt_ipgen,
6481                             &t_template->tt_t, (struct mbuf *)NULL,
6482                             tp->rcv_nxt, tp->snd_una - 1, 0);
6483                         free(t_template, M_TEMP);
6484                 }
6485         }
6486         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6487         return (1);
6488 dropit:
6489         KMOD_TCPSTAT_INC(tcps_keepdrops);
6490         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6491         tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6492         return (1);
6493 }
6494
6495 /*
6496  * Retransmit helper function, clear up all the ack
6497  * flags and take care of important book keeping.
6498  */
6499 static void
6500 rack_remxt_tmr(struct tcpcb *tp)
6501 {
6502         /*
6503          * The retransmit timer went off, all sack'd blocks must be
6504          * un-acked.
6505          */
6506         struct rack_sendmap *rsm, *trsm = NULL;
6507         struct tcp_rack *rack;
6508
6509         rack = (struct tcp_rack *)tp->t_fb_ptr;
6510         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6511         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6512         if (rack->r_state && (rack->r_state != tp->t_state))
6513                 rack_set_state(tp, rack);
6514         /*
6515          * Ideally we would like to be able to
6516          * mark SACK-PASS on anything not acked here.
6517          *
6518          * However, if we do that we would burst out
6519          * all that data 1ms apart. This would be unwise,
6520          * so for now we will just let the normal rxt timer
6521          * and tlp timer take care of it.
6522          *
6523          * Also we really need to stick them back in sequence
6524          * order. This way we send in the proper order and any
6525          * sacks that come floating in will "re-ack" the data.
6526          * To do this we zap the tmap with an INIT and then
6527          * walk through and place every rsm in the RB tree
6528          * back in its seq ordered place.
6529          */
6530         TAILQ_INIT(&rack->r_ctl.rc_tmap);
6531         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6532                 rsm->r_dupack = 0;
6533                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6534                 /* We must re-add it back to the tlist */
6535                 if (trsm == NULL) {
6536                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6537                 } else {
6538                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6539                 }
6540                 rsm->r_in_tmap = 1;
6541                 trsm = rsm;
6542                 if (rsm->r_flags & RACK_ACKED)
6543                         rsm->r_flags |= RACK_WAS_ACKED;
6544                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6545         }
6546         /* Clear the count (we just un-acked them) */
6547         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6548         rack->r_ctl.rc_sacked = 0;
6549         rack->r_ctl.rc_sacklast = NULL;
6550         rack->r_ctl.rc_agg_delayed = 0;
6551         rack->r_early = 0;
6552         rack->r_ctl.rc_agg_early = 0;
6553         rack->r_late = 0;
6554         /* Clear the tlp rtx mark */
6555         rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6556         if (rack->r_ctl.rc_resend != NULL)
6557                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6558         rack->r_ctl.rc_prr_sndcnt = 0;
6559         rack_log_to_prr(rack, 6, 0);
6560         rack->r_timer_override = 1;
6561         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6562 #ifdef NETFLIX_EXP_DETECTION
6563             || (rack->sack_attack_disable != 0)
6564 #endif
6565                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6566                 /*
6567                  * For non-sack customers new data
6568                  * needs to go out as retransmits until
6569                  * we retransmit up to snd_max.
6570                  */
6571                 rack->r_must_retran = 1;
6572                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6573                                                 rack->r_ctl.rc_sacked);
6574         }
6575         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6576 }
6577
6578 static void
6579 rack_convert_rtts(struct tcpcb *tp)
6580 {
6581         if (tp->t_srtt > 1) {
6582                 uint32_t val, frac;
6583
6584                 val = tp->t_srtt >> TCP_RTT_SHIFT;
6585                 frac = tp->t_srtt & 0x1f;
6586                 tp->t_srtt = TICKS_2_USEC(val);
6587                 /*
6588                  * frac is the fractional part of the srtt (if any)
6589                  * but its in ticks and every bit represents
6590                  * 1/32nd of a hz.
6591                  */
6592                 if (frac) {
6593                         if (hz == 1000) {
6594                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6595                         } else {
6596                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6597                         }
6598                         tp->t_srtt += frac;
6599                 }
6600         }
6601         if (tp->t_rttvar) {
6602                 uint32_t val, frac;
6603
6604                 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6605                 frac = tp->t_rttvar & 0x1f;
6606                 tp->t_rttvar = TICKS_2_USEC(val);
6607                 /*
6608                  * frac is the fractional part of the srtt (if any)
6609                  * but its in ticks and every bit represents
6610                  * 1/32nd of a hz.
6611                  */
6612                 if (frac) {
6613                         if (hz == 1000) {
6614                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6615                         } else {
6616                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6617                         }
6618                         tp->t_rttvar += frac;
6619                 }
6620         }
6621         tp->t_rxtcur = RACK_REXMTVAL(tp);
6622         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6623                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6624         }
6625         if (tp->t_rxtcur > rack_rto_max) {
6626                 tp->t_rxtcur = rack_rto_max;
6627         }
6628 }
6629
6630 static void
6631 rack_cc_conn_init(struct tcpcb *tp)
6632 {
6633         struct tcp_rack *rack;
6634         uint32_t srtt;
6635
6636         rack = (struct tcp_rack *)tp->t_fb_ptr;
6637         srtt = tp->t_srtt;
6638         cc_conn_init(tp);
6639         /*
6640          * Now convert to rack's internal format,
6641          * if required.
6642          */
6643         if ((srtt == 0) && (tp->t_srtt != 0))
6644                 rack_convert_rtts(tp);
6645         /*
6646          * We want a chance to stay in slowstart as
6647          * we create a connection. TCP spec says that
6648          * initially ssthresh is infinite. For our
6649          * purposes that is the snd_wnd.
6650          */
6651         if (tp->snd_ssthresh < tp->snd_wnd) {
6652                 tp->snd_ssthresh = tp->snd_wnd;
6653         }
6654         /*
6655          * We also want to assure a IW worth of
6656          * data can get inflight.
6657          */
6658         if (rc_init_window(rack) < tp->snd_cwnd)
6659                 tp->snd_cwnd = rc_init_window(rack);
6660 }
6661
6662 /*
6663  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6664  * we will setup to retransmit the lowest seq number outstanding.
6665  */
6666 static int
6667 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6668 {
6669         int32_t rexmt;
6670         struct inpcb *inp;
6671         int32_t retval = 0;
6672         bool isipv6;
6673
6674         inp = tp->t_inpcb;
6675         if (tp->t_timers->tt_flags & TT_STOPPED) {
6676                 return (1);
6677         }
6678         if (ctf_progress_timeout_check(tp, false)) {
6679                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6680                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6681                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6682                 return (1);
6683         }
6684         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6685         rack->r_ctl.retran_during_recovery = 0;
6686         rack->r_ctl.dsack_byte_cnt = 0;
6687         if (IN_FASTRECOVERY(tp->t_flags))
6688                 tp->t_flags |= TF_WASFRECOVERY;
6689         else
6690                 tp->t_flags &= ~TF_WASFRECOVERY;
6691         if (IN_CONGRECOVERY(tp->t_flags))
6692                 tp->t_flags |= TF_WASCRECOVERY;
6693         else
6694                 tp->t_flags &= ~TF_WASCRECOVERY;
6695         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6696             (tp->snd_una == tp->snd_max)) {
6697                 /* Nothing outstanding .. nothing to do */
6698                 return (0);
6699         }
6700         /*
6701          * Rack can only run one timer  at a time, so we cannot
6702          * run a KEEPINIT (gating SYN sending) and a retransmit
6703          * timer for the SYN. So if we are in a front state and
6704          * have a KEEPINIT timer we need to check the first transmit
6705          * against now to see if we have exceeded the KEEPINIT time
6706          * (if one is set).
6707          */
6708         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6709             (TP_KEEPINIT(tp) != 0)) {
6710                 struct rack_sendmap *rsm;
6711
6712                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6713                 if (rsm) {
6714                         /* Ok we have something outstanding to test keepinit with */
6715                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6716                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6717                                 /* We have exceeded the KEEPINIT time */
6718                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6719                                 goto drop_it;
6720                         }
6721                 }
6722         }
6723         /*
6724          * Retransmission timer went off.  Message has not been acked within
6725          * retransmit interval.  Back off to a longer retransmit interval
6726          * and retransmit one segment.
6727          */
6728         rack_remxt_tmr(tp);
6729         if ((rack->r_ctl.rc_resend == NULL) ||
6730             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6731                 /*
6732                  * If the rwnd collapsed on
6733                  * the one we are retransmitting
6734                  * it does not count against the
6735                  * rxt count.
6736                  */
6737                 tp->t_rxtshift++;
6738         }
6739         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6740                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6741 drop_it:
6742                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6743                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6744                 retval = 1;
6745                 tcp_set_inp_to_drop(rack->rc_inp,
6746                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
6747                 goto out;
6748         }
6749         if (tp->t_state == TCPS_SYN_SENT) {
6750                 /*
6751                  * If the SYN was retransmitted, indicate CWND to be limited
6752                  * to 1 segment in cc_conn_init().
6753                  */
6754                 tp->snd_cwnd = 1;
6755         } else if (tp->t_rxtshift == 1) {
6756                 /*
6757                  * first retransmit; record ssthresh and cwnd so they can be
6758                  * recovered if this turns out to be a "bad" retransmit. A
6759                  * retransmit is considered "bad" if an ACK for this segment
6760                  * is received within RTT/2 interval; the assumption here is
6761                  * that the ACK was already in flight.  See "On Estimating
6762                  * End-to-End Network Path Properties" by Allman and Paxson
6763                  * for more details.
6764                  */
6765                 tp->snd_cwnd_prev = tp->snd_cwnd;
6766                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6767                 tp->snd_recover_prev = tp->snd_recover;
6768                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6769                 tp->t_flags |= TF_PREVVALID;
6770         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6771                 tp->t_flags &= ~TF_PREVVALID;
6772         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6773         if ((tp->t_state == TCPS_SYN_SENT) ||
6774             (tp->t_state == TCPS_SYN_RECEIVED))
6775                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6776         else
6777                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6778
6779         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6780            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6781         /*
6782          * We enter the path for PLMTUD if connection is established or, if
6783          * connection is FIN_WAIT_1 status, reason for the last is that if
6784          * amount of data we send is very small, we could send it in couple
6785          * of packets and process straight to FIN. In that case we won't
6786          * catch ESTABLISHED state.
6787          */
6788 #ifdef INET6
6789         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6790 #else
6791         isipv6 = false;
6792 #endif
6793         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6794             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6795             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6796             ((tp->t_state == TCPS_ESTABLISHED) ||
6797             (tp->t_state == TCPS_FIN_WAIT_1))) {
6798                 /*
6799                  * Idea here is that at each stage of mtu probe (usually,
6800                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6801                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6802                  * should take care of that.
6803                  */
6804                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6805                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6806                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6807                     tp->t_rxtshift % 2 == 0)) {
6808                         /*
6809                          * Enter Path MTU Black-hole Detection mechanism: -
6810                          * Disable Path MTU Discovery (IP "DF" bit). -
6811                          * Reduce MTU to lower value than what we negotiated
6812                          * with peer.
6813                          */
6814                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6815                                 /* Record that we may have found a black hole. */
6816                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6817                                 /* Keep track of previous MSS. */
6818                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6819                         }
6820
6821                         /*
6822                          * Reduce the MSS to blackhole value or to the
6823                          * default in an attempt to retransmit.
6824                          */
6825 #ifdef INET6
6826                         if (isipv6 &&
6827                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6828                                 /* Use the sysctl tuneable blackhole MSS. */
6829                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6830                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6831                         } else if (isipv6) {
6832                                 /* Use the default MSS. */
6833                                 tp->t_maxseg = V_tcp_v6mssdflt;
6834                                 /*
6835                                  * Disable Path MTU Discovery when we switch
6836                                  * to minmss.
6837                                  */
6838                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6839                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6840                         }
6841 #endif
6842 #if defined(INET6) && defined(INET)
6843                         else
6844 #endif
6845 #ifdef INET
6846                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6847                                 /* Use the sysctl tuneable blackhole MSS. */
6848                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6849                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6850                         } else {
6851                                 /* Use the default MSS. */
6852                                 tp->t_maxseg = V_tcp_mssdflt;
6853                                 /*
6854                                  * Disable Path MTU Discovery when we switch
6855                                  * to minmss.
6856                                  */
6857                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6858                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6859                         }
6860 #endif
6861                 } else {
6862                         /*
6863                          * If further retransmissions are still unsuccessful
6864                          * with a lowered MTU, maybe this isn't a blackhole
6865                          * and we restore the previous MSS and blackhole
6866                          * detection flags. The limit '6' is determined by
6867                          * giving each probe stage (1448, 1188, 524) 2
6868                          * chances to recover.
6869                          */
6870                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6871                             (tp->t_rxtshift >= 6)) {
6872                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6873                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6874                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6875                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6876                         }
6877                 }
6878         }
6879         /*
6880          * Disable RFC1323 and SACK if we haven't got any response to
6881          * our third SYN to work-around some broken terminal servers
6882          * (most of which have hopefully been retired) that have bad VJ
6883          * header compression code which trashes TCP segments containing
6884          * unknown-to-them TCP options.
6885          */
6886         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6887             (tp->t_rxtshift == 3))
6888                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6889         /*
6890          * If we backed off this far, our srtt estimate is probably bogus.
6891          * Clobber it so we'll take the next rtt measurement as our srtt;
6892          * move the current srtt into rttvar to keep the current retransmit
6893          * times until then.
6894          */
6895         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6896 #ifdef INET6
6897                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6898                         in6_losing(tp->t_inpcb);
6899                 else
6900 #endif
6901                         in_losing(tp->t_inpcb);
6902                 tp->t_rttvar += tp->t_srtt;
6903                 tp->t_srtt = 0;
6904         }
6905         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6906         tp->snd_recover = tp->snd_max;
6907         tp->t_flags |= TF_ACKNOW;
6908         tp->t_rtttime = 0;
6909         rack_cong_signal(tp, CC_RTO, tp->snd_una);
6910 out:
6911         return (retval);
6912 }
6913
6914 static int
6915 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling)
6916 {
6917         int32_t ret = 0;
6918         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6919
6920         if (timers == 0) {
6921                 return (0);
6922         }
6923         if (tp->t_state == TCPS_LISTEN) {
6924                 /* no timers on listen sockets */
6925                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6926                         return (0);
6927                 return (1);
6928         }
6929         if ((timers & PACE_TMR_RACK) &&
6930             rack->rc_on_min_to) {
6931                 /*
6932                  * For the rack timer when we
6933                  * are on a min-timeout (which means rrr_conf = 3)
6934                  * we don't want to check the timer. It may
6935                  * be going off for a pace and thats ok we
6936                  * want to send the retransmit (if its ready).
6937                  *
6938                  * If its on a normal rack timer (non-min) then
6939                  * we will check if its expired.
6940                  */
6941                 goto skip_time_check;
6942         }
6943         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6944                 uint32_t left;
6945
6946                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6947                         ret = -1;
6948                         rack_log_to_processing(rack, cts, ret, 0);
6949                         return (0);
6950                 }
6951                 if (hpts_calling == 0) {
6952                         /*
6953                          * A user send or queued mbuf (sack) has called us? We
6954                          * return 0 and let the pacing guards
6955                          * deal with it if they should or
6956                          * should not cause a send.
6957                          */
6958                         ret = -2;
6959                         rack_log_to_processing(rack, cts, ret, 0);
6960                         return (0);
6961                 }
6962                 /*
6963                  * Ok our timer went off early and we are not paced false
6964                  * alarm, go back to sleep.
6965                  */
6966                 ret = -3;
6967                 left = rack->r_ctl.rc_timer_exp - cts;
6968                 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
6969                 rack_log_to_processing(rack, cts, ret, left);
6970                 return (1);
6971         }
6972 skip_time_check:
6973         rack->rc_tmr_stopped = 0;
6974         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6975         if (timers & PACE_TMR_DELACK) {
6976                 ret = rack_timeout_delack(tp, rack, cts);
6977         } else if (timers & PACE_TMR_RACK) {
6978                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6979                 rack->r_fast_output = 0;
6980                 ret = rack_timeout_rack(tp, rack, cts);
6981         } else if (timers & PACE_TMR_TLP) {
6982                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6983                 ret = rack_timeout_tlp(tp, rack, cts);
6984         } else if (timers & PACE_TMR_RXT) {
6985                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6986                 rack->r_fast_output = 0;
6987                 ret = rack_timeout_rxt(tp, rack, cts);
6988         } else if (timers & PACE_TMR_PERSIT) {
6989                 ret = rack_timeout_persist(tp, rack, cts);
6990         } else if (timers & PACE_TMR_KEEP) {
6991                 ret = rack_timeout_keepalive(tp, rack, cts);
6992         }
6993         rack_log_to_processing(rack, cts, ret, timers);
6994         return (ret);
6995 }
6996
6997 static void
6998 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6999 {
7000         struct timeval tv;
7001         uint32_t us_cts, flags_on_entry;
7002         uint8_t hpts_removed = 0;
7003
7004         flags_on_entry = rack->r_ctl.rc_hpts_flags;
7005         us_cts = tcp_get_usecs(&tv);
7006         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7007             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7008              ((tp->snd_max - tp->snd_una) == 0))) {
7009                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7010                 hpts_removed = 1;
7011                 /* If we were not delayed cancel out the flag. */
7012                 if ((tp->snd_max - tp->snd_una) == 0)
7013                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7014                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7015         }
7016         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7017                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7018                 if (rack->rc_inp->inp_in_hpts &&
7019                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7020                         /*
7021                          * Canceling timer's when we have no output being
7022                          * paced. We also must remove ourselves from the
7023                          * hpts.
7024                          */
7025                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7026                         hpts_removed = 1;
7027                 }
7028                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7029         }
7030         if (hpts_removed == 0)
7031                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7032 }
7033
7034 static void
7035 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7036 {
7037         return;
7038 }
7039
7040 static int
7041 rack_stopall(struct tcpcb *tp)
7042 {
7043         struct tcp_rack *rack;
7044         rack = (struct tcp_rack *)tp->t_fb_ptr;
7045         rack->t_timers_stopped = 1;
7046         return (0);
7047 }
7048
7049 static void
7050 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7051 {
7052         return;
7053 }
7054
7055 static int
7056 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7057 {
7058         return (0);
7059 }
7060
7061 static void
7062 rack_stop_all_timers(struct tcpcb *tp)
7063 {
7064         struct tcp_rack *rack;
7065
7066         /*
7067          * Assure no timers are running.
7068          */
7069         if (tcp_timer_active(tp, TT_PERSIST)) {
7070                 /* We enter in persists, set the flag appropriately */
7071                 rack = (struct tcp_rack *)tp->t_fb_ptr;
7072                 rack->rc_in_persist = 1;
7073         }
7074         tcp_timer_suspend(tp, TT_PERSIST);
7075         tcp_timer_suspend(tp, TT_REXMT);
7076         tcp_timer_suspend(tp, TT_KEEP);
7077         tcp_timer_suspend(tp, TT_DELACK);
7078 }
7079
7080 static void
7081 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7082     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7083 {
7084         int32_t idx;
7085         uint16_t stripped_flags;
7086
7087         rsm->r_rtr_cnt++;
7088         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7089         rsm->r_dupack = 0;
7090         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7091                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7092                 rsm->r_flags |= RACK_OVERMAX;
7093         }
7094         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7095                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7096                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7097         }
7098         idx = rsm->r_rtr_cnt - 1;
7099         rsm->r_tim_lastsent[idx] = ts;
7100         stripped_flags = rsm->r_flags & ~(RACK_SENT_SP|RACK_SENT_FP);
7101         if (rsm->r_flags & RACK_ACKED) {
7102                 /* Problably MTU discovery messing with us */
7103                 rsm->r_flags &= ~RACK_ACKED;
7104                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7105         }
7106         if (rsm->r_in_tmap) {
7107                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7108                 rsm->r_in_tmap = 0;
7109         }
7110         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7111         rsm->r_in_tmap = 1;
7112         if (rsm->r_flags & RACK_SACK_PASSED) {
7113                 /* We have retransmitted due to the SACK pass */
7114                 rsm->r_flags &= ~RACK_SACK_PASSED;
7115                 rsm->r_flags |= RACK_WAS_SACKPASS;
7116         }
7117 }
7118
7119 static uint32_t
7120 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7121     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7122 {
7123         /*
7124          * We (re-)transmitted starting at rsm->r_start for some length
7125          * (possibly less than r_end.
7126          */
7127         struct rack_sendmap *nrsm, *insret;
7128         uint32_t c_end;
7129         int32_t len;
7130
7131         len = *lenp;
7132         c_end = rsm->r_start + len;
7133         if (SEQ_GEQ(c_end, rsm->r_end)) {
7134                 /*
7135                  * We retransmitted the whole piece or more than the whole
7136                  * slopping into the next rsm.
7137                  */
7138                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7139                 if (c_end == rsm->r_end) {
7140                         *lenp = 0;
7141                         return (0);
7142                 } else {
7143                         int32_t act_len;
7144
7145                         /* Hangs over the end return whats left */
7146                         act_len = rsm->r_end - rsm->r_start;
7147                         *lenp = (len - act_len);
7148                         return (rsm->r_end);
7149                 }
7150                 /* We don't get out of this block. */
7151         }
7152         /*
7153          * Here we retransmitted less than the whole thing which means we
7154          * have to split this into what was transmitted and what was not.
7155          */
7156         nrsm = rack_alloc_full_limit(rack);
7157         if (nrsm == NULL) {
7158                 /*
7159                  * We can't get memory, so lets not proceed.
7160                  */
7161                 *lenp = 0;
7162                 return (0);
7163         }
7164         /*
7165          * So here we are going to take the original rsm and make it what we
7166          * retransmitted. nrsm will be the tail portion we did not
7167          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7168          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7169          * 1, 6 and the new piece will be 6, 11.
7170          */
7171         rack_clone_rsm(rack, nrsm, rsm, c_end);
7172         nrsm->r_dupack = 0;
7173         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7174         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7175 #ifdef INVARIANTS
7176         if (insret != NULL) {
7177                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7178                       nrsm, insret, rack, rsm);
7179         }
7180 #endif
7181         if (rsm->r_in_tmap) {
7182                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7183                 nrsm->r_in_tmap = 1;
7184         }
7185         rsm->r_flags &= (~RACK_HAS_FIN);
7186         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7187         /* Log a split of rsm into rsm and nrsm */
7188         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7189         *lenp = 0;
7190         return (0);
7191 }
7192
7193 static void
7194 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7195                 uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts,
7196                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7197 {
7198         struct tcp_rack *rack;
7199         struct rack_sendmap *rsm, *nrsm, *insret, fe;
7200         register uint32_t snd_max, snd_una;
7201
7202         /*
7203          * Add to the RACK log of packets in flight or retransmitted. If
7204          * there is a TS option we will use the TS echoed, if not we will
7205          * grab a TS.
7206          *
7207          * Retransmissions will increment the count and move the ts to its
7208          * proper place. Note that if options do not include TS's then we
7209          * won't be able to effectively use the ACK for an RTT on a retran.
7210          *
7211          * Notes about r_start and r_end. Lets consider a send starting at
7212          * sequence 1 for 10 bytes. In such an example the r_start would be
7213          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7214          * This means that r_end is actually the first sequence for the next
7215          * slot (11).
7216          *
7217          */
7218         /*
7219          * If err is set what do we do XXXrrs? should we not add the thing?
7220          * -- i.e. return if err != 0 or should we pretend we sent it? --
7221          * i.e. proceed with add ** do this for now.
7222          */
7223         INP_WLOCK_ASSERT(tp->t_inpcb);
7224         if (err)
7225                 /*
7226                  * We don't log errors -- we could but snd_max does not
7227                  * advance in this case either.
7228                  */
7229                 return;
7230
7231         if (th_flags & TH_RST) {
7232                 /*
7233                  * We don't log resets and we return immediately from
7234                  * sending
7235                  */
7236                 return;
7237         }
7238         rack = (struct tcp_rack *)tp->t_fb_ptr;
7239         snd_una = tp->snd_una;
7240         snd_max = tp->snd_max;
7241         if (th_flags & (TH_SYN | TH_FIN)) {
7242                 /*
7243                  * The call to rack_log_output is made before bumping
7244                  * snd_max. This means we can record one extra byte on a SYN
7245                  * or FIN if seq_out is adding more on and a FIN is present
7246                  * (and we are not resending).
7247                  */
7248                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7249                         len++;
7250                 if (th_flags & TH_FIN)
7251                         len++;
7252                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7253                         /*
7254                          * The add/update as not been done for the FIN/SYN
7255                          * yet.
7256                          */
7257                         snd_max = tp->snd_nxt;
7258                 }
7259         }
7260         if (SEQ_LEQ((seq_out + len), snd_una)) {
7261                 /* Are sending an old segment to induce an ack (keep-alive)? */
7262                 return;
7263         }
7264         if (SEQ_LT(seq_out, snd_una)) {
7265                 /* huh? should we panic? */
7266                 uint32_t end;
7267
7268                 end = seq_out + len;
7269                 seq_out = snd_una;
7270                 if (SEQ_GEQ(end, seq_out))
7271                         len = end - seq_out;
7272                 else
7273                         len = 0;
7274         }
7275         if (len == 0) {
7276                 /* We don't log zero window probes */
7277                 return;
7278         }
7279         rack->r_ctl.rc_time_last_sent = cts;
7280         if (IN_FASTRECOVERY(tp->t_flags)) {
7281                 rack->r_ctl.rc_prr_out += len;
7282         }
7283         /* First question is it a retransmission or new? */
7284         if (seq_out == snd_max) {
7285                 /* Its new */
7286 again:
7287                 rsm = rack_alloc(rack);
7288                 if (rsm == NULL) {
7289                         /*
7290                          * Hmm out of memory and the tcb got destroyed while
7291                          * we tried to wait.
7292                          */
7293                         return;
7294                 }
7295                 if (th_flags & TH_FIN) {
7296                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7297                 } else {
7298                         rsm->r_flags = add_flag;
7299                 }
7300                 if (hw_tls)
7301                         rsm->r_hw_tls = 1;
7302                 rsm->r_tim_lastsent[0] = cts;
7303                 rsm->r_rtr_cnt = 1;
7304                 rsm->r_rtr_bytes = 0;
7305                 if (th_flags & TH_SYN) {
7306                         /* The data space is one beyond snd_una */
7307                         rsm->r_flags |= RACK_HAS_SYN;
7308                 }
7309                 rsm->r_start = seq_out;
7310                 rsm->r_end = rsm->r_start + len;
7311                 rsm->r_dupack = 0;
7312                 /*
7313                  * save off the mbuf location that
7314                  * sndmbuf_noadv returned (which is
7315                  * where we started copying from)..
7316                  */
7317                 rsm->m = s_mb;
7318                 rsm->soff = s_moff;
7319                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7320                 if (rsm->m) {
7321                         if (rsm->m->m_len <= rsm->soff) {
7322                                 /*
7323                                  * XXXrrs Question, will this happen?
7324                                  *
7325                                  * If sbsndptr is set at the correct place
7326                                  * then s_moff should always be somewhere
7327                                  * within rsm->m. But if the sbsndptr was
7328                                  * off then that won't be true. If it occurs
7329                                  * we need to walkout to the correct location.
7330                                  */
7331                                 struct mbuf *lm;
7332
7333                                 lm = rsm->m;
7334                                 while (lm->m_len <= rsm->soff) {
7335                                         rsm->soff -= lm->m_len;
7336                                         lm = lm->m_next;
7337                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7338                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7339                                 }
7340                                 rsm->m = lm;
7341                                 counter_u64_add(rack_sbsndptr_wrong, 1);
7342                         } else
7343                                 counter_u64_add(rack_sbsndptr_right, 1);
7344                         rsm->orig_m_len = rsm->m->m_len;
7345                 } else
7346                         rsm->orig_m_len = 0;
7347                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7348                 /* Log a new rsm */
7349                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7350                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7351 #ifdef INVARIANTS
7352                 if (insret != NULL) {
7353                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7354                               nrsm, insret, rack, rsm);
7355                 }
7356 #endif
7357                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7358                 rsm->r_in_tmap = 1;
7359                 /*
7360                  * Special case detection, is there just a single
7361                  * packet outstanding when we are not in recovery?
7362                  *
7363                  * If this is true mark it so.
7364                  */
7365                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7366                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7367                         struct rack_sendmap *prsm;
7368
7369                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7370                         if (prsm)
7371                                 prsm->r_one_out_nr = 1;
7372                 }
7373                 return;
7374         }
7375         /*
7376          * If we reach here its a retransmission and we need to find it.
7377          */
7378         memset(&fe, 0, sizeof(fe));
7379 more:
7380         if (hintrsm && (hintrsm->r_start == seq_out)) {
7381                 rsm = hintrsm;
7382                 hintrsm = NULL;
7383         } else {
7384                 /* No hints sorry */
7385                 rsm = NULL;
7386         }
7387         if ((rsm) && (rsm->r_start == seq_out)) {
7388                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7389                 if (len == 0) {
7390                         return;
7391                 } else {
7392                         goto more;
7393                 }
7394         }
7395         /* Ok it was not the last pointer go through it the hard way. */
7396 refind:
7397         fe.r_start = seq_out;
7398         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7399         if (rsm) {
7400                 if (rsm->r_start == seq_out) {
7401                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7402                         if (len == 0) {
7403                                 return;
7404                         } else {
7405                                 goto refind;
7406                         }
7407                 }
7408                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7409                         /* Transmitted within this piece */
7410                         /*
7411                          * Ok we must split off the front and then let the
7412                          * update do the rest
7413                          */
7414                         nrsm = rack_alloc_full_limit(rack);
7415                         if (nrsm == NULL) {
7416                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7417                                 return;
7418                         }
7419                         /*
7420                          * copy rsm to nrsm and then trim the front of rsm
7421                          * to not include this part.
7422                          */
7423                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7424                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7425                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7426 #ifdef INVARIANTS
7427                         if (insret != NULL) {
7428                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7429                                       nrsm, insret, rack, rsm);
7430                         }
7431 #endif
7432                         if (rsm->r_in_tmap) {
7433                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7434                                 nrsm->r_in_tmap = 1;
7435                         }
7436                         rsm->r_flags &= (~RACK_HAS_FIN);
7437                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7438                         if (len == 0) {
7439                                 return;
7440                         } else if (len > 0)
7441                                 goto refind;
7442                 }
7443         }
7444         /*
7445          * Hmm not found in map did they retransmit both old and on into the
7446          * new?
7447          */
7448         if (seq_out == tp->snd_max) {
7449                 goto again;
7450         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7451 #ifdef INVARIANTS
7452                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7453                        seq_out, len, tp->snd_una, tp->snd_max);
7454                 printf("Starting Dump of all rack entries\n");
7455                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7456                         printf("rsm:%p start:%u end:%u\n",
7457                                rsm, rsm->r_start, rsm->r_end);
7458                 }
7459                 printf("Dump complete\n");
7460                 panic("seq_out not found rack:%p tp:%p",
7461                       rack, tp);
7462 #endif
7463         } else {
7464 #ifdef INVARIANTS
7465                 /*
7466                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7467                  * flag)
7468                  */
7469                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7470                       seq_out, len, tp->snd_max, tp);
7471 #endif
7472         }
7473 }
7474
7475 /*
7476  * Record one of the RTT updates from an ack into
7477  * our sample structure.
7478  */
7479
7480 static void
7481 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7482                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7483 {
7484         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7485             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7486                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7487         }
7488         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7489             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7490                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7491         }
7492         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7493             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7494                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7495             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7496                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7497         }
7498         if ((confidence == 1) &&
7499             ((rsm == NULL) ||
7500              (rsm->r_just_ret) ||
7501              (rsm->r_one_out_nr &&
7502               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7503                 /*
7504                  * If the rsm had a just return
7505                  * hit it then we can't trust the
7506                  * rtt measurement for buffer deterimination
7507                  * Note that a confidence of 2, indicates
7508                  * SACK'd which overrides the r_just_ret or
7509                  * the r_one_out_nr. If it was a CUM-ACK and
7510                  * we had only two outstanding, but get an
7511                  * ack for only 1. Then that also lowers our
7512                  * confidence.
7513                  */
7514                 confidence = 0;
7515         }
7516         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7517             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7518                 if (rack->r_ctl.rack_rs.confidence == 0) {
7519                         /*
7520                          * We take anything with no current confidence
7521                          * saved.
7522                          */
7523                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7524                         rack->r_ctl.rack_rs.confidence = confidence;
7525                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7526                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7527                         /*
7528                          * Once we have a confident number,
7529                          * we can update it with a smaller
7530                          * value since this confident number
7531                          * may include the DSACK time until
7532                          * the next segment (the second one) arrived.
7533                          */
7534                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7535                         rack->r_ctl.rack_rs.confidence = confidence;
7536                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7537                 }
7538         }
7539         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7540         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7541         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7542         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7543 }
7544
7545 /*
7546  * Collect new round-trip time estimate
7547  * and update averages and current timeout.
7548  */
7549 static void
7550 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7551 {
7552         int32_t delta;
7553         uint32_t o_srtt, o_var;
7554         int32_t hrtt_up = 0;
7555         int32_t rtt;
7556
7557         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7558                 /* No valid sample */
7559                 return;
7560         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7561                 /* We are to use the lowest RTT seen in a single ack */
7562                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7563         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7564                 /* We are to use the highest RTT seen in a single ack */
7565                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7566         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7567                 /* We are to use the average RTT seen in a single ack */
7568                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7569                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7570         } else {
7571 #ifdef INVARIANTS
7572                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7573 #endif
7574                 return;
7575         }
7576         if (rtt == 0)
7577                 rtt = 1;
7578         if (rack->rc_gp_rtt_set == 0) {
7579                 /*
7580                  * With no RTT we have to accept
7581                  * even one we are not confident of.
7582                  */
7583                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7584                 rack->rc_gp_rtt_set = 1;
7585         } else if (rack->r_ctl.rack_rs.confidence) {
7586                 /* update the running gp srtt */
7587                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7588                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7589         }
7590         if (rack->r_ctl.rack_rs.confidence) {
7591                 /*
7592                  * record the low and high for highly buffered path computation,
7593                  * we only do this if we are confident (not a retransmission).
7594                  */
7595                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7596                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7597                         hrtt_up = 1;
7598                 }
7599                 if (rack->rc_highly_buffered == 0) {
7600                         /*
7601                          * Currently once we declare a path has
7602                          * highly buffered there is no going
7603                          * back, which may be a problem...
7604                          */
7605                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7606                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7607                                                      rack->r_ctl.rc_highest_us_rtt,
7608                                                      rack->r_ctl.rc_lowest_us_rtt,
7609                                                      RACK_RTTS_SEEHBP);
7610                                 rack->rc_highly_buffered = 1;
7611                         }
7612                 }
7613         }
7614         if ((rack->r_ctl.rack_rs.confidence) ||
7615             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7616                 /*
7617                  * If we are highly confident of it <or> it was
7618                  * never retransmitted we accept it as the last us_rtt.
7619                  */
7620                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7621                 /* The lowest rtt can be set if its was not retransmited */
7622                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7623                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7624                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7625                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7626                 }
7627         }
7628         o_srtt = tp->t_srtt;
7629         o_var = tp->t_rttvar;
7630         rack = (struct tcp_rack *)tp->t_fb_ptr;
7631         if (tp->t_srtt != 0) {
7632                 /*
7633                  * We keep a simple srtt in microseconds, like our rtt
7634                  * measurement. We don't need to do any tricks with shifting
7635                  * etc. Instead we just add in 1/8th of the new measurement
7636                  * and subtract out 1/8 of the old srtt. We do the same with
7637                  * the variance after finding the absolute value of the
7638                  * difference between this sample and the current srtt.
7639                  */
7640                 delta = tp->t_srtt - rtt;
7641                 /* Take off 1/8th of the current sRTT */
7642                 tp->t_srtt -= (tp->t_srtt >> 3);
7643                 /* Add in 1/8th of the new RTT just measured */
7644                 tp->t_srtt += (rtt >> 3);
7645                 if (tp->t_srtt <= 0)
7646                         tp->t_srtt = 1;
7647                 /* Now lets make the absolute value of the variance */
7648                 if (delta < 0)
7649                         delta = -delta;
7650                 /* Subtract out 1/8th */
7651                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7652                 /* Add in 1/8th of the new variance we just saw */
7653                 tp->t_rttvar += (delta >> 3);
7654                 if (tp->t_rttvar <= 0)
7655                         tp->t_rttvar = 1;
7656                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7657                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7658         } else {
7659                 /*
7660                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7661                  * variance to half the rtt (so our first retransmit happens
7662                  * at 3*rtt).
7663                  */
7664                 tp->t_srtt = rtt;
7665                 tp->t_rttvar = rtt >> 1;
7666                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7667         }
7668         rack->rc_srtt_measure_made = 1;
7669         KMOD_TCPSTAT_INC(tcps_rttupdated);
7670         tp->t_rttupdated++;
7671 #ifdef STATS
7672         if (rack_stats_gets_ms_rtt == 0) {
7673                 /* Send in the microsecond rtt used for rxt timeout purposes */
7674                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7675         } else if (rack_stats_gets_ms_rtt == 1) {
7676                 /* Send in the millisecond rtt used for rxt timeout purposes */
7677                 int32_t ms_rtt;
7678
7679                 /* Round up */
7680                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7681                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7682         } else if (rack_stats_gets_ms_rtt == 2) {
7683                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7684                 int32_t ms_rtt;
7685
7686                 /* Round up */
7687                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7688                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7689         }  else {
7690                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7691                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7692         }
7693
7694 #endif
7695         /*
7696          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7697          * way we do the smoothing, srtt and rttvar will each average +1/2
7698          * tick of bias.  When we compute the retransmit timer, we want 1/2
7699          * tick of rounding and 1 extra tick because of +-1/2 tick
7700          * uncertainty in the firing of the timer.  The bias will give us
7701          * exactly the 1.5 tick we need.  But, because the bias is
7702          * statistical, we have to test that we don't drop below the minimum
7703          * feasible timer (which is 2 ticks).
7704          */
7705         tp->t_rxtshift = 0;
7706         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7707                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7708         rack_log_rtt_sample(rack, rtt);
7709         tp->t_softerror = 0;
7710 }
7711
7712
7713 static void
7714 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7715 {
7716         /*
7717          * Apply to filter the inbound us-rtt at us_cts.
7718          */
7719         uint32_t old_rtt;
7720
7721         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7722         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7723                                us_rtt, us_cts);
7724         if (rack->r_ctl.last_pacing_time &&
7725             rack->rc_gp_dyn_mul &&
7726             (rack->r_ctl.last_pacing_time > us_rtt))
7727                 rack->pacing_longer_than_rtt = 1;
7728         else
7729                 rack->pacing_longer_than_rtt = 0;
7730         if (old_rtt > us_rtt) {
7731                 /* We just hit a new lower rtt time */
7732                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7733                                      __LINE__, RACK_RTTS_NEWRTT);
7734                 /*
7735                  * Only count it if its lower than what we saw within our
7736                  * calculated range.
7737                  */
7738                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7739                         if (rack_probertt_lower_within &&
7740                             rack->rc_gp_dyn_mul &&
7741                             (rack->use_fixed_rate == 0) &&
7742                             (rack->rc_always_pace)) {
7743                                 /*
7744                                  * We are seeing a new lower rtt very close
7745                                  * to the time that we would have entered probe-rtt.
7746                                  * This is probably due to the fact that a peer flow
7747                                  * has entered probe-rtt. Lets go in now too.
7748                                  */
7749                                 uint32_t val;
7750
7751                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7752                                 val /= 100;
7753                                 if ((rack->in_probe_rtt == 0)  &&
7754                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7755                                         rack_enter_probertt(rack, us_cts);
7756                                 }
7757                         }
7758                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7759                 }
7760         }
7761 }
7762
7763 static int
7764 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7765     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7766 {
7767         int32_t i, all;
7768         uint32_t t, len_acked;
7769
7770         if ((rsm->r_flags & RACK_ACKED) ||
7771             (rsm->r_flags & RACK_WAS_ACKED))
7772                 /* Already done */
7773                 return (0);
7774         if (rsm->r_no_rtt_allowed) {
7775                 /* Not allowed */
7776                 return (0);
7777         }
7778         if (ack_type == CUM_ACKED) {
7779                 if (SEQ_GT(th_ack, rsm->r_end)) {
7780                         len_acked = rsm->r_end - rsm->r_start;
7781                         all = 1;
7782                 } else {
7783                         len_acked = th_ack - rsm->r_start;
7784                         all = 0;
7785                 }
7786         } else {
7787                 len_acked = rsm->r_end - rsm->r_start;
7788                 all = 0;
7789         }
7790         if (rsm->r_rtr_cnt == 1) {
7791                 uint32_t us_rtt;
7792
7793                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7794                 if ((int)t <= 0)
7795                         t = 1;
7796                 if (!tp->t_rttlow || tp->t_rttlow > t)
7797                         tp->t_rttlow = t;
7798                 if (!rack->r_ctl.rc_rack_min_rtt ||
7799                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7800                         rack->r_ctl.rc_rack_min_rtt = t;
7801                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7802                                 rack->r_ctl.rc_rack_min_rtt = 1;
7803                         }
7804                 }
7805                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7806                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7807                 else
7808                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7809                 if (us_rtt == 0)
7810                         us_rtt = 1;
7811                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7812                 if (ack_type == SACKED) {
7813                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7814                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7815                 } else {
7816                         /*
7817                          * We need to setup what our confidence
7818                          * is in this ack.
7819                          *
7820                          * If the rsm was app limited and it is
7821                          * less than a mss in length (the end
7822                          * of the send) then we have a gap. If we
7823                          * were app limited but say we were sending
7824                          * multiple MSS's then we are more confident
7825                          * int it.
7826                          *
7827                          * When we are not app-limited then we see if
7828                          * the rsm is being included in the current
7829                          * measurement, we tell this by the app_limited_needs_set
7830                          * flag.
7831                          *
7832                          * Note that being cwnd blocked is not applimited
7833                          * as well as the pacing delay between packets which
7834                          * are sending only 1 or 2 MSS's also will show up
7835                          * in the RTT. We probably need to examine this algorithm
7836                          * a bit more and enhance it to account for the delay
7837                          * between rsm's. We could do that by saving off the
7838                          * pacing delay of each rsm (in an rsm) and then
7839                          * factoring that in somehow though for now I am
7840                          * not sure how :)
7841                          */
7842                         int calc_conf = 0;
7843
7844                         if (rsm->r_flags & RACK_APP_LIMITED) {
7845                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7846                                         calc_conf = 0;
7847                                 else
7848                                         calc_conf = 1;
7849                         } else if (rack->app_limited_needs_set == 0) {
7850                                 calc_conf = 1;
7851                         } else {
7852                                 calc_conf = 0;
7853                         }
7854                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7855                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7856                                             calc_conf, rsm, rsm->r_rtr_cnt);
7857                 }
7858                 if ((rsm->r_flags & RACK_TLP) &&
7859                     (!IN_FASTRECOVERY(tp->t_flags))) {
7860                         /* Segment was a TLP and our retrans matched */
7861                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7862                                 rack->r_ctl.rc_rsm_start = tp->snd_max;
7863                                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
7864                                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
7865                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
7866                         }
7867                 }
7868                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7869                         /* New more recent rack_tmit_time */
7870                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7871                         rack->rc_rack_rtt = t;
7872                 }
7873                 return (1);
7874         }
7875         /*
7876          * We clear the soft/rxtshift since we got an ack.
7877          * There is no assurance we will call the commit() function
7878          * so we need to clear these to avoid incorrect handling.
7879          */
7880         tp->t_rxtshift = 0;
7881         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7882                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7883         tp->t_softerror = 0;
7884         if (to && (to->to_flags & TOF_TS) &&
7885             (ack_type == CUM_ACKED) &&
7886             (to->to_tsecr) &&
7887             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7888                 /*
7889                  * Now which timestamp does it match? In this block the ACK
7890                  * must be coming from a previous transmission.
7891                  */
7892                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7893                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7894                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7895                                 if ((int)t <= 0)
7896                                         t = 1;
7897                                 if ((i + 1) < rsm->r_rtr_cnt) {
7898                                         /*
7899                                          * The peer ack'd from our previous
7900                                          * transmission. We have a spurious
7901                                          * retransmission and thus we dont
7902                                          * want to update our rack_rtt.
7903                                          */
7904                                         return (0);
7905                                 }
7906                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7907                                         tp->t_rttlow = t;
7908                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7909                                         rack->r_ctl.rc_rack_min_rtt = t;
7910                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7911                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7912                                         }
7913                                 }
7914                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7915                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7916                                         /* New more recent rack_tmit_time */
7917                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7918                                         rack->rc_rack_rtt = t;
7919                                 }
7920                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7921                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7922                                                     rsm->r_rtr_cnt);
7923                                 return (1);
7924                         }
7925                 }
7926                 goto ts_not_found;
7927         } else {
7928                 /*
7929                  * Ok its a SACK block that we retransmitted. or a windows
7930                  * machine without timestamps. We can tell nothing from the
7931                  * time-stamp since its not there or the time the peer last
7932                  * recieved a segment that moved forward its cum-ack point.
7933                  */
7934 ts_not_found:
7935                 i = rsm->r_rtr_cnt - 1;
7936                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7937                 if ((int)t <= 0)
7938                         t = 1;
7939                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7940                         /*
7941                          * We retransmitted and the ack came back in less
7942                          * than the smallest rtt we have observed. We most
7943                          * likely did an improper retransmit as outlined in
7944                          * 6.2 Step 2 point 2 in the rack-draft so we
7945                          * don't want to update our rack_rtt. We in
7946                          * theory (in future) might want to think about reverting our
7947                          * cwnd state but we won't for now.
7948                          */
7949                         return (0);
7950                 } else if (rack->r_ctl.rc_rack_min_rtt) {
7951                         /*
7952                          * We retransmitted it and the retransmit did the
7953                          * job.
7954                          */
7955                         if (!rack->r_ctl.rc_rack_min_rtt ||
7956                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7957                                 rack->r_ctl.rc_rack_min_rtt = t;
7958                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7959                                         rack->r_ctl.rc_rack_min_rtt = 1;
7960                                 }
7961                         }
7962                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7963                                 /* New more recent rack_tmit_time */
7964                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7965                                 rack->rc_rack_rtt = t;
7966                         }
7967                         return (1);
7968                 }
7969         }
7970         return (0);
7971 }
7972
7973 /*
7974  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7975  */
7976 static void
7977 rack_log_sack_passed(struct tcpcb *tp,
7978     struct tcp_rack *rack, struct rack_sendmap *rsm)
7979 {
7980         struct rack_sendmap *nrsm;
7981
7982         nrsm = rsm;
7983         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7984             rack_head, r_tnext) {
7985                 if (nrsm == rsm) {
7986                         /* Skip orginal segment he is acked */
7987                         continue;
7988                 }
7989                 if (nrsm->r_flags & RACK_ACKED) {
7990                         /*
7991                          * Skip ack'd segments, though we
7992                          * should not see these, since tmap
7993                          * should not have ack'd segments.
7994                          */
7995                         continue;
7996                 }
7997                 if (nrsm->r_flags & RACK_SACK_PASSED) {
7998                         /*
7999                          * We found one that is already marked
8000                          * passed, we have been here before and
8001                          * so all others below this are marked.
8002                          */
8003                         break;
8004                 }
8005                 nrsm->r_flags |= RACK_SACK_PASSED;
8006                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8007         }
8008 }
8009
8010 static void
8011 rack_need_set_test(struct tcpcb *tp,
8012                    struct tcp_rack *rack,
8013                    struct rack_sendmap *rsm,
8014                    tcp_seq th_ack,
8015                    int line,
8016                    int use_which)
8017 {
8018
8019         if ((tp->t_flags & TF_GPUTINPROG) &&
8020             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8021                 /*
8022                  * We were app limited, and this ack
8023                  * butts up or goes beyond the point where we want
8024                  * to start our next measurement. We need
8025                  * to record the new gput_ts as here and
8026                  * possibly update the start sequence.
8027                  */
8028                 uint32_t seq, ts;
8029
8030                 if (rsm->r_rtr_cnt > 1) {
8031                         /*
8032                          * This is a retransmit, can we
8033                          * really make any assessment at this
8034                          * point?  We are not really sure of
8035                          * the timestamp, is it this or the
8036                          * previous transmission?
8037                          *
8038                          * Lets wait for something better that
8039                          * is not retransmitted.
8040                          */
8041                         return;
8042                 }
8043                 seq = tp->gput_seq;
8044                 ts = tp->gput_ts;
8045                 rack->app_limited_needs_set = 0;
8046                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8047                 /* Do we start at a new end? */
8048                 if ((use_which == RACK_USE_BEG) &&
8049                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8050                         /*
8051                          * When we get an ACK that just eats
8052                          * up some of the rsm, we set RACK_USE_BEG
8053                          * since whats at r_start (i.e. th_ack)
8054                          * is left unacked and thats where the
8055                          * measurement not starts.
8056                          */
8057                         tp->gput_seq = rsm->r_start;
8058                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8059                 }
8060                 if ((use_which == RACK_USE_END) &&
8061                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8062                             /*
8063                              * We use the end when the cumack
8064                              * is moving forward and completely
8065                              * deleting the rsm passed so basically
8066                              * r_end holds th_ack.
8067                              *
8068                              * For SACK's we also want to use the end
8069                              * since this piece just got sacked and
8070                              * we want to target anything after that
8071                              * in our measurement.
8072                              */
8073                             tp->gput_seq = rsm->r_end;
8074                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8075                 }
8076                 if (use_which == RACK_USE_END_OR_THACK) {
8077                         /*
8078                          * special case for ack moving forward,
8079                          * not a sack, we need to move all the
8080                          * way up to where this ack cum-ack moves
8081                          * to.
8082                          */
8083                         if (SEQ_GT(th_ack, rsm->r_end))
8084                                 tp->gput_seq = th_ack;
8085                         else
8086                                 tp->gput_seq = rsm->r_end;
8087                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8088                 }
8089                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8090                         /*
8091                          * We moved beyond this guy's range, re-calculate
8092                          * the new end point.
8093                          */
8094                         if (rack->rc_gp_filled == 0) {
8095                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8096                         } else {
8097                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8098                         }
8099                 }
8100                 /*
8101                  * We are moving the goal post, we may be able to clear the
8102                  * measure_saw_probe_rtt flag.
8103                  */
8104                 if ((rack->in_probe_rtt == 0) &&
8105                     (rack->measure_saw_probe_rtt) &&
8106                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8107                         rack->measure_saw_probe_rtt = 0;
8108                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8109                                            seq, tp->gput_seq, 0, 5, line, NULL);
8110                 if (rack->rc_gp_filled &&
8111                     ((tp->gput_ack - tp->gput_seq) <
8112                      max(rc_init_window(rack), (MIN_GP_WIN *
8113                                                 ctf_fixed_maxseg(tp))))) {
8114                         uint32_t ideal_amount;
8115
8116                         ideal_amount = rack_get_measure_window(tp, rack);
8117                         if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8118                                 /*
8119                                  * There is no sense of continuing this measurement
8120                                  * because its too small to gain us anything we
8121                                  * trust. Skip it and that way we can start a new
8122                                  * measurement quicker.
8123                                  */
8124                                 tp->t_flags &= ~TF_GPUTINPROG;
8125                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8126                                                            0, 0, 0, 6, __LINE__, NULL);
8127                         } else {
8128                                 /*
8129                                  * Reset the window further out.
8130                                  */
8131                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8132                         }
8133                 }
8134         }
8135 }
8136
8137 static uint32_t
8138 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8139                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8140 {
8141         uint32_t start, end, changed = 0;
8142         struct rack_sendmap stack_map;
8143         struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
8144         int32_t used_ref = 1;
8145         int moved = 0;
8146
8147         start = sack->start;
8148         end = sack->end;
8149         rsm = *prsm;
8150         memset(&fe, 0, sizeof(fe));
8151 do_rest_ofb:
8152         if ((rsm == NULL) ||
8153             (SEQ_LT(end, rsm->r_start)) ||
8154             (SEQ_GEQ(start, rsm->r_end)) ||
8155             (SEQ_LT(start, rsm->r_start))) {
8156                 /*
8157                  * We are not in the right spot,
8158                  * find the correct spot in the tree.
8159                  */
8160                 used_ref = 0;
8161                 fe.r_start = start;
8162                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8163                 moved++;
8164         }
8165         if (rsm == NULL) {
8166                 /* TSNH */
8167                 goto out;
8168         }
8169         /* Ok we have an ACK for some piece of this rsm */
8170         if (rsm->r_start != start) {
8171                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8172                         /**
8173                          * Need to split this in two pieces the before and after,
8174                          * the before remains in the map, the after must be
8175                          * added. In other words we have:
8176                          * rsm        |--------------|
8177                          * sackblk        |------->
8178                          * rsm will become
8179                          *     rsm    |---|
8180                          * and nrsm will be  the sacked piece
8181                          *     nrsm       |----------|
8182                          *
8183                          * But before we start down that path lets
8184                          * see if the sack spans over on top of
8185                          * the next guy and it is already sacked.
8186                          */
8187                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8188                         if (next && (next->r_flags & RACK_ACKED) &&
8189                             SEQ_GEQ(end, next->r_start)) {
8190                                 /**
8191                                  * So the next one is already acked, and
8192                                  * we can thus by hookery use our stack_map
8193                                  * to reflect the piece being sacked and
8194                                  * then adjust the two tree entries moving
8195                                  * the start and ends around. So we start like:
8196                                  *  rsm     |------------|             (not-acked)
8197                                  *  next                 |-----------| (acked)
8198                                  *  sackblk        |-------->
8199                                  *  We want to end like so:
8200                                  *  rsm     |------|                   (not-acked)
8201                                  *  next           |-----------------| (acked)
8202                                  *  nrsm           |-----|
8203                                  * Where nrsm is a temporary stack piece we
8204                                  * use to update all the gizmos.
8205                                  */
8206                                 /* Copy up our fudge block */
8207                                 nrsm = &stack_map;
8208                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8209                                 /* Now adjust our tree blocks */
8210                                 rsm->r_end = start;
8211                                 next->r_start = start;
8212                                 /* Now we must adjust back where next->m is */
8213                                 rack_setup_offset_for_rsm(rsm, next);
8214
8215                                 /* We don't need to adjust rsm, it did not change */
8216                                 /* Clear out the dup ack count of the remainder */
8217                                 rsm->r_dupack = 0;
8218                                 rsm->r_just_ret = 0;
8219                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8220                                 /* Now lets make sure our fudge block is right */
8221                                 nrsm->r_start = start;
8222                                 /* Now lets update all the stats and such */
8223                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8224                                 if (rack->app_limited_needs_set)
8225                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8226                                 changed += (nrsm->r_end - nrsm->r_start);
8227                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8228                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8229                                         counter_u64_add(rack_reorder_seen, 1);
8230                                         rack->r_ctl.rc_reorder_ts = cts;
8231                                 }
8232                                 /*
8233                                  * Now we want to go up from rsm (the
8234                                  * one left un-acked) to the next one
8235                                  * in the tmap. We do this so when
8236                                  * we walk backwards we include marking
8237                                  * sack-passed on rsm (The one passed in
8238                                  * is skipped since it is generally called
8239                                  * on something sacked before removing it
8240                                  * from the tmap).
8241                                  */
8242                                 if (rsm->r_in_tmap) {
8243                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8244                                         /*
8245                                          * Now that we have the next
8246                                          * one walk backwards from there.
8247                                          */
8248                                         if (nrsm && nrsm->r_in_tmap)
8249                                                 rack_log_sack_passed(tp, rack, nrsm);
8250                                 }
8251                                 /* Now are we done? */
8252                                 if (SEQ_LT(end, next->r_end) ||
8253                                     (end == next->r_end)) {
8254                                         /* Done with block */
8255                                         goto out;
8256                                 }
8257                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8258                                 counter_u64_add(rack_sack_used_next_merge, 1);
8259                                 /* Postion for the next block */
8260                                 start = next->r_end;
8261                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8262                                 if (rsm == NULL)
8263                                         goto out;
8264                         } else {
8265                                 /**
8266                                  * We can't use any hookery here, so we
8267                                  * need to split the map. We enter like
8268                                  * so:
8269                                  *  rsm      |--------|
8270                                  *  sackblk       |----->
8271                                  * We will add the new block nrsm and
8272                                  * that will be the new portion, and then
8273                                  * fall through after reseting rsm. So we
8274                                  * split and look like this:
8275                                  *  rsm      |----|
8276                                  *  sackblk       |----->
8277                                  *  nrsm          |---|
8278                                  * We then fall through reseting
8279                                  * rsm to nrsm, so the next block
8280                                  * picks it up.
8281                                  */
8282                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8283                                 if (nrsm == NULL) {
8284                                         /*
8285                                          * failed XXXrrs what can we do but loose the sack
8286                                          * info?
8287                                          */
8288                                         goto out;
8289                                 }
8290                                 counter_u64_add(rack_sack_splits, 1);
8291                                 rack_clone_rsm(rack, nrsm, rsm, start);
8292                                 rsm->r_just_ret = 0;
8293                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8294 #ifdef INVARIANTS
8295                                 if (insret != NULL) {
8296                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8297                                               nrsm, insret, rack, rsm);
8298                                 }
8299 #endif
8300                                 if (rsm->r_in_tmap) {
8301                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8302                                         nrsm->r_in_tmap = 1;
8303                                 }
8304                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8305                                 rsm->r_flags &= (~RACK_HAS_FIN);
8306                                 /* Position us to point to the new nrsm that starts the sack blk */
8307                                 rsm = nrsm;
8308                         }
8309                 } else {
8310                         /* Already sacked this piece */
8311                         counter_u64_add(rack_sack_skipped_acked, 1);
8312                         moved++;
8313                         if (end == rsm->r_end) {
8314                                 /* Done with block */
8315                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8316                                 goto out;
8317                         } else if (SEQ_LT(end, rsm->r_end)) {
8318                                 /* A partial sack to a already sacked block */
8319                                 moved++;
8320                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8321                                 goto out;
8322                         } else {
8323                                 /*
8324                                  * The end goes beyond this guy
8325                                  * repostion the start to the
8326                                  * next block.
8327                                  */
8328                                 start = rsm->r_end;
8329                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8330                                 if (rsm == NULL)
8331                                         goto out;
8332                         }
8333                 }
8334         }
8335         if (SEQ_GEQ(end, rsm->r_end)) {
8336                 /**
8337                  * The end of this block is either beyond this guy or right
8338                  * at this guy. I.e.:
8339                  *  rsm ---                 |-----|
8340                  *  end                     |-----|
8341                  *  <or>
8342                  *  end                     |---------|
8343                  */
8344                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8345                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8346                         changed += (rsm->r_end - rsm->r_start);
8347                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8348                         if (rsm->r_in_tmap) /* should be true */
8349                                 rack_log_sack_passed(tp, rack, rsm);
8350                         /* Is Reordering occuring? */
8351                         if (rsm->r_flags & RACK_SACK_PASSED) {
8352                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8353                                 counter_u64_add(rack_reorder_seen, 1);
8354                                 rack->r_ctl.rc_reorder_ts = cts;
8355                         }
8356                         if (rack->app_limited_needs_set)
8357                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8358                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8359                         rsm->r_flags |= RACK_ACKED;
8360                         rsm->r_flags &= ~RACK_TLP;
8361                         if (rsm->r_in_tmap) {
8362                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8363                                 rsm->r_in_tmap = 0;
8364                         }
8365                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8366                 } else {
8367                         counter_u64_add(rack_sack_skipped_acked, 1);
8368                         moved++;
8369                 }
8370                 if (end == rsm->r_end) {
8371                         /* This block only - done, setup for next */
8372                         goto out;
8373                 }
8374                 /*
8375                  * There is more not coverend by this rsm move on
8376                  * to the next block in the RB tree.
8377                  */
8378                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8379                 start = rsm->r_end;
8380                 rsm = nrsm;
8381                 if (rsm == NULL)
8382                         goto out;
8383                 goto do_rest_ofb;
8384         }
8385         /**
8386          * The end of this sack block is smaller than
8387          * our rsm i.e.:
8388          *  rsm ---                 |-----|
8389          *  end                     |--|
8390          */
8391         if ((rsm->r_flags & RACK_ACKED) == 0) {
8392                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8393                 if (prev && (prev->r_flags & RACK_ACKED)) {
8394                         /**
8395                          * Goal, we want the right remainder of rsm to shrink
8396                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8397                          * We want to expand prev to go all the way
8398                          * to prev->r_end <- end.
8399                          * so in the tree we have before:
8400                          *   prev     |--------|         (acked)
8401                          *   rsm               |-------| (non-acked)
8402                          *   sackblk           |-|
8403                          * We churn it so we end up with
8404                          *   prev     |----------|       (acked)
8405                          *   rsm                 |-----| (non-acked)
8406                          *   nrsm              |-| (temporary)
8407                          */
8408                         nrsm = &stack_map;
8409                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8410                         prev->r_end = end;
8411                         rsm->r_start = end;
8412                         /* Now adjust nrsm (stack copy) to be
8413                          * the one that is the small
8414                          * piece that was "sacked".
8415                          */
8416                         nrsm->r_end = end;
8417                         rsm->r_dupack = 0;
8418                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8419                         /*
8420                          * Now that the rsm has had its start moved forward
8421                          * lets go ahead and get its new place in the world.
8422                          */
8423                         rack_setup_offset_for_rsm(prev, rsm);
8424                         /*
8425                          * Now nrsm is our new little piece
8426                          * that is acked (which was merged
8427                          * to prev). Update the rtt and changed
8428                          * based on that. Also check for reordering.
8429                          */
8430                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8431                         if (rack->app_limited_needs_set)
8432                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8433                         changed += (nrsm->r_end - nrsm->r_start);
8434                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8435                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8436                                 counter_u64_add(rack_reorder_seen, 1);
8437                                 rack->r_ctl.rc_reorder_ts = cts;
8438                         }
8439                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8440                         rsm = prev;
8441                         counter_u64_add(rack_sack_used_prev_merge, 1);
8442                 } else {
8443                         /**
8444                          * This is the case where our previous
8445                          * block is not acked either, so we must
8446                          * split the block in two.
8447                          */
8448                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8449                         if (nrsm == NULL) {
8450                                 /* failed rrs what can we do but loose the sack info? */
8451                                 goto out;
8452                         }
8453                         /**
8454                          * In this case nrsm becomes
8455                          * nrsm->r_start = end;
8456                          * nrsm->r_end = rsm->r_end;
8457                          * which is un-acked.
8458                          * <and>
8459                          * rsm->r_end = nrsm->r_start;
8460                          * i.e. the remaining un-acked
8461                          * piece is left on the left
8462                          * hand side.
8463                          *
8464                          * So we start like this
8465                          * rsm      |----------| (not acked)
8466                          * sackblk  |---|
8467                          * build it so we have
8468                          * rsm      |---|         (acked)
8469                          * nrsm         |------|  (not acked)
8470                          */
8471                         counter_u64_add(rack_sack_splits, 1);
8472                         rack_clone_rsm(rack, nrsm, rsm, end);
8473                         rsm->r_flags &= (~RACK_HAS_FIN);
8474                         rsm->r_just_ret = 0;
8475                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8476 #ifdef INVARIANTS
8477                         if (insret != NULL) {
8478                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8479                                       nrsm, insret, rack, rsm);
8480                         }
8481 #endif
8482                         if (rsm->r_in_tmap) {
8483                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8484                                 nrsm->r_in_tmap = 1;
8485                         }
8486                         nrsm->r_dupack = 0;
8487                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8488                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8489                         changed += (rsm->r_end - rsm->r_start);
8490                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8491                         if (rsm->r_in_tmap) /* should be true */
8492                                 rack_log_sack_passed(tp, rack, rsm);
8493                         /* Is Reordering occuring? */
8494                         if (rsm->r_flags & RACK_SACK_PASSED) {
8495                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8496                                 counter_u64_add(rack_reorder_seen, 1);
8497                                 rack->r_ctl.rc_reorder_ts = cts;
8498                         }
8499                         if (rack->app_limited_needs_set)
8500                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8501                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8502                         rsm->r_flags |= RACK_ACKED;
8503                         rsm->r_flags &= ~RACK_TLP;
8504                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8505                         if (rsm->r_in_tmap) {
8506                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8507                                 rsm->r_in_tmap = 0;
8508                         }
8509                 }
8510         } else if (start != end){
8511                 /*
8512                  * The block was already acked.
8513                  */
8514                 counter_u64_add(rack_sack_skipped_acked, 1);
8515                 moved++;
8516         }
8517 out:
8518         if (rsm && (rsm->r_flags & RACK_ACKED)) {
8519                 /*
8520                  * Now can we merge where we worked
8521                  * with either the previous or
8522                  * next block?
8523                  */
8524                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8525                 while (next) {
8526                     if (next->r_flags & RACK_ACKED) {
8527                         /* yep this and next can be merged */
8528                         rsm = rack_merge_rsm(rack, rsm, next);
8529                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8530                     } else
8531                             break;
8532                 }
8533                 /* Now what about the previous? */
8534                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8535                 while (prev) {
8536                     if (prev->r_flags & RACK_ACKED) {
8537                         /* yep the previous and this can be merged */
8538                         rsm = rack_merge_rsm(rack, prev, rsm);
8539                         prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8540                     } else
8541                             break;
8542                 }
8543         }
8544         if (used_ref == 0) {
8545                 counter_u64_add(rack_sack_proc_all, 1);
8546         } else {
8547                 counter_u64_add(rack_sack_proc_short, 1);
8548         }
8549         /* Save off the next one for quick reference. */
8550         if (rsm)
8551                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8552         else
8553                 nrsm = NULL;
8554         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8555         /* Pass back the moved. */
8556         *moved_two = moved;
8557         return (changed);
8558 }
8559
8560 static void inline
8561 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8562 {
8563         struct rack_sendmap *tmap;
8564
8565         tmap = NULL;
8566         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8567                 /* Its no longer sacked, mark it so */
8568                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8569 #ifdef INVARIANTS
8570                 if (rsm->r_in_tmap) {
8571                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8572                               rack, rsm, rsm->r_flags);
8573                 }
8574 #endif
8575                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8576                 /* Rebuild it into our tmap */
8577                 if (tmap == NULL) {
8578                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8579                         tmap = rsm;
8580                 } else {
8581                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8582                         tmap = rsm;
8583                 }
8584                 tmap->r_in_tmap = 1;
8585                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8586         }
8587         /*
8588          * Now lets possibly clear the sack filter so we start
8589          * recognizing sacks that cover this area.
8590          */
8591         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8592
8593 }
8594
8595 static void
8596 rack_do_decay(struct tcp_rack *rack)
8597 {
8598         struct timeval res;
8599
8600 #define timersub(tvp, uvp, vvp)                                         \
8601         do {                                                            \
8602                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8603                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8604                 if ((vvp)->tv_usec < 0) {                               \
8605                         (vvp)->tv_sec--;                                \
8606                         (vvp)->tv_usec += 1000000;                      \
8607                 }                                                       \
8608         } while (0)
8609
8610         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8611 #undef timersub
8612
8613         rack->r_ctl.input_pkt++;
8614         if ((rack->rc_in_persist) ||
8615             (res.tv_sec >= 1) ||
8616             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8617                 /*
8618                  * Check for decay of non-SAD,
8619                  * we want all SAD detection metrics to
8620                  * decay 1/4 per second (or more) passed.
8621                  */
8622                 uint32_t pkt_delta;
8623
8624                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8625                 /* Update our saved tracking values */
8626                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8627                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8628                 /* Now do we escape without decay? */
8629 #ifdef NETFLIX_EXP_DETECTION
8630                 if (rack->rc_in_persist ||
8631                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8632                     (pkt_delta < tcp_sad_low_pps)){
8633                         /*
8634                          * We don't decay idle connections
8635                          * or ones that have a low input pps.
8636                          */
8637                         return;
8638                 }
8639                 /* Decay the counters */
8640                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8641                                                         tcp_sad_decay_val);
8642                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8643                                                          tcp_sad_decay_val);
8644                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8645                                                                tcp_sad_decay_val);
8646                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8647                                                                 tcp_sad_decay_val);
8648 #endif
8649         }
8650 }
8651
8652 static void
8653 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8654 {
8655         struct rack_sendmap *rsm, *rm;
8656
8657         /*
8658          * The ACK point is advancing to th_ack, we must drop off
8659          * the packets in the rack log and calculate any eligble
8660          * RTT's.
8661          */
8662         rack->r_wanted_output = 1;
8663 more:
8664         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8665         if (rsm == NULL) {
8666                 if ((th_ack - 1) == tp->iss) {
8667                         /*
8668                          * For the SYN incoming case we will not
8669                          * have called tcp_output for the sending of
8670                          * the SYN, so there will be no map. All
8671                          * other cases should probably be a panic.
8672                          */
8673                         return;
8674                 }
8675                 if (tp->t_flags & TF_SENTFIN) {
8676                         /* if we sent a FIN we often will not have map */
8677                         return;
8678                 }
8679 #ifdef INVARIANTS
8680                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8681                       tp,
8682                       tp->t_state, th_ack, rack,
8683                       tp->snd_una, tp->snd_max, tp->snd_nxt);
8684 #endif
8685                 return;
8686         }
8687         if (SEQ_LT(th_ack, rsm->r_start)) {
8688                 /* Huh map is missing this */
8689 #ifdef INVARIANTS
8690                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8691                        rsm->r_start,
8692                        th_ack, tp->t_state, rack->r_state);
8693 #endif
8694                 return;
8695         }
8696         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8697         /* Now do we consume the whole thing? */
8698         if (SEQ_GEQ(th_ack, rsm->r_end)) {
8699                 /* Its all consumed. */
8700                 uint32_t left;
8701                 uint8_t newly_acked;
8702
8703                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8704                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8705                 rsm->r_rtr_bytes = 0;
8706                 /* Record the time of highest cumack sent */
8707                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8708                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8709 #ifdef INVARIANTS
8710                 if (rm != rsm) {
8711                         panic("removing head in rack:%p rsm:%p rm:%p",
8712                               rack, rsm, rm);
8713                 }
8714 #endif
8715                 if (rsm->r_in_tmap) {
8716                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8717                         rsm->r_in_tmap = 0;
8718                 }
8719                 newly_acked = 1;
8720                 if (rsm->r_flags & RACK_ACKED) {
8721                         /*
8722                          * It was acked on the scoreboard -- remove
8723                          * it from total
8724                          */
8725                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8726                         newly_acked = 0;
8727                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
8728                         /*
8729                          * There are segments ACKED on the
8730                          * scoreboard further up. We are seeing
8731                          * reordering.
8732                          */
8733                         rsm->r_flags &= ~RACK_SACK_PASSED;
8734                         counter_u64_add(rack_reorder_seen, 1);
8735                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8736                         rsm->r_flags |= RACK_ACKED;
8737                         rack->r_ctl.rc_reorder_ts = cts;
8738                         if (rack->r_ent_rec_ns) {
8739                                 /*
8740                                  * We have sent no more, and we saw an sack
8741                                  * then ack arrive.
8742                                  */
8743                                 rack->r_might_revert = 1;
8744                         }
8745                 }
8746                 if ((rsm->r_flags & RACK_TO_REXT) &&
8747                     (tp->t_flags & TF_RCVD_TSTMP) &&
8748                     (to->to_flags & TOF_TS) &&
8749                     (tp->t_flags & TF_PREVVALID)) {
8750                         /*
8751                          * We can use the timestamp to see
8752                          * if this retransmission was from the
8753                          * first transmit. If so we made a mistake.
8754                          */
8755                         tp->t_flags &= ~TF_PREVVALID;
8756                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
8757                                 /* The first transmit is what this ack is for */
8758                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack);
8759                         }
8760                 }
8761                 left = th_ack - rsm->r_end;
8762                 if (rack->app_limited_needs_set && newly_acked)
8763                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
8764                 /* Free back to zone */
8765                 rack_free(rack, rsm);
8766                 if (left) {
8767                         goto more;
8768                 }
8769                 /* Check for reneging */
8770                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8771                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
8772                         /*
8773                          * The peer has moved snd_una up to
8774                          * the edge of this send, i.e. one
8775                          * that it had previously acked. The only
8776                          * way that can be true if the peer threw
8777                          * away data (space issues) that it had
8778                          * previously sacked (else it would have
8779                          * given us snd_una up to (rsm->r_end).
8780                          * We need to undo the acked markings here.
8781                          *
8782                          * Note we have to look to make sure th_ack is
8783                          * our rsm->r_start in case we get an old ack
8784                          * where th_ack is behind snd_una.
8785                          */
8786                         rack_peer_reneges(rack, rsm, th_ack);
8787                 }
8788                 return;
8789         }
8790         if (rsm->r_flags & RACK_ACKED) {
8791                 /*
8792                  * It was acked on the scoreboard -- remove it from
8793                  * total for the part being cum-acked.
8794                  */
8795                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
8796         }
8797         /*
8798          * Clear the dup ack count for
8799          * the piece that remains.
8800          */
8801         rsm->r_dupack = 0;
8802         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8803         if (rsm->r_rtr_bytes) {
8804                 /*
8805                  * It was retransmitted adjust the
8806                  * sack holes for what was acked.
8807                  */
8808                 int ack_am;
8809
8810                 ack_am = (th_ack - rsm->r_start);
8811                 if (ack_am >= rsm->r_rtr_bytes) {
8812                         rack->r_ctl.rc_holes_rxt -= ack_am;
8813                         rsm->r_rtr_bytes -= ack_am;
8814                 }
8815         }
8816         /*
8817          * Update where the piece starts and record
8818          * the time of send of highest cumack sent.
8819          */
8820         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8821         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
8822         /* Now we need to move our offset forward too */
8823         if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
8824                 /* Fix up the orig_m_len and possibly the mbuf offset */
8825                 rack_adjust_orig_mlen(rsm);
8826         }
8827         rsm->soff += (th_ack - rsm->r_start);
8828         rsm->r_start = th_ack;
8829         /* Now do we need to move the mbuf fwd too? */
8830         if (rsm->m) {
8831                 while (rsm->soff >= rsm->m->m_len) {
8832                         rsm->soff -= rsm->m->m_len;
8833                         rsm->m = rsm->m->m_next;
8834                         KASSERT((rsm->m != NULL),
8835                                 (" nrsm:%p hit at soff:%u null m",
8836                                  rsm, rsm->soff));
8837                 }
8838                 rsm->orig_m_len = rsm->m->m_len;
8839         }
8840         if (rack->app_limited_needs_set)
8841                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
8842 }
8843
8844 static void
8845 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
8846 {
8847         struct rack_sendmap *rsm;
8848         int sack_pass_fnd = 0;
8849
8850         if (rack->r_might_revert) {
8851                 /*
8852                  * Ok we have reordering, have not sent anything, we
8853                  * might want to revert the congestion state if nothing
8854                  * further has SACK_PASSED on it. Lets check.
8855                  *
8856                  * We also get here when we have DSACKs come in for
8857                  * all the data that we FR'd. Note that a rxt or tlp
8858                  * timer clears this from happening.
8859                  */
8860
8861                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
8862                         if (rsm->r_flags & RACK_SACK_PASSED) {
8863                                 sack_pass_fnd = 1;
8864                                 break;
8865                         }
8866                 }
8867                 if (sack_pass_fnd == 0) {
8868                         /*
8869                          * We went into recovery
8870                          * incorrectly due to reordering!
8871                          */
8872                         int orig_cwnd;
8873
8874                         rack->r_ent_rec_ns = 0;
8875                         orig_cwnd = tp->snd_cwnd;
8876                         tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec;
8877                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
8878                         tp->snd_recover = tp->snd_una;
8879                         rack_log_to_prr(rack, 14, orig_cwnd);
8880                         EXIT_RECOVERY(tp->t_flags);
8881                 }
8882                 rack->r_might_revert = 0;
8883         }
8884 }
8885
8886 #ifdef NETFLIX_EXP_DETECTION
8887 static void
8888 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
8889 {
8890         if ((rack->do_detection || tcp_force_detection) &&
8891             tcp_sack_to_ack_thresh &&
8892             tcp_sack_to_move_thresh &&
8893             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
8894                 /*
8895                  * We have thresholds set to find
8896                  * possible attackers and disable sack.
8897                  * Check them.
8898                  */
8899                 uint64_t ackratio, moveratio, movetotal;
8900
8901                 /* Log detecting */
8902                 rack_log_sad(rack, 1);
8903                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
8904                 ackratio *= (uint64_t)(1000);
8905                 if (rack->r_ctl.ack_count)
8906                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
8907                 else {
8908                         /* We really should not hit here */
8909                         ackratio = 1000;
8910                 }
8911                 if ((rack->sack_attack_disable == 0) &&
8912                     (ackratio > rack_highest_sack_thresh_seen))
8913                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
8914                 movetotal = rack->r_ctl.sack_moved_extra;
8915                 movetotal += rack->r_ctl.sack_noextra_move;
8916                 moveratio = rack->r_ctl.sack_moved_extra;
8917                 moveratio *= (uint64_t)1000;
8918                 if (movetotal)
8919                         moveratio /= movetotal;
8920                 else {
8921                         /* No moves, thats pretty good */
8922                         moveratio = 0;
8923                 }
8924                 if ((rack->sack_attack_disable == 0) &&
8925                     (moveratio > rack_highest_move_thresh_seen))
8926                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
8927                 if (rack->sack_attack_disable == 0) {
8928                         if ((ackratio > tcp_sack_to_ack_thresh) &&
8929                             (moveratio > tcp_sack_to_move_thresh)) {
8930                                 /* Disable sack processing */
8931                                 rack->sack_attack_disable = 1;
8932                                 if (rack->r_rep_attack == 0) {
8933                                         rack->r_rep_attack = 1;
8934                                         counter_u64_add(rack_sack_attacks_detected, 1);
8935                                 }
8936                                 if (tcp_attack_on_turns_on_logging) {
8937                                         /*
8938                                          * Turn on logging, used for debugging
8939                                          * false positives.
8940                                          */
8941                                         rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
8942                                 }
8943                                 /* Clamp the cwnd at flight size */
8944                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
8945                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
8946                                 rack_log_sad(rack, 2);
8947                         }
8948                 } else {
8949                         /* We are sack-disabled check for false positives */
8950                         if ((ackratio <= tcp_restoral_thresh) ||
8951                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
8952                                 rack->sack_attack_disable = 0;
8953                                 rack_log_sad(rack, 3);
8954                                 /* Restart counting */
8955                                 rack->r_ctl.sack_count = 0;
8956                                 rack->r_ctl.sack_moved_extra = 0;
8957                                 rack->r_ctl.sack_noextra_move = 1;
8958                                 rack->r_ctl.ack_count = max(1,
8959                                       (bytes_this_ack / segsiz));
8960
8961                                 if (rack->r_rep_reverse == 0) {
8962                                         rack->r_rep_reverse = 1;
8963                                         counter_u64_add(rack_sack_attacks_reversed, 1);
8964                                 }
8965                                 /* Restore the cwnd */
8966                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
8967                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
8968                         }
8969                 }
8970         }
8971 }
8972 #endif
8973
8974 static void
8975 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
8976 {
8977
8978         uint32_t am;
8979
8980         if (SEQ_GT(end, start))
8981                 am = end - start;
8982         else
8983                 am = 0;
8984         /*
8985          * We keep track of how many DSACK blocks we get
8986          * after a recovery incident.
8987          */
8988         rack->r_ctl.dsack_byte_cnt += am;
8989         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
8990             rack->r_ctl.retran_during_recovery &&
8991             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
8992                 /*
8993                  * False recovery most likely culprit is reordering. If
8994                  * nothing else is missing we need to revert.
8995                  */
8996                 rack->r_might_revert = 1;
8997                 rack_handle_might_revert(rack->rc_tp, rack);
8998                 rack->r_might_revert = 0;
8999                 rack->r_ctl.retran_during_recovery = 0;
9000                 rack->r_ctl.dsack_byte_cnt = 0;
9001         }
9002 }
9003
9004 static void
9005 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9006 {
9007         /* Deal with changed and PRR here (in recovery only) */
9008         uint32_t pipe, snd_una;
9009
9010         rack->r_ctl.rc_prr_delivered += changed;
9011
9012         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9013                 /*
9014                  * It is all outstanding, we are application limited
9015                  * and thus we don't need more room to send anything.
9016                  * Note we use tp->snd_una here and not th_ack because
9017                  * the data as yet not been cut from the sb.
9018                  */
9019                 rack->r_ctl.rc_prr_sndcnt = 0;
9020                 return;
9021         }
9022         /* Compute prr_sndcnt */
9023         if (SEQ_GT(tp->snd_una, th_ack)) {
9024                 snd_una = tp->snd_una;
9025         } else {
9026                 snd_una = th_ack;
9027         }
9028         pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9029         if (pipe > tp->snd_ssthresh) {
9030                 long sndcnt;
9031
9032                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9033                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9034                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9035                 else {
9036                         rack->r_ctl.rc_prr_sndcnt = 0;
9037                         rack_log_to_prr(rack, 9, 0);
9038                         sndcnt = 0;
9039                 }
9040                 sndcnt++;
9041                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9042                         sndcnt -= rack->r_ctl.rc_prr_out;
9043                 else
9044                         sndcnt = 0;
9045                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9046                 rack_log_to_prr(rack, 10, 0);
9047         } else {
9048                 uint32_t limit;
9049
9050                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9051                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9052                 else
9053                         limit = 0;
9054                 if (changed > limit)
9055                         limit = changed;
9056                 limit += ctf_fixed_maxseg(tp);
9057                 if (tp->snd_ssthresh > pipe) {
9058                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9059                         rack_log_to_prr(rack, 11, 0);
9060                 } else {
9061                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9062                         rack_log_to_prr(rack, 12, 0);
9063                 }
9064         }
9065 }
9066
9067 static void
9068 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9069 {
9070         uint32_t changed;
9071         struct tcp_rack *rack;
9072         struct rack_sendmap *rsm;
9073         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9074         register uint32_t th_ack;
9075         int32_t i, j, k, num_sack_blks = 0;
9076         uint32_t cts, acked, ack_point, sack_changed = 0;
9077         int loop_start = 0, moved_two = 0;
9078         uint32_t tsused;
9079
9080
9081         INP_WLOCK_ASSERT(tp->t_inpcb);
9082         if (th->th_flags & TH_RST) {
9083                 /* We don't log resets */
9084                 return;
9085         }
9086         rack = (struct tcp_rack *)tp->t_fb_ptr;
9087         cts = tcp_get_usecs(NULL);
9088         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9089         changed = 0;
9090         th_ack = th->th_ack;
9091         if (rack->sack_attack_disable == 0)
9092                 rack_do_decay(rack);
9093         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9094                 /*
9095                  * You only get credit for
9096                  * MSS and greater (and you get extra
9097                  * credit for larger cum-ack moves).
9098                  */
9099                 int ac;
9100
9101                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9102                 rack->r_ctl.ack_count += ac;
9103                 counter_u64_add(rack_ack_total, ac);
9104         }
9105         if (rack->r_ctl.ack_count > 0xfff00000) {
9106                 /*
9107                  * reduce the number to keep us under
9108                  * a uint32_t.
9109                  */
9110                 rack->r_ctl.ack_count /= 2;
9111                 rack->r_ctl.sack_count /= 2;
9112         }
9113         if (SEQ_GT(th_ack, tp->snd_una)) {
9114                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9115                 tp->t_acktime = ticks;
9116         }
9117         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9118                 changed = th_ack - rsm->r_start;
9119         if (changed) {
9120                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9121         }
9122         if ((to->to_flags & TOF_SACK) == 0) {
9123                 /* We are done nothing left and no sack. */
9124                 rack_handle_might_revert(tp, rack);
9125                 /*
9126                  * For cases where we struck a dup-ack
9127                  * with no SACK, add to the changes so
9128                  * PRR will work right.
9129                  */
9130                 if (dup_ack_struck && (changed == 0)) {
9131                         changed += ctf_fixed_maxseg(rack->rc_tp);
9132                 }
9133                 goto out;
9134         }
9135         /* Sack block processing */
9136         if (SEQ_GT(th_ack, tp->snd_una))
9137                 ack_point = th_ack;
9138         else
9139                 ack_point = tp->snd_una;
9140         for (i = 0; i < to->to_nsacks; i++) {
9141                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9142                       &sack, sizeof(sack));
9143                 sack.start = ntohl(sack.start);
9144                 sack.end = ntohl(sack.end);
9145                 if (SEQ_GT(sack.end, sack.start) &&
9146                     SEQ_GT(sack.start, ack_point) &&
9147                     SEQ_LT(sack.start, tp->snd_max) &&
9148                     SEQ_GT(sack.end, ack_point) &&
9149                     SEQ_LEQ(sack.end, tp->snd_max)) {
9150                         sack_blocks[num_sack_blks] = sack;
9151                         num_sack_blks++;
9152 #ifdef NETFLIX_STATS
9153                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9154                            SEQ_LEQ(sack.end, th_ack)) {
9155                         /*
9156                          * Its a D-SACK block.
9157                          */
9158                         tcp_record_dsack(sack.start, sack.end);
9159 #endif
9160                         rack_note_dsack(rack, sack.start, sack.end);
9161                 }
9162         }
9163         /*
9164          * Sort the SACK blocks so we can update the rack scoreboard with
9165          * just one pass.
9166          */
9167         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9168                                          num_sack_blks, th->th_ack);
9169         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9170         if (num_sack_blks == 0) {
9171                 /* Nothing to sack (DSACKs?) */
9172                 goto out_with_totals;
9173         }
9174         if (num_sack_blks < 2) {
9175                 /* Only one, we don't need to sort */
9176                 goto do_sack_work;
9177         }
9178         /* Sort the sacks */
9179         for (i = 0; i < num_sack_blks; i++) {
9180                 for (j = i + 1; j < num_sack_blks; j++) {
9181                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9182                                 sack = sack_blocks[i];
9183                                 sack_blocks[i] = sack_blocks[j];
9184                                 sack_blocks[j] = sack;
9185                         }
9186                 }
9187         }
9188         /*
9189          * Now are any of the sack block ends the same (yes some
9190          * implementations send these)?
9191          */
9192 again:
9193         if (num_sack_blks == 0)
9194                 goto out_with_totals;
9195         if (num_sack_blks > 1) {
9196                 for (i = 0; i < num_sack_blks; i++) {
9197                         for (j = i + 1; j < num_sack_blks; j++) {
9198                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9199                                         /*
9200                                          * Ok these two have the same end we
9201                                          * want the smallest end and then
9202                                          * throw away the larger and start
9203                                          * again.
9204                                          */
9205                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9206                                                 /*
9207                                                  * The second block covers
9208                                                  * more area use that
9209                                                  */
9210                                                 sack_blocks[i].start = sack_blocks[j].start;
9211                                         }
9212                                         /*
9213                                          * Now collapse out the dup-sack and
9214                                          * lower the count
9215                                          */
9216                                         for (k = (j + 1); k < num_sack_blks; k++) {
9217                                                 sack_blocks[j].start = sack_blocks[k].start;
9218                                                 sack_blocks[j].end = sack_blocks[k].end;
9219                                                 j++;
9220                                         }
9221                                         num_sack_blks--;
9222                                         goto again;
9223                                 }
9224                         }
9225                 }
9226         }
9227 do_sack_work:
9228         /*
9229          * First lets look to see if
9230          * we have retransmitted and
9231          * can use the transmit next?
9232          */
9233         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9234         if (rsm &&
9235             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9236             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9237                 /*
9238                  * We probably did the FR and the next
9239                  * SACK in continues as we would expect.
9240                  */
9241                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9242                 if (acked) {
9243                         rack->r_wanted_output = 1;
9244                         changed += acked;
9245                         sack_changed += acked;
9246                 }
9247                 if (num_sack_blks == 1) {
9248                         /*
9249                          * This is what we would expect from
9250                          * a normal implementation to happen
9251                          * after we have retransmitted the FR,
9252                          * i.e the sack-filter pushes down
9253                          * to 1 block and the next to be retransmitted
9254                          * is the sequence in the sack block (has more
9255                          * are acked). Count this as ACK'd data to boost
9256                          * up the chances of recovering any false positives.
9257                          */
9258                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9259                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9260                         counter_u64_add(rack_express_sack, 1);
9261                         if (rack->r_ctl.ack_count > 0xfff00000) {
9262                                 /*
9263                                  * reduce the number to keep us under
9264                                  * a uint32_t.
9265                                  */
9266                                 rack->r_ctl.ack_count /= 2;
9267                                 rack->r_ctl.sack_count /= 2;
9268                         }
9269                         goto out_with_totals;
9270                 } else {
9271                         /*
9272                          * Start the loop through the
9273                          * rest of blocks, past the first block.
9274                          */
9275                         moved_two = 0;
9276                         loop_start = 1;
9277                 }
9278         }
9279         /* Its a sack of some sort */
9280         rack->r_ctl.sack_count++;
9281         if (rack->r_ctl.sack_count > 0xfff00000) {
9282                 /*
9283                  * reduce the number to keep us under
9284                  * a uint32_t.
9285                  */
9286                 rack->r_ctl.ack_count /= 2;
9287                 rack->r_ctl.sack_count /= 2;
9288         }
9289         counter_u64_add(rack_sack_total, 1);
9290         if (rack->sack_attack_disable) {
9291                 /* An attacker disablement is in place */
9292                 if (num_sack_blks > 1) {
9293                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9294                         rack->r_ctl.sack_moved_extra++;
9295                         counter_u64_add(rack_move_some, 1);
9296                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9297                                 rack->r_ctl.sack_moved_extra /= 2;
9298                                 rack->r_ctl.sack_noextra_move /= 2;
9299                         }
9300                 }
9301                 goto out;
9302         }
9303         rsm = rack->r_ctl.rc_sacklast;
9304         for (i = loop_start; i < num_sack_blks; i++) {
9305                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9306                 if (acked) {
9307                         rack->r_wanted_output = 1;
9308                         changed += acked;
9309                         sack_changed += acked;
9310                 }
9311                 if (moved_two) {
9312                         /*
9313                          * If we did not get a SACK for at least a MSS and
9314                          * had to move at all, or if we moved more than our
9315                          * threshold, it counts against the "extra" move.
9316                          */
9317                         rack->r_ctl.sack_moved_extra += moved_two;
9318                         counter_u64_add(rack_move_some, 1);
9319                 } else {
9320                         /*
9321                          * else we did not have to move
9322                          * any more than we would expect.
9323                          */
9324                         rack->r_ctl.sack_noextra_move++;
9325                         counter_u64_add(rack_move_none, 1);
9326                 }
9327                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9328                         /*
9329                          * If the SACK was not a full MSS then
9330                          * we add to sack_count the number of
9331                          * MSS's (or possibly more than
9332                          * a MSS if its a TSO send) we had to skip by.
9333                          */
9334                         rack->r_ctl.sack_count += moved_two;
9335                         counter_u64_add(rack_sack_total, moved_two);
9336                 }
9337                 /*
9338                  * Now we need to setup for the next
9339                  * round. First we make sure we won't
9340                  * exceed the size of our uint32_t on
9341                  * the various counts, and then clear out
9342                  * moved_two.
9343                  */
9344                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9345                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9346                         rack->r_ctl.sack_moved_extra /= 2;
9347                         rack->r_ctl.sack_noextra_move /= 2;
9348                 }
9349                 if (rack->r_ctl.sack_count > 0xfff00000) {
9350                         rack->r_ctl.ack_count /= 2;
9351                         rack->r_ctl.sack_count /= 2;
9352                 }
9353                 moved_two = 0;
9354         }
9355 out_with_totals:
9356         if (num_sack_blks > 1) {
9357                 /*
9358                  * You get an extra stroke if
9359                  * you have more than one sack-blk, this
9360                  * could be where we are skipping forward
9361                  * and the sack-filter is still working, or
9362                  * it could be an attacker constantly
9363                  * moving us.
9364                  */
9365                 rack->r_ctl.sack_moved_extra++;
9366                 counter_u64_add(rack_move_some, 1);
9367         }
9368 out:
9369 #ifdef NETFLIX_EXP_DETECTION
9370         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9371 #endif
9372         if (changed) {
9373                 /* Something changed cancel the rack timer */
9374                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9375         }
9376         tsused = tcp_get_usecs(NULL);
9377         rsm = tcp_rack_output(tp, rack, tsused);
9378         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9379             rsm) {
9380                 /* Enter recovery */
9381                 rack->r_ctl.rc_rsm_start = rsm->r_start;
9382                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
9383                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
9384                 entered_recovery = 1;
9385                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9386                 /*
9387                  * When we enter recovery we need to assure we send
9388                  * one packet.
9389                  */
9390                 if (rack->rack_no_prr == 0) {
9391                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9392                         rack_log_to_prr(rack, 8, 0);
9393                 }
9394                 rack->r_timer_override = 1;
9395                 rack->r_early = 0;
9396                 rack->r_ctl.rc_agg_early = 0;
9397         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9398                    rsm &&
9399                    (rack->r_rr_config == 3)) {
9400                 /*
9401                  * Assure we can output and we get no
9402                  * remembered pace time except the retransmit.
9403                  */
9404                 rack->r_timer_override = 1;
9405                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9406                 rack->r_ctl.rc_resend = rsm;
9407         }
9408         if (IN_FASTRECOVERY(tp->t_flags) &&
9409             (rack->rack_no_prr == 0) &&
9410             (entered_recovery == 0)) {
9411                 rack_update_prr(tp, rack, changed, th_ack);
9412                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9413                      ((rack->rc_inp->inp_in_hpts == 0) &&
9414                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9415                         /*
9416                          * If you are pacing output you don't want
9417                          * to override.
9418                          */
9419                         rack->r_early = 0;
9420                         rack->r_ctl.rc_agg_early = 0;
9421                         rack->r_timer_override = 1;
9422                 }
9423         }
9424 }
9425
9426 static void
9427 rack_strike_dupack(struct tcp_rack *rack)
9428 {
9429         struct rack_sendmap *rsm;
9430
9431         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9432         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9433                 rsm = TAILQ_NEXT(rsm, r_tnext);
9434         }
9435         if (rsm && (rsm->r_dupack < 0xff)) {
9436                 rsm->r_dupack++;
9437                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9438                         struct timeval tv;
9439                         uint32_t cts;
9440                         /*
9441                          * Here we see if we need to retransmit. For
9442                          * a SACK type connection if enough time has passed
9443                          * we will get a return of the rsm. For a non-sack
9444                          * connection we will get the rsm returned if the
9445                          * dupack value is 3 or more.
9446                          */
9447                         cts = tcp_get_usecs(&tv);
9448                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9449                         if (rack->r_ctl.rc_resend != NULL) {
9450                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9451                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9452                                                          rack->rc_tp->snd_una);
9453                                 }
9454                                 rack->r_wanted_output = 1;
9455                                 rack->r_timer_override = 1;
9456                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9457                         }
9458                 } else {
9459                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9460                 }
9461         }
9462 }
9463
9464 static void
9465 rack_check_bottom_drag(struct tcpcb *tp,
9466                        struct tcp_rack *rack,
9467                        struct socket *so, int32_t acked)
9468 {
9469         uint32_t segsiz, minseg;
9470
9471         segsiz = ctf_fixed_maxseg(tp);
9472         minseg = segsiz;
9473
9474         if (tp->snd_max == tp->snd_una) {
9475                 /*
9476                  * We are doing dynamic pacing and we are way
9477                  * under. Basically everything got acked while
9478                  * we were still waiting on the pacer to expire.
9479                  *
9480                  * This means we need to boost the b/w in
9481                  * addition to any earlier boosting of
9482                  * the multipler.
9483                  */
9484                 rack->rc_dragged_bottom = 1;
9485                 rack_validate_multipliers_at_or_above100(rack);
9486                 /*
9487                  * Lets use the segment bytes acked plus
9488                  * the lowest RTT seen as the basis to
9489                  * form a b/w estimate. This will be off
9490                  * due to the fact that the true estimate
9491                  * should be around 1/2 the time of the RTT
9492                  * but we can settle for that.
9493                  */
9494                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9495                     acked) {
9496                         uint64_t bw, calc_bw, rtt;
9497
9498                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9499                         if (rtt == 0) {
9500                                 /* no us sample is there a ms one? */
9501                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9502                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9503                                 } else {
9504                                         goto no_measurement;
9505                                 }
9506                         }
9507                         bw = acked;
9508                         calc_bw = bw * 1000000;
9509                         calc_bw /= rtt;
9510                         if (rack->r_ctl.last_max_bw &&
9511                             (rack->r_ctl.last_max_bw < calc_bw)) {
9512                                 /*
9513                                  * If we have a last calculated max bw
9514                                  * enforce it.
9515                                  */
9516                                 calc_bw = rack->r_ctl.last_max_bw;
9517                         }
9518                         /* now plop it in */
9519                         if (rack->rc_gp_filled == 0) {
9520                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9521                                         /*
9522                                          * If we have no measurement
9523                                          * don't let us set in more than
9524                                          * 1.2Mbps. If we are still too
9525                                          * low after pacing with this we
9526                                          * will hopefully have a max b/w
9527                                          * available to sanity check things.
9528                                          */
9529                                         calc_bw = ONE_POINT_TWO_MEG;
9530                                 }
9531                                 rack->r_ctl.rc_rtt_diff = 0;
9532                                 rack->r_ctl.gp_bw = calc_bw;
9533                                 rack->rc_gp_filled = 1;
9534                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9535                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9536                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9537                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9538                                 rack->r_ctl.rc_rtt_diff = 0;
9539                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9540                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9541                                 rack->r_ctl.gp_bw = calc_bw;
9542                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9543                         } else
9544                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9545                         if ((rack->gp_ready == 0) &&
9546                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9547                                 /* We have enough measurements now */
9548                                 rack->gp_ready = 1;
9549                                 rack_set_cc_pacing(rack);
9550                                 if (rack->defer_options)
9551                                         rack_apply_deferred_options(rack);
9552                         }
9553                         /*
9554                          * For acks over 1mss we do a extra boost to simulate
9555                          * where we would get 2 acks (we want 110 for the mul).
9556                          */
9557                         if (acked > segsiz)
9558                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9559                 } else {
9560                         /*
9561                          * zero rtt possibly?, settle for just an old increase.
9562                          */
9563 no_measurement:
9564                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9565                 }
9566         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9567                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9568                                                minseg)) &&
9569                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9570                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9571                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9572                     (segsiz * rack_req_segs))) {
9573                 /*
9574                  * We are doing dynamic GP pacing and
9575                  * we have everything except 1MSS or less
9576                  * bytes left out. We are still pacing away.
9577                  * And there is data that could be sent, This
9578                  * means we are inserting delayed ack time in
9579                  * our measurements because we are pacing too slow.
9580                  */
9581                 rack_validate_multipliers_at_or_above100(rack);
9582                 rack->rc_dragged_bottom = 1;
9583                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9584         }
9585 }
9586
9587
9588
9589 static void
9590 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9591 {
9592         /*
9593          * The fast output path is enabled and we
9594          * have moved the cumack forward. Lets see if
9595          * we can expand forward the fast path length by
9596          * that amount. What we would ideally like to
9597          * do is increase the number of bytes in the
9598          * fast path block (left_to_send) by the
9599          * acked amount. However we have to gate that
9600          * by two factors:
9601          * 1) The amount outstanding and the rwnd of the peer
9602          *    (i.e. we don't want to exceed the rwnd of the peer).
9603          *    <and>
9604          * 2) The amount of data left in the socket buffer (i.e.
9605          *    we can't send beyond what is in the buffer).
9606          *
9607          * Note that this does not take into account any increase
9608          * in the cwnd. We will only extend the fast path by
9609          * what was acked.
9610          */
9611         uint32_t new_total, gating_val;
9612
9613         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9614         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9615                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9616         if (new_total <= gating_val) {
9617                 /* We can increase left_to_send by the acked amount */
9618                 counter_u64_add(rack_extended_rfo, 1);
9619                 rack->r_ctl.fsb.left_to_send = new_total;
9620                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9621                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
9622                          rack, rack->r_ctl.fsb.left_to_send,
9623                          sbavail(&rack->rc_inp->inp_socket->so_snd),
9624                          (tp->snd_max - tp->snd_una)));
9625
9626         }
9627 }
9628
9629 static void
9630 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9631 {
9632         /*
9633          * Here any sendmap entry that points to the
9634          * beginning mbuf must be adjusted to the correct
9635          * offset. This must be called with:
9636          * 1) The socket buffer locked
9637          * 2) snd_una adjusted to its new postion.
9638          *
9639          * Note that (2) implies rack_ack_received has also
9640          * been called.
9641          *
9642          * We grab the first mbuf in the socket buffer and
9643          * then go through the front of the sendmap, recalculating
9644          * the stored offset for any sendmap entry that has
9645          * that mbuf. We must use the sb functions to do this
9646          * since its possible an add was done has well as
9647          * the subtraction we may have just completed. This should
9648          * not be a penalty though, since we just referenced the sb
9649          * to go in and trim off the mbufs that we freed (of course
9650          * there will be a penalty for the sendmap references though).
9651          */
9652         struct mbuf *m;
9653         struct rack_sendmap *rsm;
9654
9655         SOCKBUF_LOCK_ASSERT(sb);
9656         m = sb->sb_mb;
9657         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9658         if ((rsm == NULL) || (m == NULL)) {
9659                 /* Nothing outstanding */
9660                 return;
9661         }
9662         while (rsm->m && (rsm->m == m)) {
9663                 /* one to adjust */
9664 #ifdef INVARIANTS
9665                 struct mbuf *tm;
9666                 uint32_t soff;
9667
9668                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9669                 if (rsm->orig_m_len != m->m_len) {
9670                         rack_adjust_orig_mlen(rsm);
9671                 }
9672                 if (rsm->soff != soff) {
9673                         /*
9674                          * This is not a fatal error, we anticipate it
9675                          * might happen (the else code), so we count it here
9676                          * so that under invariant we can see that it really
9677                          * does happen.
9678                          */
9679                         counter_u64_add(rack_adjust_map_bw, 1);
9680                 }
9681                 rsm->m = tm;
9682                 rsm->soff = soff;
9683                 if (tm)
9684                         rsm->orig_m_len = rsm->m->m_len;
9685                 else
9686                         rsm->orig_m_len = 0;
9687 #else
9688                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
9689                 if (rsm->m)
9690                         rsm->orig_m_len = rsm->m->m_len;
9691                 else
9692                         rsm->orig_m_len = 0;
9693 #endif
9694                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
9695                               rsm);
9696                 if (rsm == NULL)
9697                         break;
9698         }
9699 }
9700
9701 /*
9702  * Return value of 1, we do not need to call rack_process_data().
9703  * return value of 0, rack_process_data can be called.
9704  * For ret_val if its 0 the TCP is locked, if its non-zero
9705  * its unlocked and probably unsafe to touch the TCB.
9706  */
9707 static int
9708 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9709     struct tcpcb *tp, struct tcpopt *to,
9710     uint32_t tiwin, int32_t tlen,
9711     int32_t * ofia, int32_t thflags, int32_t *ret_val)
9712 {
9713         int32_t ourfinisacked = 0;
9714         int32_t nsegs, acked_amount;
9715         int32_t acked;
9716         struct mbuf *mfree;
9717         struct tcp_rack *rack;
9718         int32_t under_pacing = 0;
9719         int32_t recovery = 0;
9720
9721         rack = (struct tcp_rack *)tp->t_fb_ptr;
9722         if (SEQ_GT(th->th_ack, tp->snd_max)) {
9723                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
9724                                       &rack->r_ctl.challenge_ack_ts,
9725                                       &rack->r_ctl.challenge_ack_cnt);
9726                 rack->r_wanted_output = 1;
9727                 return (1);
9728         }
9729         if (rack->gp_ready &&
9730             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
9731                 under_pacing = 1;
9732         }
9733         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
9734                 int in_rec, dup_ack_struck = 0;
9735
9736                 in_rec = IN_FASTRECOVERY(tp->t_flags);
9737                 if (rack->rc_in_persist) {
9738                         tp->t_rxtshift = 0;
9739                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9740                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9741                 }
9742                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) {
9743                         rack_strike_dupack(rack);
9744                         dup_ack_struck = 1;
9745                 }
9746                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
9747         }
9748         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
9749                 /*
9750                  * Old ack, behind (or duplicate to) the last one rcv'd
9751                  * Note: We mark reordering is occuring if its
9752                  * less than and we have not closed our window.
9753                  */
9754                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
9755                         counter_u64_add(rack_reorder_seen, 1);
9756                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9757                 }
9758                 return (0);
9759         }
9760         /*
9761          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
9762          * something we sent.
9763          */
9764         if (tp->t_flags & TF_NEEDSYN) {
9765                 /*
9766                  * T/TCP: Connection was half-synchronized, and our SYN has
9767                  * been ACK'd (so connection is now fully synchronized).  Go
9768                  * to non-starred state, increment snd_una for ACK of SYN,
9769                  * and check if we can do window scaling.
9770                  */
9771                 tp->t_flags &= ~TF_NEEDSYN;
9772                 tp->snd_una++;
9773                 /* Do window scaling? */
9774                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9775                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9776                         tp->rcv_scale = tp->request_r_scale;
9777                         /* Send window already scaled. */
9778                 }
9779         }
9780         nsegs = max(1, m->m_pkthdr.lro_nsegs);
9781         INP_WLOCK_ASSERT(tp->t_inpcb);
9782
9783         acked = BYTES_THIS_ACK(tp, th);
9784         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
9785         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
9786         /*
9787          * If we just performed our first retransmit, and the ACK arrives
9788          * within our recovery window, then it was a mistake to do the
9789          * retransmit in the first place.  Recover our original cwnd and
9790          * ssthresh, and proceed to transmit where we left off.
9791          */
9792         if ((tp->t_flags & TF_PREVVALID) &&
9793             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
9794                 tp->t_flags &= ~TF_PREVVALID;
9795                 if (tp->t_rxtshift == 1 &&
9796                     (int)(ticks - tp->t_badrxtwin) < 0)
9797                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
9798         }
9799         if (acked) {
9800                 /* assure we are not backed off */
9801                 tp->t_rxtshift = 0;
9802                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9803                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9804                 rack->rc_tlp_in_progress = 0;
9805                 rack->r_ctl.rc_tlp_cnt_out = 0;
9806                 /*
9807                  * If it is the RXT timer we want to
9808                  * stop it, so we can restart a TLP.
9809                  */
9810                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
9811                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9812 #ifdef NETFLIX_HTTP_LOGGING
9813                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
9814 #endif
9815         }
9816         /*
9817          * If we have a timestamp reply, update smoothed round trip time. If
9818          * no timestamp is present but transmit timer is running and timed
9819          * sequence number was acked, update smoothed round trip time. Since
9820          * we now have an rtt measurement, cancel the timer backoff (cf.,
9821          * Phil Karn's retransmit alg.). Recompute the initial retransmit
9822          * timer.
9823          *
9824          * Some boxes send broken timestamp replies during the SYN+ACK
9825          * phase, ignore timestamps of 0 or we could calculate a huge RTT
9826          * and blow up the retransmit timer.
9827          */
9828         /*
9829          * If all outstanding data is acked, stop retransmit timer and
9830          * remember to restart (more output or persist). If there is more
9831          * data to be acked, restart retransmit timer, using current
9832          * (possibly backed-off) value.
9833          */
9834         if (acked == 0) {
9835                 if (ofia)
9836                         *ofia = ourfinisacked;
9837                 return (0);
9838         }
9839         if (IN_RECOVERY(tp->t_flags)) {
9840                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
9841                     (SEQ_LT(th->th_ack, tp->snd_max))) {
9842                         tcp_rack_partialack(tp);
9843                 } else {
9844                         rack_post_recovery(tp, th->th_ack);
9845                         recovery = 1;
9846                 }
9847         }
9848         /*
9849          * Let the congestion control algorithm update congestion control
9850          * related information. This typically means increasing the
9851          * congestion window.
9852          */
9853         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
9854         SOCKBUF_LOCK(&so->so_snd);
9855         acked_amount = min(acked, (int)sbavail(&so->so_snd));
9856         tp->snd_wnd -= acked_amount;
9857         mfree = sbcut_locked(&so->so_snd, acked_amount);
9858         if ((sbused(&so->so_snd) == 0) &&
9859             (acked > acked_amount) &&
9860             (tp->t_state >= TCPS_FIN_WAIT_1) &&
9861             (tp->t_flags & TF_SENTFIN)) {
9862                 /*
9863                  * We must be sure our fin
9864                  * was sent and acked (we can be
9865                  * in FIN_WAIT_1 without having
9866                  * sent the fin).
9867                  */
9868                 ourfinisacked = 1;
9869         }
9870         tp->snd_una = th->th_ack;
9871         if (acked_amount && sbavail(&so->so_snd))
9872                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
9873         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
9874         /* NB: sowwakeup_locked() does an implicit unlock. */
9875         sowwakeup_locked(so);
9876         m_freem(mfree);
9877         if (SEQ_GT(tp->snd_una, tp->snd_recover))
9878                 tp->snd_recover = tp->snd_una;
9879
9880         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
9881                 tp->snd_nxt = tp->snd_una;
9882         }
9883         if (under_pacing &&
9884             (rack->use_fixed_rate == 0) &&
9885             (rack->in_probe_rtt == 0) &&
9886             rack->rc_gp_dyn_mul &&
9887             rack->rc_always_pace) {
9888                 /* Check if we are dragging bottom */
9889                 rack_check_bottom_drag(tp, rack, so, acked);
9890         }
9891         if (tp->snd_una == tp->snd_max) {
9892                 /* Nothing left outstanding */
9893                 tp->t_flags &= ~TF_PREVVALID;
9894                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
9895                 rack->r_ctl.retran_during_recovery = 0;
9896                 rack->r_ctl.dsack_byte_cnt = 0;
9897                 if (rack->r_ctl.rc_went_idle_time == 0)
9898                         rack->r_ctl.rc_went_idle_time = 1;
9899                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
9900                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
9901                         tp->t_acktime = 0;
9902                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9903                 /* Set need output so persist might get set */
9904                 rack->r_wanted_output = 1;
9905                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
9906                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
9907                     (sbavail(&so->so_snd) == 0) &&
9908                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
9909                         /*
9910                          * The socket was gone and the
9911                          * peer sent data (now or in the past), time to
9912                          * reset him.
9913                          */
9914                         *ret_val = 1;
9915                         /* tcp_close will kill the inp pre-log the Reset */
9916                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9917                         tp = tcp_close(tp);
9918                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
9919                         return (1);
9920                 }
9921         }
9922         if (ofia)
9923                 *ofia = ourfinisacked;
9924         return (0);
9925 }
9926
9927 static void
9928 rack_collapsed_window(struct tcp_rack *rack)
9929 {
9930         /*
9931          * Now we must walk the
9932          * send map and divide the
9933          * ones left stranded. These
9934          * guys can't cause us to abort
9935          * the connection and are really
9936          * "unsent". However if a buggy
9937          * client actually did keep some
9938          * of the data i.e. collapsed the win
9939          * and refused to ack and then opened
9940          * the win and acked that data. We would
9941          * get into an ack war, the simplier
9942          * method then of just pretending we
9943          * did not send those segments something
9944          * won't work.
9945          */
9946         struct rack_sendmap *rsm, *nrsm, fe, *insret;
9947         tcp_seq max_seq;
9948
9949         max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
9950         memset(&fe, 0, sizeof(fe));
9951         fe.r_start = max_seq;
9952         /* Find the first seq past or at maxseq */
9953         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
9954         if (rsm == NULL) {
9955                 /* Nothing to do strange */
9956                 rack->rc_has_collapsed = 0;
9957                 return;
9958         }
9959         /*
9960          * Now do we need to split at
9961          * the collapse point?
9962          */
9963         if (SEQ_GT(max_seq, rsm->r_start)) {
9964                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9965                 if (nrsm == NULL) {
9966                         /* We can't get a rsm, mark all? */
9967                         nrsm = rsm;
9968                         goto no_split;
9969                 }
9970                 /* Clone it */
9971                 rack_clone_rsm(rack, nrsm, rsm, max_seq);
9972                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
9973 #ifdef INVARIANTS
9974                 if (insret != NULL) {
9975                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
9976                               nrsm, insret, rack, rsm);
9977                 }
9978 #endif
9979                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
9980                 if (rsm->r_in_tmap) {
9981                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9982                         nrsm->r_in_tmap = 1;
9983                 }
9984                 /*
9985                  * Set in the new RSM as the
9986                  * collapsed starting point
9987                  */
9988                 rsm = nrsm;
9989         }
9990 no_split:
9991         counter_u64_add(rack_collapsed_win, 1);
9992         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
9993                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
9994         }
9995         rack->rc_has_collapsed = 1;
9996 }
9997
9998 static void
9999 rack_un_collapse_window(struct tcp_rack *rack)
10000 {
10001         struct rack_sendmap *rsm;
10002
10003         RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
10004                 if (rsm->r_flags & RACK_RWND_COLLAPSED)
10005                         rsm->r_flags &= ~RACK_RWND_COLLAPSED;
10006                 else
10007                         break;
10008         }
10009         rack->rc_has_collapsed = 0;
10010 }
10011
10012 static void
10013 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10014                         int32_t tlen, int32_t tfo_syn)
10015 {
10016         if (DELAY_ACK(tp, tlen) || tfo_syn) {
10017                 if (rack->rc_dack_mode &&
10018                     (tlen > 500) &&
10019                     (rack->rc_dack_toggle == 1)) {
10020                         goto no_delayed_ack;
10021                 }
10022                 rack_timer_cancel(tp, rack,
10023                                   rack->r_ctl.rc_rcvtime, __LINE__);
10024                 tp->t_flags |= TF_DELACK;
10025         } else {
10026 no_delayed_ack:
10027                 rack->r_wanted_output = 1;
10028                 tp->t_flags |= TF_ACKNOW;
10029                 if (rack->rc_dack_mode) {
10030                         if (tp->t_flags & TF_DELACK)
10031                                 rack->rc_dack_toggle = 1;
10032                         else
10033                                 rack->rc_dack_toggle = 0;
10034                 }
10035         }
10036 }
10037
10038 static void
10039 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10040 {
10041         /*
10042          * If fast output is in progress, lets validate that
10043          * the new window did not shrink on us and make it
10044          * so fast output should end.
10045          */
10046         if (rack->r_fast_output) {
10047                 uint32_t out;
10048
10049                 /*
10050                  * Calculate what we will send if left as is
10051                  * and compare that to our send window.
10052                  */
10053                 out = ctf_outstanding(tp);
10054                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10055                         /* ok we have an issue */
10056                         if (out >= tp->snd_wnd) {
10057                                 /* Turn off fast output the window is met or collapsed */
10058                                 rack->r_fast_output = 0;
10059                         } else {
10060                                 /* we have some room left */
10061                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10062                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10063                                         /* If not at least 1 full segment never mind */
10064                                         rack->r_fast_output = 0;
10065                                 }
10066                         }
10067                 }
10068         }
10069 }
10070
10071
10072 /*
10073  * Return value of 1, the TCB is unlocked and most
10074  * likely gone, return value of 0, the TCP is still
10075  * locked.
10076  */
10077 static int
10078 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10079     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10080     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10081 {
10082         /*
10083          * Update window information. Don't look at window if no ACK: TAC's
10084          * send garbage on first SYN.
10085          */
10086         int32_t nsegs;
10087         int32_t tfo_syn;
10088         struct tcp_rack *rack;
10089
10090         rack = (struct tcp_rack *)tp->t_fb_ptr;
10091         INP_WLOCK_ASSERT(tp->t_inpcb);
10092         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10093         if ((thflags & TH_ACK) &&
10094             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10095             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10096             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10097                 /* keep track of pure window updates */
10098                 if (tlen == 0 &&
10099                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10100                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10101                 tp->snd_wnd = tiwin;
10102                 rack_validate_fo_sendwin_up(tp, rack);
10103                 tp->snd_wl1 = th->th_seq;
10104                 tp->snd_wl2 = th->th_ack;
10105                 if (tp->snd_wnd > tp->max_sndwnd)
10106                         tp->max_sndwnd = tp->snd_wnd;
10107                 rack->r_wanted_output = 1;
10108         } else if (thflags & TH_ACK) {
10109                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10110                         tp->snd_wnd = tiwin;
10111                         rack_validate_fo_sendwin_up(tp, rack);
10112                         tp->snd_wl1 = th->th_seq;
10113                         tp->snd_wl2 = th->th_ack;
10114                 }
10115         }
10116         if (tp->snd_wnd < ctf_outstanding(tp))
10117                 /* The peer collapsed the window */
10118                 rack_collapsed_window(rack);
10119         else if (rack->rc_has_collapsed)
10120                 rack_un_collapse_window(rack);
10121         /* Was persist timer active and now we have window space? */
10122         if ((rack->rc_in_persist != 0) &&
10123             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10124                                 rack->r_ctl.rc_pace_min_segs))) {
10125                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10126                 tp->snd_nxt = tp->snd_max;
10127                 /* Make sure we output to start the timer */
10128                 rack->r_wanted_output = 1;
10129         }
10130         /* Do we enter persists? */
10131         if ((rack->rc_in_persist == 0) &&
10132             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10133             TCPS_HAVEESTABLISHED(tp->t_state) &&
10134             (tp->snd_max == tp->snd_una) &&
10135             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10136             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10137                 /*
10138                  * Here the rwnd is less than
10139                  * the pacing size, we are established,
10140                  * nothing is outstanding, and there is
10141                  * data to send. Enter persists.
10142                  */
10143                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10144         }
10145         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10146                 m_freem(m);
10147                 return (0);
10148         }
10149         /*
10150          * don't process the URG bit, ignore them drag
10151          * along the up.
10152          */
10153         tp->rcv_up = tp->rcv_nxt;
10154         INP_WLOCK_ASSERT(tp->t_inpcb);
10155
10156         /*
10157          * Process the segment text, merging it into the TCP sequencing
10158          * queue, and arranging for acknowledgment of receipt if necessary.
10159          * This process logically involves adjusting tp->rcv_wnd as data is
10160          * presented to the user (this happens in tcp_usrreq.c, case
10161          * PRU_RCVD).  If a FIN has already been received on this connection
10162          * then we just ignore the text.
10163          */
10164         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10165                    IS_FASTOPEN(tp->t_flags));
10166         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10167             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10168                 tcp_seq save_start = th->th_seq;
10169                 tcp_seq save_rnxt  = tp->rcv_nxt;
10170                 int     save_tlen  = tlen;
10171
10172                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10173                 /*
10174                  * Insert segment which includes th into TCP reassembly
10175                  * queue with control block tp.  Set thflags to whether
10176                  * reassembly now includes a segment with FIN.  This handles
10177                  * the common case inline (segment is the next to be
10178                  * received on an established connection, and the queue is
10179                  * empty), avoiding linkage into and removal from the queue
10180                  * and repetition of various conversions. Set DELACK for
10181                  * segments received in order, but ack immediately when
10182                  * segments are out of order (so fast retransmit can work).
10183                  */
10184                 if (th->th_seq == tp->rcv_nxt &&
10185                     SEGQ_EMPTY(tp) &&
10186                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10187                     tfo_syn)) {
10188 #ifdef NETFLIX_SB_LIMITS
10189                         u_int mcnt, appended;
10190
10191                         if (so->so_rcv.sb_shlim) {
10192                                 mcnt = m_memcnt(m);
10193                                 appended = 0;
10194                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10195                                     CFO_NOSLEEP, NULL) == false) {
10196                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10197                                         m_freem(m);
10198                                         return (0);
10199                                 }
10200                         }
10201 #endif
10202                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10203                         tp->rcv_nxt += tlen;
10204                         if (tlen &&
10205                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10206                             (tp->t_fbyte_in == 0)) {
10207                                 tp->t_fbyte_in = ticks;
10208                                 if (tp->t_fbyte_in == 0)
10209                                         tp->t_fbyte_in = 1;
10210                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10211                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10212                         }
10213                         thflags = th->th_flags & TH_FIN;
10214                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10215                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10216                         SOCKBUF_LOCK(&so->so_rcv);
10217                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10218                                 m_freem(m);
10219                         } else
10220 #ifdef NETFLIX_SB_LIMITS
10221                                 appended =
10222 #endif
10223                                         sbappendstream_locked(&so->so_rcv, m, 0);
10224
10225                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10226                         /* NB: sorwakeup_locked() does an implicit unlock. */
10227                         sorwakeup_locked(so);
10228 #ifdef NETFLIX_SB_LIMITS
10229                         if (so->so_rcv.sb_shlim && appended != mcnt)
10230                                 counter_fo_release(so->so_rcv.sb_shlim,
10231                                     mcnt - appended);
10232 #endif
10233                 } else {
10234                         /*
10235                          * XXX: Due to the header drop above "th" is
10236                          * theoretically invalid by now.  Fortunately
10237                          * m_adj() doesn't actually frees any mbufs when
10238                          * trimming from the head.
10239                          */
10240                         tcp_seq temp = save_start;
10241
10242                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10243                         tp->t_flags |= TF_ACKNOW;
10244                         if (tp->t_flags & TF_WAKESOR) {
10245                                 tp->t_flags &= ~TF_WAKESOR;
10246                                 /* NB: sorwakeup_locked() does an implicit unlock. */
10247                                 sorwakeup_locked(so);
10248                         }
10249                 }
10250                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10251                     (save_tlen > 0) &&
10252                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10253                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10254                                 /*
10255                                  * DSACK actually handled in the fastpath
10256                                  * above.
10257                                  */
10258                                 RACK_OPTS_INC(tcp_sack_path_1);
10259                                 tcp_update_sack_list(tp, save_start,
10260                                     save_start + save_tlen);
10261                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10262                                 if ((tp->rcv_numsacks >= 1) &&
10263                                     (tp->sackblks[0].end == save_start)) {
10264                                         /*
10265                                          * Partial overlap, recorded at todrop
10266                                          * above.
10267                                          */
10268                                         RACK_OPTS_INC(tcp_sack_path_2a);
10269                                         tcp_update_sack_list(tp,
10270                                             tp->sackblks[0].start,
10271                                             tp->sackblks[0].end);
10272                                 } else {
10273                                         RACK_OPTS_INC(tcp_sack_path_2b);
10274                                         tcp_update_dsack_list(tp, save_start,
10275                                             save_start + save_tlen);
10276                                 }
10277                         } else if (tlen >= save_tlen) {
10278                                 /* Update of sackblks. */
10279                                 RACK_OPTS_INC(tcp_sack_path_3);
10280                                 tcp_update_dsack_list(tp, save_start,
10281                                     save_start + save_tlen);
10282                         } else if (tlen > 0) {
10283                                 RACK_OPTS_INC(tcp_sack_path_4);
10284                                 tcp_update_dsack_list(tp, save_start,
10285                                     save_start + tlen);
10286                         }
10287                 }
10288         } else {
10289                 m_freem(m);
10290                 thflags &= ~TH_FIN;
10291         }
10292
10293         /*
10294          * If FIN is received ACK the FIN and let the user know that the
10295          * connection is closing.
10296          */
10297         if (thflags & TH_FIN) {
10298                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10299                         /* The socket upcall is handled by socantrcvmore. */
10300                         socantrcvmore(so);
10301                         /*
10302                          * If connection is half-synchronized (ie NEEDSYN
10303                          * flag on) then delay ACK, so it may be piggybacked
10304                          * when SYN is sent. Otherwise, since we received a
10305                          * FIN then no more input can be expected, send ACK
10306                          * now.
10307                          */
10308                         if (tp->t_flags & TF_NEEDSYN) {
10309                                 rack_timer_cancel(tp, rack,
10310                                     rack->r_ctl.rc_rcvtime, __LINE__);
10311                                 tp->t_flags |= TF_DELACK;
10312                         } else {
10313                                 tp->t_flags |= TF_ACKNOW;
10314                         }
10315                         tp->rcv_nxt++;
10316                 }
10317                 switch (tp->t_state) {
10318                         /*
10319                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10320                          * CLOSE_WAIT state.
10321                          */
10322                 case TCPS_SYN_RECEIVED:
10323                         tp->t_starttime = ticks;
10324                         /* FALLTHROUGH */
10325                 case TCPS_ESTABLISHED:
10326                         rack_timer_cancel(tp, rack,
10327                             rack->r_ctl.rc_rcvtime, __LINE__);
10328                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10329                         break;
10330
10331                         /*
10332                          * If still in FIN_WAIT_1 STATE FIN has not been
10333                          * acked so enter the CLOSING state.
10334                          */
10335                 case TCPS_FIN_WAIT_1:
10336                         rack_timer_cancel(tp, rack,
10337                             rack->r_ctl.rc_rcvtime, __LINE__);
10338                         tcp_state_change(tp, TCPS_CLOSING);
10339                         break;
10340
10341                         /*
10342                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10343                          * starting the time-wait timer, turning off the
10344                          * other standard timers.
10345                          */
10346                 case TCPS_FIN_WAIT_2:
10347                         rack_timer_cancel(tp, rack,
10348                             rack->r_ctl.rc_rcvtime, __LINE__);
10349                         tcp_twstart(tp);
10350                         return (1);
10351                 }
10352         }
10353         /*
10354          * Return any desired output.
10355          */
10356         if ((tp->t_flags & TF_ACKNOW) ||
10357             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10358                 rack->r_wanted_output = 1;
10359         }
10360         INP_WLOCK_ASSERT(tp->t_inpcb);
10361         return (0);
10362 }
10363
10364 /*
10365  * Here nothing is really faster, its just that we
10366  * have broken out the fast-data path also just like
10367  * the fast-ack.
10368  */
10369 static int
10370 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10371     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10372     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10373 {
10374         int32_t nsegs;
10375         int32_t newsize = 0;    /* automatic sockbuf scaling */
10376         struct tcp_rack *rack;
10377 #ifdef NETFLIX_SB_LIMITS
10378         u_int mcnt, appended;
10379 #endif
10380 #ifdef TCPDEBUG
10381         /*
10382          * The size of tcp_saveipgen must be the size of the max ip header,
10383          * now IPv6.
10384          */
10385         u_char tcp_saveipgen[IP6_HDR_LEN];
10386         struct tcphdr tcp_savetcp;
10387         short ostate = 0;
10388
10389 #endif
10390         /*
10391          * If last ACK falls within this segment's sequence numbers, record
10392          * the timestamp. NOTE that the test is modified according to the
10393          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10394          */
10395         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10396                 return (0);
10397         }
10398         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10399                 return (0);
10400         }
10401         if (tiwin && tiwin != tp->snd_wnd) {
10402                 return (0);
10403         }
10404         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10405                 return (0);
10406         }
10407         if (__predict_false((to->to_flags & TOF_TS) &&
10408             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10409                 return (0);
10410         }
10411         if (__predict_false((th->th_ack != tp->snd_una))) {
10412                 return (0);
10413         }
10414         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10415                 return (0);
10416         }
10417         if ((to->to_flags & TOF_TS) != 0 &&
10418             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10419                 tp->ts_recent_age = tcp_ts_getticks();
10420                 tp->ts_recent = to->to_tsval;
10421         }
10422         rack = (struct tcp_rack *)tp->t_fb_ptr;
10423         /*
10424          * This is a pure, in-sequence data packet with nothing on the
10425          * reassembly queue and we have enough buffer space to take it.
10426          */
10427         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10428
10429 #ifdef NETFLIX_SB_LIMITS
10430         if (so->so_rcv.sb_shlim) {
10431                 mcnt = m_memcnt(m);
10432                 appended = 0;
10433                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10434                     CFO_NOSLEEP, NULL) == false) {
10435                         counter_u64_add(tcp_sb_shlim_fails, 1);
10436                         m_freem(m);
10437                         return (1);
10438                 }
10439         }
10440 #endif
10441         /* Clean receiver SACK report if present */
10442         if (tp->rcv_numsacks)
10443                 tcp_clean_sackreport(tp);
10444         KMOD_TCPSTAT_INC(tcps_preddat);
10445         tp->rcv_nxt += tlen;
10446         if (tlen &&
10447             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10448             (tp->t_fbyte_in == 0)) {
10449                 tp->t_fbyte_in = ticks;
10450                 if (tp->t_fbyte_in == 0)
10451                         tp->t_fbyte_in = 1;
10452                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10453                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10454         }
10455         /*
10456          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10457          */
10458         tp->snd_wl1 = th->th_seq;
10459         /*
10460          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10461          */
10462         tp->rcv_up = tp->rcv_nxt;
10463         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10464         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10465 #ifdef TCPDEBUG
10466         if (so->so_options & SO_DEBUG)
10467                 tcp_trace(TA_INPUT, ostate, tp,
10468                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
10469 #endif
10470         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10471
10472         /* Add data to socket buffer. */
10473         SOCKBUF_LOCK(&so->so_rcv);
10474         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10475                 m_freem(m);
10476         } else {
10477                 /*
10478                  * Set new socket buffer size. Give up when limit is
10479                  * reached.
10480                  */
10481                 if (newsize)
10482                         if (!sbreserve_locked(&so->so_rcv,
10483                             newsize, so, NULL))
10484                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10485                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10486 #ifdef NETFLIX_SB_LIMITS
10487                 appended =
10488 #endif
10489                         sbappendstream_locked(&so->so_rcv, m, 0);
10490                 ctf_calc_rwin(so, tp);
10491         }
10492         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10493         /* NB: sorwakeup_locked() does an implicit unlock. */
10494         sorwakeup_locked(so);
10495 #ifdef NETFLIX_SB_LIMITS
10496         if (so->so_rcv.sb_shlim && mcnt != appended)
10497                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10498 #endif
10499         rack_handle_delayed_ack(tp, rack, tlen, 0);
10500         if (tp->snd_una == tp->snd_max)
10501                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10502         return (1);
10503 }
10504
10505 /*
10506  * This subfunction is used to try to highly optimize the
10507  * fast path. We again allow window updates that are
10508  * in sequence to remain in the fast-path. We also add
10509  * in the __predict's to attempt to help the compiler.
10510  * Note that if we return a 0, then we can *not* process
10511  * it and the caller should push the packet into the
10512  * slow-path.
10513  */
10514 static int
10515 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10516     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10517     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10518 {
10519         int32_t acked;
10520         int32_t nsegs;
10521 #ifdef TCPDEBUG
10522         /*
10523          * The size of tcp_saveipgen must be the size of the max ip header,
10524          * now IPv6.
10525          */
10526         u_char tcp_saveipgen[IP6_HDR_LEN];
10527         struct tcphdr tcp_savetcp;
10528         short ostate = 0;
10529 #endif
10530         int32_t under_pacing = 0;
10531         struct tcp_rack *rack;
10532
10533         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10534                 /* Old ack, behind (or duplicate to) the last one rcv'd */
10535                 return (0);
10536         }
10537         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10538                 /* Above what we have sent? */
10539                 return (0);
10540         }
10541         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10542                 /* We are retransmitting */
10543                 return (0);
10544         }
10545         if (__predict_false(tiwin == 0)) {
10546                 /* zero window */
10547                 return (0);
10548         }
10549         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10550                 /* We need a SYN or a FIN, unlikely.. */
10551                 return (0);
10552         }
10553         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10554                 /* Timestamp is behind .. old ack with seq wrap? */
10555                 return (0);
10556         }
10557         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10558                 /* Still recovering */
10559                 return (0);
10560         }
10561         rack = (struct tcp_rack *)tp->t_fb_ptr;
10562         if (rack->r_ctl.rc_sacked) {
10563                 /* We have sack holes on our scoreboard */
10564                 return (0);
10565         }
10566         /* Ok if we reach here, we can process a fast-ack */
10567         if (rack->gp_ready &&
10568             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10569                 under_pacing = 1;
10570         }
10571         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10572         rack_log_ack(tp, to, th, 0, 0);
10573         /* Did the window get updated? */
10574         if (tiwin != tp->snd_wnd) {
10575                 tp->snd_wnd = tiwin;
10576                 rack_validate_fo_sendwin_up(tp, rack);
10577                 tp->snd_wl1 = th->th_seq;
10578                 if (tp->snd_wnd > tp->max_sndwnd)
10579                         tp->max_sndwnd = tp->snd_wnd;
10580         }
10581         /* Do we exit persists? */
10582         if ((rack->rc_in_persist != 0) &&
10583             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10584                                rack->r_ctl.rc_pace_min_segs))) {
10585                 rack_exit_persist(tp, rack, cts);
10586         }
10587         /* Do we enter persists? */
10588         if ((rack->rc_in_persist == 0) &&
10589             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10590             TCPS_HAVEESTABLISHED(tp->t_state) &&
10591             (tp->snd_max == tp->snd_una) &&
10592             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10593             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10594                 /*
10595                  * Here the rwnd is less than
10596                  * the pacing size, we are established,
10597                  * nothing is outstanding, and there is
10598                  * data to send. Enter persists.
10599                  */
10600                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10601         }
10602         /*
10603          * If last ACK falls within this segment's sequence numbers, record
10604          * the timestamp. NOTE that the test is modified according to the
10605          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10606          */
10607         if ((to->to_flags & TOF_TS) != 0 &&
10608             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10609                 tp->ts_recent_age = tcp_ts_getticks();
10610                 tp->ts_recent = to->to_tsval;
10611         }
10612         /*
10613          * This is a pure ack for outstanding data.
10614          */
10615         KMOD_TCPSTAT_INC(tcps_predack);
10616
10617         /*
10618          * "bad retransmit" recovery.
10619          */
10620         if ((tp->t_flags & TF_PREVVALID) &&
10621             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10622                 tp->t_flags &= ~TF_PREVVALID;
10623                 if (tp->t_rxtshift == 1 &&
10624                     (int)(ticks - tp->t_badrxtwin) < 0)
10625                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10626         }
10627         /*
10628          * Recalculate the transmit timer / rtt.
10629          *
10630          * Some boxes send broken timestamp replies during the SYN+ACK
10631          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10632          * and blow up the retransmit timer.
10633          */
10634         acked = BYTES_THIS_ACK(tp, th);
10635
10636 #ifdef TCP_HHOOK
10637         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10638         hhook_run_tcp_est_in(tp, th, to);
10639 #endif
10640         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10641         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10642         if (acked) {
10643                 struct mbuf *mfree;
10644
10645                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
10646                 SOCKBUF_LOCK(&so->so_snd);
10647                 mfree = sbcut_locked(&so->so_snd, acked);
10648                 tp->snd_una = th->th_ack;
10649                 /* Note we want to hold the sb lock through the sendmap adjust */
10650                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10651                 /* Wake up the socket if we have room to write more */
10652                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10653                 sowwakeup_locked(so);
10654                 m_freem(mfree);
10655                 tp->t_rxtshift = 0;
10656                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10657                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10658                 rack->rc_tlp_in_progress = 0;
10659                 rack->r_ctl.rc_tlp_cnt_out = 0;
10660                 /*
10661                  * If it is the RXT timer we want to
10662                  * stop it, so we can restart a TLP.
10663                  */
10664                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10665                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10666 #ifdef NETFLIX_HTTP_LOGGING
10667                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10668 #endif
10669         }
10670         /*
10671          * Let the congestion control algorithm update congestion control
10672          * related information. This typically means increasing the
10673          * congestion window.
10674          */
10675         if (tp->snd_wnd < ctf_outstanding(tp)) {
10676                 /* The peer collapsed the window */
10677                 rack_collapsed_window(rack);
10678         } else if (rack->rc_has_collapsed)
10679                 rack_un_collapse_window(rack);
10680
10681         /*
10682          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
10683          */
10684         tp->snd_wl2 = th->th_ack;
10685         tp->t_dupacks = 0;
10686         m_freem(m);
10687         /* ND6_HINT(tp);         *//* Some progress has been made. */
10688
10689         /*
10690          * If all outstanding data are acked, stop retransmit timer,
10691          * otherwise restart timer using current (possibly backed-off)
10692          * value. If process is waiting for space, wakeup/selwakeup/signal.
10693          * If data are ready to send, let tcp_output decide between more
10694          * output or persist.
10695          */
10696 #ifdef TCPDEBUG
10697         if (so->so_options & SO_DEBUG)
10698                 tcp_trace(TA_INPUT, ostate, tp,
10699                     (void *)tcp_saveipgen,
10700                     &tcp_savetcp, 0);
10701 #endif
10702         if (under_pacing &&
10703             (rack->use_fixed_rate == 0) &&
10704             (rack->in_probe_rtt == 0) &&
10705             rack->rc_gp_dyn_mul &&
10706             rack->rc_always_pace) {
10707                 /* Check if we are dragging bottom */
10708                 rack_check_bottom_drag(tp, rack, so, acked);
10709         }
10710         if (tp->snd_una == tp->snd_max) {
10711                 tp->t_flags &= ~TF_PREVVALID;
10712                 rack->r_ctl.retran_during_recovery = 0;
10713                 rack->r_ctl.dsack_byte_cnt = 0;
10714                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10715                 if (rack->r_ctl.rc_went_idle_time == 0)
10716                         rack->r_ctl.rc_went_idle_time = 1;
10717                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10718                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10719                         tp->t_acktime = 0;
10720                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10721         }
10722         if (acked && rack->r_fast_output)
10723                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
10724         if (sbavail(&so->so_snd)) {
10725                 rack->r_wanted_output = 1;
10726         }
10727         return (1);
10728 }
10729
10730 /*
10731  * Return value of 1, the TCB is unlocked and most
10732  * likely gone, return value of 0, the TCP is still
10733  * locked.
10734  */
10735 static int
10736 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
10737     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10738     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10739 {
10740         int32_t ret_val = 0;
10741         int32_t todrop;
10742         int32_t ourfinisacked = 0;
10743         struct tcp_rack *rack;
10744
10745         ctf_calc_rwin(so, tp);
10746         /*
10747          * If the state is SYN_SENT: if seg contains an ACK, but not for our
10748          * SYN, drop the input. if seg contains a RST, then drop the
10749          * connection. if seg does not contain SYN, then drop it. Otherwise
10750          * this is an acceptable SYN segment initialize tp->rcv_nxt and
10751          * tp->irs if seg contains ack then advance tp->snd_una if seg
10752          * contains an ECE and ECN support is enabled, the stream is ECN
10753          * capable. if SYN has been acked change to ESTABLISHED else
10754          * SYN_RCVD state arrange for segment to be acked (eventually)
10755          * continue processing rest of data/controls.
10756          */
10757         if ((thflags & TH_ACK) &&
10758             (SEQ_LEQ(th->th_ack, tp->iss) ||
10759             SEQ_GT(th->th_ack, tp->snd_max))) {
10760                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10761                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10762                 return (1);
10763         }
10764         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
10765                 TCP_PROBE5(connect__refused, NULL, tp,
10766                     mtod(m, const char *), tp, th);
10767                 tp = tcp_drop(tp, ECONNREFUSED);
10768                 ctf_do_drop(m, tp);
10769                 return (1);
10770         }
10771         if (thflags & TH_RST) {
10772                 ctf_do_drop(m, tp);
10773                 return (1);
10774         }
10775         if (!(thflags & TH_SYN)) {
10776                 ctf_do_drop(m, tp);
10777                 return (1);
10778         }
10779         tp->irs = th->th_seq;
10780         tcp_rcvseqinit(tp);
10781         rack = (struct tcp_rack *)tp->t_fb_ptr;
10782         if (thflags & TH_ACK) {
10783                 int tfo_partial = 0;
10784
10785                 KMOD_TCPSTAT_INC(tcps_connects);
10786                 soisconnected(so);
10787 #ifdef MAC
10788                 mac_socketpeer_set_from_mbuf(m, so);
10789 #endif
10790                 /* Do window scaling on this connection? */
10791                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10792                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10793                         tp->rcv_scale = tp->request_r_scale;
10794                 }
10795                 tp->rcv_adv += min(tp->rcv_wnd,
10796                     TCP_MAXWIN << tp->rcv_scale);
10797                 /*
10798                  * If not all the data that was sent in the TFO SYN
10799                  * has been acked, resend the remainder right away.
10800                  */
10801                 if (IS_FASTOPEN(tp->t_flags) &&
10802                     (tp->snd_una != tp->snd_max)) {
10803                         tp->snd_nxt = th->th_ack;
10804                         tfo_partial = 1;
10805                 }
10806                 /*
10807                  * If there's data, delay ACK; if there's also a FIN ACKNOW
10808                  * will be turned on later.
10809                  */
10810                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
10811                         rack_timer_cancel(tp, rack,
10812                                           rack->r_ctl.rc_rcvtime, __LINE__);
10813                         tp->t_flags |= TF_DELACK;
10814                 } else {
10815                         rack->r_wanted_output = 1;
10816                         tp->t_flags |= TF_ACKNOW;
10817                         rack->rc_dack_toggle = 0;
10818                 }
10819                 if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
10820                     (V_tcp_do_ecn == 1)) {
10821                         tp->t_flags2 |= TF2_ECN_PERMIT;
10822                         KMOD_TCPSTAT_INC(tcps_ecn_shs);
10823                 }
10824                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
10825                         /*
10826                          * We advance snd_una for the
10827                          * fast open case. If th_ack is
10828                          * acknowledging data beyond
10829                          * snd_una we can't just call
10830                          * ack-processing since the
10831                          * data stream in our send-map
10832                          * will start at snd_una + 1 (one
10833                          * beyond the SYN). If its just
10834                          * equal we don't need to do that
10835                          * and there is no send_map.
10836                          */
10837                         tp->snd_una++;
10838                 }
10839                 /*
10840                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
10841                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
10842                  */
10843                 tp->t_starttime = ticks;
10844                 if (tp->t_flags & TF_NEEDFIN) {
10845                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
10846                         tp->t_flags &= ~TF_NEEDFIN;
10847                         thflags &= ~TH_SYN;
10848                 } else {
10849                         tcp_state_change(tp, TCPS_ESTABLISHED);
10850                         TCP_PROBE5(connect__established, NULL, tp,
10851                             mtod(m, const char *), tp, th);
10852                         rack_cc_conn_init(tp);
10853                 }
10854         } else {
10855                 /*
10856                  * Received initial SYN in SYN-SENT[*] state => simultaneous
10857                  * open.  If segment contains CC option and there is a
10858                  * cached CC, apply TAO test. If it succeeds, connection is *
10859                  * half-synchronized. Otherwise, do 3-way handshake:
10860                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
10861                  * there was no CC option, clear cached CC value.
10862                  */
10863                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
10864                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
10865         }
10866         INP_WLOCK_ASSERT(tp->t_inpcb);
10867         /*
10868          * Advance th->th_seq to correspond to first data byte. If data,
10869          * trim to stay within window, dropping FIN if necessary.
10870          */
10871         th->th_seq++;
10872         if (tlen > tp->rcv_wnd) {
10873                 todrop = tlen - tp->rcv_wnd;
10874                 m_adj(m, -todrop);
10875                 tlen = tp->rcv_wnd;
10876                 thflags &= ~TH_FIN;
10877                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
10878                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
10879         }
10880         tp->snd_wl1 = th->th_seq - 1;
10881         tp->rcv_up = th->th_seq;
10882         /*
10883          * Client side of transaction: already sent SYN and data. If the
10884          * remote host used T/TCP to validate the SYN, our data will be
10885          * ACK'd; if so, enter normal data segment processing in the middle
10886          * of step 5, ack processing. Otherwise, goto step 6.
10887          */
10888         if (thflags & TH_ACK) {
10889                 /* For syn-sent we need to possibly update the rtt */
10890                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
10891                         uint32_t t, mcts;
10892
10893                         mcts = tcp_ts_getticks();
10894                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
10895                         if (!tp->t_rttlow || tp->t_rttlow > t)
10896                                 tp->t_rttlow = t;
10897                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
10898                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
10899                         tcp_rack_xmit_timer_commit(rack, tp);
10900                 }
10901                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
10902                         return (ret_val);
10903                 /* We may have changed to FIN_WAIT_1 above */
10904                 if (tp->t_state == TCPS_FIN_WAIT_1) {
10905                         /*
10906                          * In FIN_WAIT_1 STATE in addition to the processing
10907                          * for the ESTABLISHED state if our FIN is now
10908                          * acknowledged then enter FIN_WAIT_2.
10909                          */
10910                         if (ourfinisacked) {
10911                                 /*
10912                                  * If we can't receive any more data, then
10913                                  * closing user can proceed. Starting the
10914                                  * timer is contrary to the specification,
10915                                  * but if we don't get a FIN we'll hang
10916                                  * forever.
10917                                  *
10918                                  * XXXjl: we should release the tp also, and
10919                                  * use a compressed state.
10920                                  */
10921                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10922                                         soisdisconnected(so);
10923                                         tcp_timer_activate(tp, TT_2MSL,
10924                                             (tcp_fast_finwait2_recycle ?
10925                                             tcp_finwait2_timeout :
10926                                             TP_MAXIDLE(tp)));
10927                                 }
10928                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
10929                         }
10930                 }
10931         }
10932         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10933            tiwin, thflags, nxt_pkt));
10934 }
10935
10936 /*
10937  * Return value of 1, the TCB is unlocked and most
10938  * likely gone, return value of 0, the TCP is still
10939  * locked.
10940  */
10941 static int
10942 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
10943     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10944     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10945 {
10946         struct tcp_rack *rack;
10947         int32_t ret_val = 0;
10948         int32_t ourfinisacked = 0;
10949
10950         ctf_calc_rwin(so, tp);
10951         if ((thflags & TH_ACK) &&
10952             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
10953             SEQ_GT(th->th_ack, tp->snd_max))) {
10954                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10955                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10956                 return (1);
10957         }
10958         rack = (struct tcp_rack *)tp->t_fb_ptr;
10959         if (IS_FASTOPEN(tp->t_flags)) {
10960                 /*
10961                  * When a TFO connection is in SYN_RECEIVED, the
10962                  * only valid packets are the initial SYN, a
10963                  * retransmit/copy of the initial SYN (possibly with
10964                  * a subset of the original data), a valid ACK, a
10965                  * FIN, or a RST.
10966                  */
10967                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
10968                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10969                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10970                         return (1);
10971                 } else if (thflags & TH_SYN) {
10972                         /* non-initial SYN is ignored */
10973                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
10974                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
10975                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
10976                                 ctf_do_drop(m, NULL);
10977                                 return (0);
10978                         }
10979                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
10980                         ctf_do_drop(m, NULL);
10981                         return (0);
10982                 }
10983         }
10984         if ((thflags & TH_RST) ||
10985             (tp->t_fin_is_rst && (thflags & TH_FIN)))
10986                 return (ctf_process_rst(m, th, so, tp));
10987         /*
10988          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10989          * it's less than ts_recent, drop it.
10990          */
10991         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10992             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10993                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10994                         return (ret_val);
10995         }
10996         /*
10997          * In the SYN-RECEIVED state, validate that the packet belongs to
10998          * this connection before trimming the data to fit the receive
10999          * window.  Check the sequence number versus IRS since we know the
11000          * sequence numbers haven't wrapped.  This is a partial fix for the
11001          * "LAND" DoS attack.
11002          */
11003         if (SEQ_LT(th->th_seq, tp->irs)) {
11004                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11005                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11006                 return (1);
11007         }
11008         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11009                               &rack->r_ctl.challenge_ack_ts,
11010                               &rack->r_ctl.challenge_ack_cnt)) {
11011                 return (ret_val);
11012         }
11013         /*
11014          * If last ACK falls within this segment's sequence numbers, record
11015          * its timestamp. NOTE: 1) That the test incorporates suggestions
11016          * from the latest proposal of the tcplw@cray.com list (Braden
11017          * 1993/04/26). 2) That updating only on newer timestamps interferes
11018          * with our earlier PAWS tests, so this check should be solely
11019          * predicated on the sequence space of this segment. 3) That we
11020          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11021          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11022          * SEG.Len, This modified check allows us to overcome RFC1323's
11023          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11024          * p.869. In such cases, we can still calculate the RTT correctly
11025          * when RCV.NXT == Last.ACK.Sent.
11026          */
11027         if ((to->to_flags & TOF_TS) != 0 &&
11028             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11029             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11030             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11031                 tp->ts_recent_age = tcp_ts_getticks();
11032                 tp->ts_recent = to->to_tsval;
11033         }
11034         tp->snd_wnd = tiwin;
11035         rack_validate_fo_sendwin_up(tp, rack);
11036         /*
11037          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11038          * is on (half-synchronized state), then queue data for later
11039          * processing; else drop segment and return.
11040          */
11041         if ((thflags & TH_ACK) == 0) {
11042                 if (IS_FASTOPEN(tp->t_flags)) {
11043                         rack_cc_conn_init(tp);
11044                 }
11045                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11046                     tiwin, thflags, nxt_pkt));
11047         }
11048         KMOD_TCPSTAT_INC(tcps_connects);
11049         soisconnected(so);
11050         /* Do window scaling? */
11051         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11052             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11053                 tp->rcv_scale = tp->request_r_scale;
11054         }
11055         /*
11056          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11057          * FIN-WAIT-1
11058          */
11059         tp->t_starttime = ticks;
11060         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11061                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11062                 tp->t_tfo_pending = NULL;
11063         }
11064         if (tp->t_flags & TF_NEEDFIN) {
11065                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11066                 tp->t_flags &= ~TF_NEEDFIN;
11067         } else {
11068                 tcp_state_change(tp, TCPS_ESTABLISHED);
11069                 TCP_PROBE5(accept__established, NULL, tp,
11070                     mtod(m, const char *), tp, th);
11071                 /*
11072                  * TFO connections call cc_conn_init() during SYN
11073                  * processing.  Calling it again here for such connections
11074                  * is not harmless as it would undo the snd_cwnd reduction
11075                  * that occurs when a TFO SYN|ACK is retransmitted.
11076                  */
11077                 if (!IS_FASTOPEN(tp->t_flags))
11078                         rack_cc_conn_init(tp);
11079         }
11080         /*
11081          * Account for the ACK of our SYN prior to
11082          * regular ACK processing below, except for
11083          * simultaneous SYN, which is handled later.
11084          */
11085         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11086                 tp->snd_una++;
11087         /*
11088          * If segment contains data or ACK, will call tcp_reass() later; if
11089          * not, do so now to pass queued data to user.
11090          */
11091         if (tlen == 0 && (thflags & TH_FIN) == 0) {
11092                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11093                     (struct mbuf *)0);
11094                 if (tp->t_flags & TF_WAKESOR) {
11095                         tp->t_flags &= ~TF_WAKESOR;
11096                         /* NB: sorwakeup_locked() does an implicit unlock. */
11097                         sorwakeup_locked(so);
11098                 }
11099         }
11100         tp->snd_wl1 = th->th_seq - 1;
11101         /* For syn-recv we need to possibly update the rtt */
11102         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11103                 uint32_t t, mcts;
11104
11105                 mcts = tcp_ts_getticks();
11106                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11107                 if (!tp->t_rttlow || tp->t_rttlow > t)
11108                         tp->t_rttlow = t;
11109                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11110                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11111                 tcp_rack_xmit_timer_commit(rack, tp);
11112         }
11113         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11114                 return (ret_val);
11115         }
11116         if (tp->t_state == TCPS_FIN_WAIT_1) {
11117                 /* We could have went to FIN_WAIT_1 (or EST) above */
11118                 /*
11119                  * In FIN_WAIT_1 STATE in addition to the processing for the
11120                  * ESTABLISHED state if our FIN is now acknowledged then
11121                  * enter FIN_WAIT_2.
11122                  */
11123                 if (ourfinisacked) {
11124                         /*
11125                          * If we can't receive any more data, then closing
11126                          * user can proceed. Starting the timer is contrary
11127                          * to the specification, but if we don't get a FIN
11128                          * we'll hang forever.
11129                          *
11130                          * XXXjl: we should release the tp also, and use a
11131                          * compressed state.
11132                          */
11133                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11134                                 soisdisconnected(so);
11135                                 tcp_timer_activate(tp, TT_2MSL,
11136                                     (tcp_fast_finwait2_recycle ?
11137                                     tcp_finwait2_timeout :
11138                                     TP_MAXIDLE(tp)));
11139                         }
11140                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11141                 }
11142         }
11143         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11144             tiwin, thflags, nxt_pkt));
11145 }
11146
11147 /*
11148  * Return value of 1, the TCB is unlocked and most
11149  * likely gone, return value of 0, the TCP is still
11150  * locked.
11151  */
11152 static int
11153 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11154     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11155     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11156 {
11157         int32_t ret_val = 0;
11158         struct tcp_rack *rack;
11159
11160         /*
11161          * Header prediction: check for the two common cases of a
11162          * uni-directional data xfer.  If the packet has no control flags,
11163          * is in-sequence, the window didn't change and we're not
11164          * retransmitting, it's a candidate.  If the length is zero and the
11165          * ack moved forward, we're the sender side of the xfer.  Just free
11166          * the data acked & wake any higher level process that was blocked
11167          * waiting for space.  If the length is non-zero and the ack didn't
11168          * move, we're the receiver side.  If we're getting packets in-order
11169          * (the reassembly queue is empty), add the data toc The socket
11170          * buffer and note that we need a delayed ack. Make sure that the
11171          * hidden state-flags are also off. Since we check for
11172          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11173          */
11174         rack = (struct tcp_rack *)tp->t_fb_ptr;
11175         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11176             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11177             __predict_true(SEGQ_EMPTY(tp)) &&
11178             __predict_true(th->th_seq == tp->rcv_nxt)) {
11179                 if (tlen == 0) {
11180                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11181                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11182                                 return (0);
11183                         }
11184                 } else {
11185                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11186                             tiwin, nxt_pkt, iptos)) {
11187                                 return (0);
11188                         }
11189                 }
11190         }
11191         ctf_calc_rwin(so, tp);
11192
11193         if ((thflags & TH_RST) ||
11194             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11195                 return (ctf_process_rst(m, th, so, tp));
11196
11197         /*
11198          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11199          * synchronized state.
11200          */
11201         if (thflags & TH_SYN) {
11202                 ctf_challenge_ack(m, th, tp, &ret_val);
11203                 return (ret_val);
11204         }
11205         /*
11206          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11207          * it's less than ts_recent, drop it.
11208          */
11209         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11210             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11211                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11212                         return (ret_val);
11213         }
11214         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11215                               &rack->r_ctl.challenge_ack_ts,
11216                               &rack->r_ctl.challenge_ack_cnt)) {
11217                 return (ret_val);
11218         }
11219         /*
11220          * If last ACK falls within this segment's sequence numbers, record
11221          * its timestamp. NOTE: 1) That the test incorporates suggestions
11222          * from the latest proposal of the tcplw@cray.com list (Braden
11223          * 1993/04/26). 2) That updating only on newer timestamps interferes
11224          * with our earlier PAWS tests, so this check should be solely
11225          * predicated on the sequence space of this segment. 3) That we
11226          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11227          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11228          * SEG.Len, This modified check allows us to overcome RFC1323's
11229          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11230          * p.869. In such cases, we can still calculate the RTT correctly
11231          * when RCV.NXT == Last.ACK.Sent.
11232          */
11233         if ((to->to_flags & TOF_TS) != 0 &&
11234             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11235             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11236             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11237                 tp->ts_recent_age = tcp_ts_getticks();
11238                 tp->ts_recent = to->to_tsval;
11239         }
11240         /*
11241          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11242          * is on (half-synchronized state), then queue data for later
11243          * processing; else drop segment and return.
11244          */
11245         if ((thflags & TH_ACK) == 0) {
11246                 if (tp->t_flags & TF_NEEDSYN) {
11247                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11248                             tiwin, thflags, nxt_pkt));
11249
11250                 } else if (tp->t_flags & TF_ACKNOW) {
11251                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11252                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11253                         return (ret_val);
11254                 } else {
11255                         ctf_do_drop(m, NULL);
11256                         return (0);
11257                 }
11258         }
11259         /*
11260          * Ack processing.
11261          */
11262         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11263                 return (ret_val);
11264         }
11265         if (sbavail(&so->so_snd)) {
11266                 if (ctf_progress_timeout_check(tp, true)) {
11267                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11268                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11269                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11270                         return (1);
11271                 }
11272         }
11273         /* State changes only happen in rack_process_data() */
11274         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11275             tiwin, thflags, nxt_pkt));
11276 }
11277
11278 /*
11279  * Return value of 1, the TCB is unlocked and most
11280  * likely gone, return value of 0, the TCP is still
11281  * locked.
11282  */
11283 static int
11284 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11285     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11286     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11287 {
11288         int32_t ret_val = 0;
11289         struct tcp_rack *rack;
11290
11291         rack = (struct tcp_rack *)tp->t_fb_ptr;
11292         ctf_calc_rwin(so, tp);
11293         if ((thflags & TH_RST) ||
11294             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11295                 return (ctf_process_rst(m, th, so, tp));
11296         /*
11297          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11298          * synchronized state.
11299          */
11300         if (thflags & TH_SYN) {
11301                 ctf_challenge_ack(m, th, tp, &ret_val);
11302                 return (ret_val);
11303         }
11304         /*
11305          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11306          * it's less than ts_recent, drop it.
11307          */
11308         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11309             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11310                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11311                         return (ret_val);
11312         }
11313         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11314                               &rack->r_ctl.challenge_ack_ts,
11315                               &rack->r_ctl.challenge_ack_cnt)) {
11316                 return (ret_val);
11317         }
11318         /*
11319          * If last ACK falls within this segment's sequence numbers, record
11320          * its timestamp. NOTE: 1) That the test incorporates suggestions
11321          * from the latest proposal of the tcplw@cray.com list (Braden
11322          * 1993/04/26). 2) That updating only on newer timestamps interferes
11323          * with our earlier PAWS tests, so this check should be solely
11324          * predicated on the sequence space of this segment. 3) That we
11325          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11326          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11327          * SEG.Len, This modified check allows us to overcome RFC1323's
11328          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11329          * p.869. In such cases, we can still calculate the RTT correctly
11330          * when RCV.NXT == Last.ACK.Sent.
11331          */
11332         if ((to->to_flags & TOF_TS) != 0 &&
11333             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11334             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11335             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11336                 tp->ts_recent_age = tcp_ts_getticks();
11337                 tp->ts_recent = to->to_tsval;
11338         }
11339         /*
11340          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11341          * is on (half-synchronized state), then queue data for later
11342          * processing; else drop segment and return.
11343          */
11344         if ((thflags & TH_ACK) == 0) {
11345                 if (tp->t_flags & TF_NEEDSYN) {
11346                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11347                             tiwin, thflags, nxt_pkt));
11348
11349                 } else if (tp->t_flags & TF_ACKNOW) {
11350                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11351                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11352                         return (ret_val);
11353                 } else {
11354                         ctf_do_drop(m, NULL);
11355                         return (0);
11356                 }
11357         }
11358         /*
11359          * Ack processing.
11360          */
11361         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11362                 return (ret_val);
11363         }
11364         if (sbavail(&so->so_snd)) {
11365                 if (ctf_progress_timeout_check(tp, true)) {
11366                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11367                                                 tp, tick, PROGRESS_DROP, __LINE__);
11368                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11369                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11370                         return (1);
11371                 }
11372         }
11373         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11374             tiwin, thflags, nxt_pkt));
11375 }
11376
11377 static int
11378 rack_check_data_after_close(struct mbuf *m,
11379     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11380 {
11381         struct tcp_rack *rack;
11382
11383         rack = (struct tcp_rack *)tp->t_fb_ptr;
11384         if (rack->rc_allow_data_af_clo == 0) {
11385         close_now:
11386                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11387                 /* tcp_close will kill the inp pre-log the Reset */
11388                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11389                 tp = tcp_close(tp);
11390                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11391                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11392                 return (1);
11393         }
11394         if (sbavail(&so->so_snd) == 0)
11395                 goto close_now;
11396         /* Ok we allow data that is ignored and a followup reset */
11397         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11398         tp->rcv_nxt = th->th_seq + *tlen;
11399         tp->t_flags2 |= TF2_DROP_AF_DATA;
11400         rack->r_wanted_output = 1;
11401         *tlen = 0;
11402         return (0);
11403 }
11404
11405 /*
11406  * Return value of 1, the TCB is unlocked and most
11407  * likely gone, return value of 0, the TCP is still
11408  * locked.
11409  */
11410 static int
11411 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11412     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11413     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11414 {
11415         int32_t ret_val = 0;
11416         int32_t ourfinisacked = 0;
11417         struct tcp_rack *rack;
11418
11419         rack = (struct tcp_rack *)tp->t_fb_ptr;
11420         ctf_calc_rwin(so, tp);
11421
11422         if ((thflags & TH_RST) ||
11423             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11424                 return (ctf_process_rst(m, th, so, tp));
11425         /*
11426          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11427          * synchronized state.
11428          */
11429         if (thflags & TH_SYN) {
11430                 ctf_challenge_ack(m, th, tp, &ret_val);
11431                 return (ret_val);
11432         }
11433         /*
11434          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11435          * it's less than ts_recent, drop it.
11436          */
11437         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11438             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11439                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11440                         return (ret_val);
11441         }
11442         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11443                               &rack->r_ctl.challenge_ack_ts,
11444                               &rack->r_ctl.challenge_ack_cnt)) {
11445                 return (ret_val);
11446         }
11447         /*
11448          * If new data are received on a connection after the user processes
11449          * are gone, then RST the other end.
11450          */
11451         if ((so->so_state & SS_NOFDREF) && tlen) {
11452                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11453                         return (1);
11454         }
11455         /*
11456          * If last ACK falls within this segment's sequence numbers, record
11457          * its timestamp. NOTE: 1) That the test incorporates suggestions
11458          * from the latest proposal of the tcplw@cray.com list (Braden
11459          * 1993/04/26). 2) That updating only on newer timestamps interferes
11460          * with our earlier PAWS tests, so this check should be solely
11461          * predicated on the sequence space of this segment. 3) That we
11462          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11463          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11464          * SEG.Len, This modified check allows us to overcome RFC1323's
11465          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11466          * p.869. In such cases, we can still calculate the RTT correctly
11467          * when RCV.NXT == Last.ACK.Sent.
11468          */
11469         if ((to->to_flags & TOF_TS) != 0 &&
11470             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11471             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11472             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11473                 tp->ts_recent_age = tcp_ts_getticks();
11474                 tp->ts_recent = to->to_tsval;
11475         }
11476         /*
11477          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11478          * is on (half-synchronized state), then queue data for later
11479          * processing; else drop segment and return.
11480          */
11481         if ((thflags & TH_ACK) == 0) {
11482                 if (tp->t_flags & TF_NEEDSYN) {
11483                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11484                             tiwin, thflags, nxt_pkt));
11485                 } else if (tp->t_flags & TF_ACKNOW) {
11486                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11487                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11488                         return (ret_val);
11489                 } else {
11490                         ctf_do_drop(m, NULL);
11491                         return (0);
11492                 }
11493         }
11494         /*
11495          * Ack processing.
11496          */
11497         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11498                 return (ret_val);
11499         }
11500         if (ourfinisacked) {
11501                 /*
11502                  * If we can't receive any more data, then closing user can
11503                  * proceed. Starting the timer is contrary to the
11504                  * specification, but if we don't get a FIN we'll hang
11505                  * forever.
11506                  *
11507                  * XXXjl: we should release the tp also, and use a
11508                  * compressed state.
11509                  */
11510                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11511                         soisdisconnected(so);
11512                         tcp_timer_activate(tp, TT_2MSL,
11513                             (tcp_fast_finwait2_recycle ?
11514                             tcp_finwait2_timeout :
11515                             TP_MAXIDLE(tp)));
11516                 }
11517                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11518         }
11519         if (sbavail(&so->so_snd)) {
11520                 if (ctf_progress_timeout_check(tp, true)) {
11521                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11522                                                 tp, tick, PROGRESS_DROP, __LINE__);
11523                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11524                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11525                         return (1);
11526                 }
11527         }
11528         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11529             tiwin, thflags, nxt_pkt));
11530 }
11531
11532 /*
11533  * Return value of 1, the TCB is unlocked and most
11534  * likely gone, return value of 0, the TCP is still
11535  * locked.
11536  */
11537 static int
11538 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11539     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11540     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11541 {
11542         int32_t ret_val = 0;
11543         int32_t ourfinisacked = 0;
11544         struct tcp_rack *rack;
11545
11546         rack = (struct tcp_rack *)tp->t_fb_ptr;
11547         ctf_calc_rwin(so, tp);
11548
11549         if ((thflags & TH_RST) ||
11550             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11551                 return (ctf_process_rst(m, th, so, tp));
11552         /*
11553          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11554          * synchronized state.
11555          */
11556         if (thflags & TH_SYN) {
11557                 ctf_challenge_ack(m, th, tp, &ret_val);
11558                 return (ret_val);
11559         }
11560         /*
11561          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11562          * it's less than ts_recent, drop it.
11563          */
11564         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11565             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11566                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11567                         return (ret_val);
11568         }
11569         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11570                               &rack->r_ctl.challenge_ack_ts,
11571                               &rack->r_ctl.challenge_ack_cnt)) {
11572                 return (ret_val);
11573         }
11574         /*
11575          * If new data are received on a connection after the user processes
11576          * are gone, then RST the other end.
11577          */
11578         if ((so->so_state & SS_NOFDREF) && tlen) {
11579                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11580                         return (1);
11581         }
11582         /*
11583          * If last ACK falls within this segment's sequence numbers, record
11584          * its timestamp. NOTE: 1) That the test incorporates suggestions
11585          * from the latest proposal of the tcplw@cray.com list (Braden
11586          * 1993/04/26). 2) That updating only on newer timestamps interferes
11587          * with our earlier PAWS tests, so this check should be solely
11588          * predicated on the sequence space of this segment. 3) That we
11589          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11590          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11591          * SEG.Len, This modified check allows us to overcome RFC1323's
11592          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11593          * p.869. In such cases, we can still calculate the RTT correctly
11594          * when RCV.NXT == Last.ACK.Sent.
11595          */
11596         if ((to->to_flags & TOF_TS) != 0 &&
11597             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11598             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11599             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11600                 tp->ts_recent_age = tcp_ts_getticks();
11601                 tp->ts_recent = to->to_tsval;
11602         }
11603         /*
11604          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11605          * is on (half-synchronized state), then queue data for later
11606          * processing; else drop segment and return.
11607          */
11608         if ((thflags & TH_ACK) == 0) {
11609                 if (tp->t_flags & TF_NEEDSYN) {
11610                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11611                             tiwin, thflags, nxt_pkt));
11612                 } else if (tp->t_flags & TF_ACKNOW) {
11613                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11614                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11615                         return (ret_val);
11616                 } else {
11617                         ctf_do_drop(m, NULL);
11618                         return (0);
11619                 }
11620         }
11621         /*
11622          * Ack processing.
11623          */
11624         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11625                 return (ret_val);
11626         }
11627         if (ourfinisacked) {
11628                 tcp_twstart(tp);
11629                 m_freem(m);
11630                 return (1);
11631         }
11632         if (sbavail(&so->so_snd)) {
11633                 if (ctf_progress_timeout_check(tp, true)) {
11634                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11635                                                 tp, tick, PROGRESS_DROP, __LINE__);
11636                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11637                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11638                         return (1);
11639                 }
11640         }
11641         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11642             tiwin, thflags, nxt_pkt));
11643 }
11644
11645 /*
11646  * Return value of 1, the TCB is unlocked and most
11647  * likely gone, return value of 0, the TCP is still
11648  * locked.
11649  */
11650 static int
11651 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11652     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11653     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11654 {
11655         int32_t ret_val = 0;
11656         int32_t ourfinisacked = 0;
11657         struct tcp_rack *rack;
11658
11659         rack = (struct tcp_rack *)tp->t_fb_ptr;
11660         ctf_calc_rwin(so, tp);
11661
11662         if ((thflags & TH_RST) ||
11663             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11664                 return (ctf_process_rst(m, th, so, tp));
11665         /*
11666          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11667          * synchronized state.
11668          */
11669         if (thflags & TH_SYN) {
11670                 ctf_challenge_ack(m, th, tp, &ret_val);
11671                 return (ret_val);
11672         }
11673         /*
11674          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11675          * it's less than ts_recent, drop it.
11676          */
11677         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11678             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11679                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11680                         return (ret_val);
11681         }
11682         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11683                               &rack->r_ctl.challenge_ack_ts,
11684                               &rack->r_ctl.challenge_ack_cnt)) {
11685                 return (ret_val);
11686         }
11687         /*
11688          * If new data are received on a connection after the user processes
11689          * are gone, then RST the other end.
11690          */
11691         if ((so->so_state & SS_NOFDREF) && tlen) {
11692                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11693                         return (1);
11694         }
11695         /*
11696          * If last ACK falls within this segment's sequence numbers, record
11697          * its timestamp. NOTE: 1) That the test incorporates suggestions
11698          * from the latest proposal of the tcplw@cray.com list (Braden
11699          * 1993/04/26). 2) That updating only on newer timestamps interferes
11700          * with our earlier PAWS tests, so this check should be solely
11701          * predicated on the sequence space of this segment. 3) That we
11702          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11703          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11704          * SEG.Len, This modified check allows us to overcome RFC1323's
11705          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11706          * p.869. In such cases, we can still calculate the RTT correctly
11707          * when RCV.NXT == Last.ACK.Sent.
11708          */
11709         if ((to->to_flags & TOF_TS) != 0 &&
11710             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11711             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11712             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11713                 tp->ts_recent_age = tcp_ts_getticks();
11714                 tp->ts_recent = to->to_tsval;
11715         }
11716         /*
11717          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11718          * is on (half-synchronized state), then queue data for later
11719          * processing; else drop segment and return.
11720          */
11721         if ((thflags & TH_ACK) == 0) {
11722                 if (tp->t_flags & TF_NEEDSYN) {
11723                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11724                             tiwin, thflags, nxt_pkt));
11725                 } else if (tp->t_flags & TF_ACKNOW) {
11726                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11727                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11728                         return (ret_val);
11729                 } else {
11730                         ctf_do_drop(m, NULL);
11731                         return (0);
11732                 }
11733         }
11734         /*
11735          * case TCPS_LAST_ACK: Ack processing.
11736          */
11737         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11738                 return (ret_val);
11739         }
11740         if (ourfinisacked) {
11741                 tp = tcp_close(tp);
11742                 ctf_do_drop(m, tp);
11743                 return (1);
11744         }
11745         if (sbavail(&so->so_snd)) {
11746                 if (ctf_progress_timeout_check(tp, true)) {
11747                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11748                                                 tp, tick, PROGRESS_DROP, __LINE__);
11749                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11750                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11751                         return (1);
11752                 }
11753         }
11754         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11755             tiwin, thflags, nxt_pkt));
11756 }
11757
11758 /*
11759  * Return value of 1, the TCB is unlocked and most
11760  * likely gone, return value of 0, the TCP is still
11761  * locked.
11762  */
11763 static int
11764 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
11765     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11766     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11767 {
11768         int32_t ret_val = 0;
11769         int32_t ourfinisacked = 0;
11770         struct tcp_rack *rack;
11771
11772         rack = (struct tcp_rack *)tp->t_fb_ptr;
11773         ctf_calc_rwin(so, tp);
11774
11775         /* Reset receive buffer auto scaling when not in bulk receive mode. */
11776         if ((thflags & TH_RST) ||
11777             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11778                 return (ctf_process_rst(m, th, so, tp));
11779         /*
11780          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11781          * synchronized state.
11782          */
11783         if (thflags & TH_SYN) {
11784                 ctf_challenge_ack(m, th, tp, &ret_val);
11785                 return (ret_val);
11786         }
11787         /*
11788          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11789          * it's less than ts_recent, drop it.
11790          */
11791         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11792             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11793                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11794                         return (ret_val);
11795         }
11796         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11797                               &rack->r_ctl.challenge_ack_ts,
11798                               &rack->r_ctl.challenge_ack_cnt)) {
11799                 return (ret_val);
11800         }
11801         /*
11802          * If new data are received on a connection after the user processes
11803          * are gone, then RST the other end.
11804          */
11805         if ((so->so_state & SS_NOFDREF) &&
11806             tlen) {
11807                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11808                         return (1);
11809         }
11810         /*
11811          * If last ACK falls within this segment's sequence numbers, record
11812          * its timestamp. NOTE: 1) That the test incorporates suggestions
11813          * from the latest proposal of the tcplw@cray.com list (Braden
11814          * 1993/04/26). 2) That updating only on newer timestamps interferes
11815          * with our earlier PAWS tests, so this check should be solely
11816          * predicated on the sequence space of this segment. 3) That we
11817          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11818          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11819          * SEG.Len, This modified check allows us to overcome RFC1323's
11820          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11821          * p.869. In such cases, we can still calculate the RTT correctly
11822          * when RCV.NXT == Last.ACK.Sent.
11823          */
11824         if ((to->to_flags & TOF_TS) != 0 &&
11825             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11826             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11827             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11828                 tp->ts_recent_age = tcp_ts_getticks();
11829                 tp->ts_recent = to->to_tsval;
11830         }
11831         /*
11832          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11833          * is on (half-synchronized state), then queue data for later
11834          * processing; else drop segment and return.
11835          */
11836         if ((thflags & TH_ACK) == 0) {
11837                 if (tp->t_flags & TF_NEEDSYN) {
11838                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11839                             tiwin, thflags, nxt_pkt));
11840                 } else if (tp->t_flags & TF_ACKNOW) {
11841                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11842                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11843                         return (ret_val);
11844                 } else {
11845                         ctf_do_drop(m, NULL);
11846                         return (0);
11847                 }
11848         }
11849         /*
11850          * Ack processing.
11851          */
11852         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11853                 return (ret_val);
11854         }
11855         if (sbavail(&so->so_snd)) {
11856                 if (ctf_progress_timeout_check(tp, true)) {
11857                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11858                                                 tp, tick, PROGRESS_DROP, __LINE__);
11859                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11860                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11861                         return (1);
11862                 }
11863         }
11864         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11865             tiwin, thflags, nxt_pkt));
11866 }
11867
11868 static void inline
11869 rack_clear_rate_sample(struct tcp_rack *rack)
11870 {
11871         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
11872         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
11873         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
11874 }
11875
11876 static void
11877 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
11878 {
11879         uint64_t bw_est, rate_wanted;
11880         int chged = 0;
11881         uint32_t user_max, orig_min, orig_max;
11882
11883         orig_min = rack->r_ctl.rc_pace_min_segs;
11884         orig_max = rack->r_ctl.rc_pace_max_segs;
11885         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
11886         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
11887                 chged = 1;
11888         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
11889         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
11890                 if (user_max != rack->r_ctl.rc_pace_max_segs)
11891                         chged = 1;
11892         }
11893         if (rack->rc_force_max_seg) {
11894                 rack->r_ctl.rc_pace_max_segs = user_max;
11895         } else if (rack->use_fixed_rate) {
11896                 bw_est = rack_get_bw(rack);
11897                 if ((rack->r_ctl.crte == NULL) ||
11898                     (bw_est != rack->r_ctl.crte->rate)) {
11899                         rack->r_ctl.rc_pace_max_segs = user_max;
11900                 } else {
11901                         /* We are pacing right at the hardware rate */
11902                         uint32_t segsiz;
11903
11904                         segsiz = min(ctf_fixed_maxseg(tp),
11905                                      rack->r_ctl.rc_pace_min_segs);
11906                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
11907                                                            tp, bw_est, segsiz, 0,
11908                                                            rack->r_ctl.crte, NULL);
11909                 }
11910         } else if (rack->rc_always_pace) {
11911                 if (rack->r_ctl.gp_bw ||
11912 #ifdef NETFLIX_PEAKRATE
11913                     rack->rc_tp->t_maxpeakrate ||
11914 #endif
11915                     rack->r_ctl.init_rate) {
11916                         /* We have a rate of some sort set */
11917                         uint32_t  orig;
11918
11919                         bw_est = rack_get_bw(rack);
11920                         orig = rack->r_ctl.rc_pace_max_segs;
11921                         if (fill_override)
11922                                 rate_wanted = *fill_override;
11923                         else
11924                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
11925                         if (rate_wanted) {
11926                                 /* We have something */
11927                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
11928                                                                                    rate_wanted,
11929                                                                                    ctf_fixed_maxseg(rack->rc_tp));
11930                         } else
11931                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
11932                         if (orig != rack->r_ctl.rc_pace_max_segs)
11933                                 chged = 1;
11934                 } else if ((rack->r_ctl.gp_bw == 0) &&
11935                            (rack->r_ctl.rc_pace_max_segs == 0)) {
11936                         /*
11937                          * If we have nothing limit us to bursting
11938                          * out IW sized pieces.
11939                          */
11940                         chged = 1;
11941                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
11942                 }
11943         }
11944         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
11945                 chged = 1;
11946                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
11947         }
11948         if (chged)
11949                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
11950 }
11951
11952
11953 static void
11954 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
11955 {
11956 #ifdef INET6
11957         struct ip6_hdr *ip6 = NULL;
11958 #endif
11959 #ifdef INET
11960         struct ip *ip = NULL;
11961 #endif
11962         struct udphdr *udp = NULL;
11963
11964         /* Ok lets fill in the fast block, it can only be used with no IP options! */
11965 #ifdef INET6
11966         if (rack->r_is_v6) {
11967                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
11968                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
11969                 if (tp->t_port) {
11970                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11971                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
11972                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11973                         udp->uh_dport = tp->t_port;
11974                         rack->r_ctl.fsb.udp = udp;
11975                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11976                 } else
11977                 {
11978                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
11979                         rack->r_ctl.fsb.udp = NULL;
11980                 }
11981                 tcpip_fillheaders(rack->rc_inp,
11982                                   tp->t_port,
11983                                   ip6, rack->r_ctl.fsb.th);
11984         } else
11985 #endif                          /* INET6 */
11986         {
11987                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
11988                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
11989                 if (tp->t_port) {
11990                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11991                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
11992                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11993                         udp->uh_dport = tp->t_port;
11994                         rack->r_ctl.fsb.udp = udp;
11995                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11996                 } else
11997                 {
11998                         rack->r_ctl.fsb.udp = NULL;
11999                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12000                 }
12001                 tcpip_fillheaders(rack->rc_inp,
12002                                   tp->t_port,
12003                                   ip, rack->r_ctl.fsb.th);
12004         }
12005         rack->r_fsb_inited = 1;
12006 }
12007
12008 static int
12009 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12010 {
12011         /*
12012          * Allocate the larger of spaces V6 if available else just
12013          * V4 and include udphdr (overbook)
12014          */
12015 #ifdef INET6
12016         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12017 #else
12018         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12019 #endif
12020         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12021                                             M_TCPFSB, M_NOWAIT|M_ZERO);
12022         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12023                 return (ENOMEM);
12024         }
12025         rack->r_fsb_inited = 0;
12026         return (0);
12027 }
12028
12029 static int
12030 rack_init(struct tcpcb *tp)
12031 {
12032         struct tcp_rack *rack = NULL;
12033         struct rack_sendmap *insret;
12034         uint32_t iwin, snt, us_cts;
12035         int err;
12036
12037         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12038         if (tp->t_fb_ptr == NULL) {
12039                 /*
12040                  * We need to allocate memory but cant. The INP and INP_INFO
12041                  * locks and they are recusive (happens during setup. So a
12042                  * scheme to drop the locks fails :(
12043                  *
12044                  */
12045                 return (ENOMEM);
12046         }
12047         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12048
12049         rack = (struct tcp_rack *)tp->t_fb_ptr;
12050         RB_INIT(&rack->r_ctl.rc_mtree);
12051         TAILQ_INIT(&rack->r_ctl.rc_free);
12052         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12053         rack->rc_tp = tp;
12054         rack->rc_inp = tp->t_inpcb;
12055         /* Set the flag */
12056         rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12057         /* Probably not needed but lets be sure */
12058         rack_clear_rate_sample(rack);
12059         /*
12060          * Save off the default values, socket options will poke
12061          * at these if pacing is not on or we have not yet
12062          * reached where pacing is on (gp_ready/fixed enabled).
12063          * When they get set into the CC module (when gp_ready
12064          * is enabled or we enable fixed) then we will set these
12065          * values into the CC and place in here the old values
12066          * so we have a restoral. Then we will set the flag
12067          * rc_pacing_cc_set. That way whenever we turn off pacing
12068          * or switch off this stack, we will know to go restore
12069          * the saved values.
12070          */
12071         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12072         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12073         /* We want abe like behavior as well */
12074         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
12075         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12076         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12077         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12078         if (use_rack_rr)
12079                 rack->use_rack_rr = 1;
12080         if (V_tcp_delack_enabled)
12081                 tp->t_delayed_ack = 1;
12082         else
12083                 tp->t_delayed_ack = 0;
12084 #ifdef TCP_ACCOUNTING
12085         if (rack_tcp_accounting) {
12086                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12087         }
12088 #endif
12089         if (rack_enable_shared_cwnd)
12090                 rack->rack_enable_scwnd = 1;
12091         rack->rc_user_set_max_segs = rack_hptsi_segments;
12092         rack->rc_force_max_seg = 0;
12093         if (rack_use_imac_dack)
12094                 rack->rc_dack_mode = 1;
12095         TAILQ_INIT(&rack->r_ctl.opt_list);
12096         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12097         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12098         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12099         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12100         rack->r_ctl.rc_highest_us_rtt = 0;
12101         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12102         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12103         if (rack_use_cmp_acks)
12104                 rack->r_use_cmp_ack = 1;
12105         if (rack_disable_prr)
12106                 rack->rack_no_prr = 1;
12107         if (rack_gp_no_rec_chg)
12108                 rack->rc_gp_no_rec_chg = 1;
12109         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12110                 rack->rc_always_pace = 1;
12111                 if (rack->use_fixed_rate || rack->gp_ready)
12112                         rack_set_cc_pacing(rack);
12113         } else
12114                 rack->rc_always_pace = 0;
12115         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12116                 rack->r_mbuf_queue = 1;
12117         else
12118                 rack->r_mbuf_queue = 0;
12119         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12120                 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12121         else
12122                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12123         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12124         if (rack_limits_scwnd)
12125                 rack->r_limit_scw = 1;
12126         else
12127                 rack->r_limit_scw = 0;
12128         rack->rc_labc = V_tcp_abc_l_var;
12129         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12130         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12131         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12132         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12133         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12134         rack->r_ctl.rc_min_to = rack_min_to;
12135         microuptime(&rack->r_ctl.act_rcv_time);
12136         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12137         rack->r_running_late = 0;
12138         rack->r_running_early = 0;
12139         rack->rc_init_win = rack_default_init_window;
12140         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12141         if (rack_hw_up_only)
12142                 rack->r_up_only = 1;
12143         if (rack_do_dyn_mul) {
12144                 /* When dynamic adjustment is on CA needs to start at 100% */
12145                 rack->rc_gp_dyn_mul = 1;
12146                 if (rack_do_dyn_mul >= 100)
12147                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12148         } else
12149                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12150         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12151         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12152         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12153         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12154                                 rack_probertt_filter_life);
12155         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12156         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12157         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12158         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12159         rack->r_ctl.rc_time_probertt_starts = 0;
12160         /* We require at least one measurement, even if the sysctl is 0 */
12161         if (rack_req_measurements)
12162                 rack->r_ctl.req_measurements = rack_req_measurements;
12163         else
12164                 rack->r_ctl.req_measurements = 1;
12165         if (rack_enable_hw_pacing)
12166                 rack->rack_hdw_pace_ena = 1;
12167         if (rack_hw_rate_caps)
12168                 rack->r_rack_hw_rate_caps = 1;
12169         /* Do we force on detection? */
12170 #ifdef NETFLIX_EXP_DETECTION
12171         if (tcp_force_detection)
12172                 rack->do_detection = 1;
12173         else
12174 #endif
12175                 rack->do_detection = 0;
12176         if (rack_non_rxt_use_cr)
12177                 rack->rack_rec_nonrxt_use_cr = 1;
12178         err = rack_init_fsb(tp, rack);
12179         if (err) {
12180                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12181                 tp->t_fb_ptr = NULL;
12182                 return (err);
12183         }
12184         if (tp->snd_una != tp->snd_max) {
12185                 /* Create a send map for the current outstanding data */
12186                 struct rack_sendmap *rsm;
12187
12188                 rsm = rack_alloc(rack);
12189                 if (rsm == NULL) {
12190                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12191                         tp->t_fb_ptr = NULL;
12192                         return (ENOMEM);
12193                 }
12194                 rsm->r_no_rtt_allowed = 1;
12195                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12196                 rsm->r_rtr_cnt = 1;
12197                 rsm->r_rtr_bytes = 0;
12198                 if (tp->t_flags & TF_SENTFIN) {
12199                         rsm->r_end = tp->snd_max - 1;
12200                         rsm->r_flags |= RACK_HAS_FIN;
12201                 } else {
12202                         rsm->r_end = tp->snd_max;
12203                 }
12204                 if (tp->snd_una == tp->iss) {
12205                         /* The data space is one beyond snd_una */
12206                         rsm->r_flags |= RACK_HAS_SYN;
12207                         rsm->r_start = tp->iss;
12208                         rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12209                 } else
12210                         rsm->r_start = tp->snd_una;
12211                 rsm->r_dupack = 0;
12212                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12213                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12214                         if (rsm->m)
12215                                 rsm->orig_m_len = rsm->m->m_len;
12216                         else
12217                                 rsm->orig_m_len = 0;
12218                 } else {
12219                         /*
12220                          * This can happen if we have a stand-alone FIN or
12221                          *  SYN.
12222                          */
12223                         rsm->m = NULL;
12224                         rsm->orig_m_len = 0;
12225                         rsm->soff = 0;
12226                 }
12227                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12228 #ifdef INVARIANTS
12229                 if (insret != NULL) {
12230                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12231                               insret, rack, rsm);
12232                 }
12233 #endif
12234                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12235                 rsm->r_in_tmap = 1;
12236         }
12237         /*
12238          * Timers in Rack are kept in microseconds so lets
12239          * convert any initial incoming variables
12240          * from ticks into usecs. Note that we
12241          * also change the values of t_srtt and t_rttvar, if
12242          * they are non-zero. They are kept with a 5
12243          * bit decimal so we have to carefully convert
12244          * these to get the full precision.
12245          */
12246         rack_convert_rtts(tp);
12247         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12248         if (rack_def_profile)
12249                 rack_set_profile(rack, rack_def_profile);
12250         /* Cancel the GP measurement in progress */
12251         tp->t_flags &= ~TF_GPUTINPROG;
12252         if (SEQ_GT(tp->snd_max, tp->iss))
12253                 snt = tp->snd_max - tp->iss;
12254         else
12255                 snt = 0;
12256         iwin = rc_init_window(rack);
12257         if (snt < iwin) {
12258                 /* We are not past the initial window
12259                  * so we need to make sure cwnd is
12260                  * correct.
12261                  */
12262                 if (tp->snd_cwnd < iwin)
12263                         tp->snd_cwnd = iwin;
12264                 /*
12265                  * If we are within the initial window
12266                  * we want ssthresh to be unlimited. Setting
12267                  * it to the rwnd (which the default stack does
12268                  * and older racks) is not really a good idea
12269                  * since we want to be in SS and grow both the
12270                  * cwnd and the rwnd (via dynamic rwnd growth). If
12271                  * we set it to the rwnd then as the peer grows its
12272                  * rwnd we will be stuck in CA and never hit SS.
12273                  *
12274                  * Its far better to raise it up high (this takes the
12275                  * risk that there as been a loss already, probably
12276                  * we should have an indicator in all stacks of loss
12277                  * but we don't), but considering the normal use this
12278                  * is a risk worth taking. The consequences of not
12279                  * hitting SS are far worse than going one more time
12280                  * into it early on (before we have sent even a IW).
12281                  * It is highly unlikely that we will have had a loss
12282                  * before getting the IW out.
12283                  */
12284                 tp->snd_ssthresh = 0xffffffff;
12285         }
12286         rack_stop_all_timers(tp);
12287         /* Lets setup the fsb block */
12288         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12289         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12290                              __LINE__, RACK_RTTS_INIT);
12291         return (0);
12292 }
12293
12294 static int
12295 rack_handoff_ok(struct tcpcb *tp)
12296 {
12297         if ((tp->t_state == TCPS_CLOSED) ||
12298             (tp->t_state == TCPS_LISTEN)) {
12299                 /* Sure no problem though it may not stick */
12300                 return (0);
12301         }
12302         if ((tp->t_state == TCPS_SYN_SENT) ||
12303             (tp->t_state == TCPS_SYN_RECEIVED)) {
12304                 /*
12305                  * We really don't know if you support sack,
12306                  * you have to get to ESTAB or beyond to tell.
12307                  */
12308                 return (EAGAIN);
12309         }
12310         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12311                 /*
12312                  * Rack will only send a FIN after all data is acknowledged.
12313                  * So in this case we have more data outstanding. We can't
12314                  * switch stacks until either all data and only the FIN
12315                  * is left (in which case rack_init() now knows how
12316                  * to deal with that) <or> all is acknowledged and we
12317                  * are only left with incoming data, though why you
12318                  * would want to switch to rack after all data is acknowledged
12319                  * I have no idea (rrs)!
12320                  */
12321                 return (EAGAIN);
12322         }
12323         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12324                 return (0);
12325         }
12326         /*
12327          * If we reach here we don't do SACK on this connection so we can
12328          * never do rack.
12329          */
12330         return (EINVAL);
12331 }
12332
12333
12334 static void
12335 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12336 {
12337         int ack_cmp = 0;
12338
12339         if (tp->t_fb_ptr) {
12340                 struct tcp_rack *rack;
12341                 struct rack_sendmap *rsm, *nrsm, *rm;
12342
12343                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12344                 if (tp->t_in_pkt) {
12345                         /*
12346                          * It is unsafe to process the packets since a
12347                          * reset may be lurking in them (its rare but it
12348                          * can occur). If we were to find a RST, then we
12349                          * would end up dropping the connection and the
12350                          * INP lock, so when we return the caller (tcp_usrreq)
12351                          * will blow up when it trys to unlock the inp.
12352                          */
12353                         struct mbuf *save, *m;
12354
12355                         m = tp->t_in_pkt;
12356                         tp->t_in_pkt = NULL;
12357                         tp->t_tail_pkt = NULL;
12358                         while (m) {
12359                                 save = m->m_nextpkt;
12360                                 m->m_nextpkt = NULL;
12361                                 m_freem(m);
12362                                 m = save;
12363                         }
12364                         if ((tp->t_inpcb) &&
12365                             (tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP))
12366                                 ack_cmp = 1;
12367                         if (ack_cmp) {
12368                                 /* Total if we used large or small (if ack-cmp was used). */
12369                                 if (rack->rc_inp->inp_flags2 & INP_MBUF_L_ACKS)
12370                                         counter_u64_add(rack_large_ackcmp, 1);
12371                                 else
12372                                         counter_u64_add(rack_small_ackcmp, 1);
12373                         }
12374                 }
12375                 tp->t_flags &= ~TF_FORCEDATA;
12376 #ifdef NETFLIX_SHARED_CWND
12377                 if (rack->r_ctl.rc_scw) {
12378                         uint32_t limit;
12379
12380                         if (rack->r_limit_scw)
12381                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12382                         else
12383                                 limit = 0;
12384                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12385                                                   rack->r_ctl.rc_scw_index,
12386                                                   limit);
12387                         rack->r_ctl.rc_scw = NULL;
12388                 }
12389 #endif
12390                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12391                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12392                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12393                         rack->r_ctl.fsb.th = NULL;
12394                 }
12395                 /* Convert back to ticks, with  */
12396                 if (tp->t_srtt > 1) {
12397                         uint32_t val, frac;
12398
12399                         val = USEC_2_TICKS(tp->t_srtt);
12400                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12401                         tp->t_srtt = val << TCP_RTT_SHIFT;
12402                         /*
12403                          * frac is the fractional part here is left
12404                          * over from converting to hz and shifting.
12405                          * We need to convert this to the 5 bit
12406                          * remainder.
12407                          */
12408                         if (frac) {
12409                                 if (hz == 1000) {
12410                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12411                                 } else {
12412                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12413                                 }
12414                                 tp->t_srtt += frac;
12415                         }
12416                 }
12417                 if (tp->t_rttvar) {
12418                         uint32_t val, frac;
12419
12420                         val = USEC_2_TICKS(tp->t_rttvar);
12421                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12422                         tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12423                         /*
12424                          * frac is the fractional part here is left
12425                          * over from converting to hz and shifting.
12426                          * We need to convert this to the 5 bit
12427                          * remainder.
12428                          */
12429                         if (frac) {
12430                                 if (hz == 1000) {
12431                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12432                                 } else {
12433                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12434                                 }
12435                                 tp->t_rttvar += frac;
12436                         }
12437                 }
12438                 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12439                 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12440                 if (rack->rc_always_pace) {
12441                         tcp_decrement_paced_conn();
12442                         rack_undo_cc_pacing(rack);
12443                         rack->rc_always_pace = 0;
12444                 }
12445                 /* Clean up any options if they were not applied */
12446                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12447                         struct deferred_opt_list *dol;
12448
12449                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12450                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12451                         free(dol, M_TCPDO);
12452                 }
12453                 /* rack does not use force data but other stacks may clear it */
12454                 if (rack->r_ctl.crte != NULL) {
12455                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12456                         rack->rack_hdrw_pacing = 0;
12457                         rack->r_ctl.crte = NULL;
12458                 }
12459 #ifdef TCP_BLACKBOX
12460                 tcp_log_flowend(tp);
12461 #endif
12462                 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12463                         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12464 #ifdef INVARIANTS
12465                         if (rm != rsm) {
12466                                 panic("At fini, rack:%p rsm:%p rm:%p",
12467                                       rack, rsm, rm);
12468                         }
12469 #endif
12470                         uma_zfree(rack_zone, rsm);
12471                 }
12472                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12473                 while (rsm) {
12474                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12475                         uma_zfree(rack_zone, rsm);
12476                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12477                 }
12478                 rack->rc_free_cnt = 0;
12479                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12480                 tp->t_fb_ptr = NULL;
12481         }
12482         if (tp->t_inpcb) {
12483                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12484                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12485                 tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12486                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
12487                 /* Cancel the GP measurement in progress */
12488                 tp->t_flags &= ~TF_GPUTINPROG;
12489                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
12490         }
12491         /* Make sure snd_nxt is correctly set */
12492         tp->snd_nxt = tp->snd_max;
12493 }
12494
12495 static void
12496 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12497 {
12498         if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12499                 rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12500         }
12501         switch (tp->t_state) {
12502         case TCPS_SYN_SENT:
12503                 rack->r_state = TCPS_SYN_SENT;
12504                 rack->r_substate = rack_do_syn_sent;
12505                 break;
12506         case TCPS_SYN_RECEIVED:
12507                 rack->r_state = TCPS_SYN_RECEIVED;
12508                 rack->r_substate = rack_do_syn_recv;
12509                 break;
12510         case TCPS_ESTABLISHED:
12511                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12512                 rack->r_state = TCPS_ESTABLISHED;
12513                 rack->r_substate = rack_do_established;
12514                 break;
12515         case TCPS_CLOSE_WAIT:
12516                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12517                 rack->r_state = TCPS_CLOSE_WAIT;
12518                 rack->r_substate = rack_do_close_wait;
12519                 break;
12520         case TCPS_FIN_WAIT_1:
12521                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12522                 rack->r_state = TCPS_FIN_WAIT_1;
12523                 rack->r_substate = rack_do_fin_wait_1;
12524                 break;
12525         case TCPS_CLOSING:
12526                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12527                 rack->r_state = TCPS_CLOSING;
12528                 rack->r_substate = rack_do_closing;
12529                 break;
12530         case TCPS_LAST_ACK:
12531                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12532                 rack->r_state = TCPS_LAST_ACK;
12533                 rack->r_substate = rack_do_lastack;
12534                 break;
12535         case TCPS_FIN_WAIT_2:
12536                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12537                 rack->r_state = TCPS_FIN_WAIT_2;
12538                 rack->r_substate = rack_do_fin_wait_2;
12539                 break;
12540         case TCPS_LISTEN:
12541         case TCPS_CLOSED:
12542         case TCPS_TIME_WAIT:
12543         default:
12544                 break;
12545         };
12546         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12547                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12548
12549 }
12550
12551 static void
12552 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12553 {
12554         /*
12555          * We received an ack, and then did not
12556          * call send or were bounced out due to the
12557          * hpts was running. Now a timer is up as well, is
12558          * it the right timer?
12559          */
12560         struct rack_sendmap *rsm;
12561         int tmr_up;
12562
12563         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12564         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12565                 return;
12566         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12567         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12568             (tmr_up == PACE_TMR_RXT)) {
12569                 /* Should be an RXT */
12570                 return;
12571         }
12572         if (rsm == NULL) {
12573                 /* Nothing outstanding? */
12574                 if (tp->t_flags & TF_DELACK) {
12575                         if (tmr_up == PACE_TMR_DELACK)
12576                                 /* We are supposed to have delayed ack up and we do */
12577                                 return;
12578                 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12579                         /*
12580                          * if we hit enobufs then we would expect the possiblity
12581                          * of nothing outstanding and the RXT up (and the hptsi timer).
12582                          */
12583                         return;
12584                 } else if (((V_tcp_always_keepalive ||
12585                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12586                             (tp->t_state <= TCPS_CLOSING)) &&
12587                            (tmr_up == PACE_TMR_KEEP) &&
12588                            (tp->snd_max == tp->snd_una)) {
12589                         /* We should have keep alive up and we do */
12590                         return;
12591                 }
12592         }
12593         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12594                    ((tmr_up == PACE_TMR_TLP) ||
12595                     (tmr_up == PACE_TMR_RACK) ||
12596                     (tmr_up == PACE_TMR_RXT))) {
12597                 /*
12598                  * Either a Rack, TLP or RXT is fine if  we
12599                  * have outstanding data.
12600                  */
12601                 return;
12602         } else if (tmr_up == PACE_TMR_DELACK) {
12603                 /*
12604                  * If the delayed ack was going to go off
12605                  * before the rtx/tlp/rack timer were going to
12606                  * expire, then that would be the timer in control.
12607                  * Note we don't check the time here trusting the
12608                  * code is correct.
12609                  */
12610                 return;
12611         }
12612         /*
12613          * Ok the timer originally started is not what we want now.
12614          * We will force the hpts to be stopped if any, and restart
12615          * with the slot set to what was in the saved slot.
12616          */
12617         if (rack->rc_inp->inp_in_hpts) {
12618                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12619                         uint32_t us_cts;
12620
12621                         us_cts = tcp_get_usecs(NULL);
12622                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12623                                 rack->r_early = 1;
12624                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12625                         }
12626                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12627                 }
12628                 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
12629         }
12630         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12631         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12632 }
12633
12634
12635 static void
12636 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)
12637 {
12638         tp->snd_wnd = tiwin;
12639         rack_validate_fo_sendwin_up(tp, rack);
12640         tp->snd_wl1 = seq;
12641         tp->snd_wl2 = ack;
12642         if (tp->snd_wnd > tp->max_sndwnd)
12643                 tp->max_sndwnd = tp->snd_wnd;
12644         if (tp->snd_wnd < (tp->snd_max - high_seq)) {
12645                 /* The peer collapsed the window */
12646                 rack_collapsed_window(rack);
12647         } else if (rack->rc_has_collapsed)
12648                 rack_un_collapse_window(rack);
12649         /* Do we exit persists? */
12650         if ((rack->rc_in_persist != 0) &&
12651             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12652                                 rack->r_ctl.rc_pace_min_segs))) {
12653                 rack_exit_persist(tp, rack, cts);
12654         }
12655         /* Do we enter persists? */
12656         if ((rack->rc_in_persist == 0) &&
12657             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12658             TCPS_HAVEESTABLISHED(tp->t_state) &&
12659             (tp->snd_max == tp->snd_una) &&
12660             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
12661             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
12662                 /*
12663                  * Here the rwnd is less than
12664                  * the pacing size, we are established,
12665                  * nothing is outstanding, and there is
12666                  * data to send. Enter persists.
12667                  */
12668                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12669         }
12670 }
12671
12672 static void
12673 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
12674 {
12675
12676         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
12677                 union tcp_log_stackspecific log;
12678                 struct timeval ltv;
12679                 char tcp_hdr_buf[60];
12680                 struct tcphdr *th;
12681                 struct timespec ts;
12682                 uint32_t orig_snd_una;
12683                 uint8_t xx = 0;
12684
12685 #ifdef NETFLIX_HTTP_LOGGING
12686                 struct http_sendfile_track *http_req;
12687
12688                 if (SEQ_GT(ae->ack, tp->snd_una)) {
12689                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
12690                 } else {
12691                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
12692                 }
12693 #endif
12694                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
12695                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
12696                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
12697                 if (rack->rack_no_prr == 0)
12698                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
12699                 else
12700                         log.u_bbr.flex1 = 0;
12701                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
12702                 log.u_bbr.use_lt_bw <<= 1;
12703                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
12704                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
12705                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
12706                 log.u_bbr.pkts_out = tp->t_maxseg;
12707                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
12708                 log.u_bbr.flex7 = 1;
12709                 log.u_bbr.lost = ae->flags;
12710                 log.u_bbr.cwnd_gain = ackval;
12711                 log.u_bbr.pacing_gain = 0x2;
12712                 if (ae->flags & TSTMP_HDWR) {
12713                         /* Record the hardware timestamp if present */
12714                         log.u_bbr.flex3 = M_TSTMP;
12715                         ts.tv_sec = ae->timestamp / 1000000000;
12716                         ts.tv_nsec = ae->timestamp % 1000000000;
12717                         ltv.tv_sec = ts.tv_sec;
12718                         ltv.tv_usec = ts.tv_nsec / 1000;
12719                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
12720                 } else if (ae->flags & TSTMP_LRO) {
12721                         /* Record the LRO the arrival timestamp */
12722                         log.u_bbr.flex3 = M_TSTMP_LRO;
12723                         ts.tv_sec = ae->timestamp / 1000000000;
12724                         ts.tv_nsec = ae->timestamp % 1000000000;
12725                         ltv.tv_sec = ts.tv_sec;
12726                         ltv.tv_usec = ts.tv_nsec / 1000;
12727                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
12728                 }
12729                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
12730                 /* Log the rcv time */
12731                 log.u_bbr.delRate = ae->timestamp;
12732 #ifdef NETFLIX_HTTP_LOGGING
12733                 log.u_bbr.applimited = tp->t_http_closed;
12734                 log.u_bbr.applimited <<= 8;
12735                 log.u_bbr.applimited |= tp->t_http_open;
12736                 log.u_bbr.applimited <<= 8;
12737                 log.u_bbr.applimited |= tp->t_http_req;
12738                 if (http_req) {
12739                         /* Copy out any client req info */
12740                         /* seconds */
12741                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
12742                         /* useconds */
12743                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
12744                         log.u_bbr.rttProp = http_req->timestamp;
12745                         log.u_bbr.cur_del_rate = http_req->start;
12746                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
12747                                 log.u_bbr.flex8 |= 1;
12748                         } else {
12749                                 log.u_bbr.flex8 |= 2;
12750                                 log.u_bbr.bw_inuse = http_req->end;
12751                         }
12752                         log.u_bbr.flex6 = http_req->start_seq;
12753                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
12754                                 log.u_bbr.flex8 |= 4;
12755                                 log.u_bbr.epoch = http_req->end_seq;
12756                         }
12757                 }
12758 #endif
12759                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
12760                 th = (struct tcphdr *)tcp_hdr_buf;
12761                 th->th_seq = ae->seq;
12762                 th->th_ack = ae->ack;
12763                 th->th_win = ae->win;
12764                 /* Now fill in the ports */
12765                 th->th_sport = tp->t_inpcb->inp_fport;
12766                 th->th_dport = tp->t_inpcb->inp_lport;
12767                 th->th_flags = ae->flags & 0xff;
12768                 /* Now do we have a timestamp option? */
12769                 if (ae->flags & HAS_TSTMP) {
12770                         u_char *cp;
12771                         uint32_t val;
12772
12773                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
12774                         cp = (u_char *)(th + 1);
12775                         *cp = TCPOPT_NOP;
12776                         cp++;
12777                         *cp = TCPOPT_NOP;
12778                         cp++;
12779                         *cp = TCPOPT_TIMESTAMP;
12780                         cp++;
12781                         *cp = TCPOLEN_TIMESTAMP;
12782                         cp++;
12783                         val = htonl(ae->ts_value);
12784                         bcopy((char *)&val,
12785                               (char *)cp, sizeof(uint32_t));
12786                         val = htonl(ae->ts_echo);
12787                         bcopy((char *)&val,
12788                               (char *)(cp + 4), sizeof(uint32_t));
12789                 } else
12790                         th->th_off = (sizeof(struct tcphdr) >> 2);
12791
12792                 /*
12793                  * For sane logging we need to play a little trick.
12794                  * If the ack were fully processed we would have moved
12795                  * snd_una to high_seq, but since compressed acks are
12796                  * processed in two phases, at this point (logging) snd_una
12797                  * won't be advanced. So we would see multiple acks showing
12798                  * the advancement. We can prevent that by "pretending" that
12799                  * snd_una was advanced and then un-advancing it so that the
12800                  * logging code has the right value for tlb_snd_una.
12801                  */
12802                 if (tp->snd_una != high_seq) {
12803                         orig_snd_una = tp->snd_una;
12804                         tp->snd_una = high_seq;
12805                         xx = 1;
12806                 } else
12807                         xx = 0;
12808                 TCP_LOG_EVENTP(tp, th,
12809                                &tp->t_inpcb->inp_socket->so_rcv,
12810                                &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
12811                                0, &log, true, &ltv);
12812                 if (xx) {
12813                         tp->snd_una = orig_snd_una;
12814                 }
12815         }
12816
12817 }
12818
12819 static int
12820 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
12821 {
12822         /*
12823          * Handle a "special" compressed ack mbuf. Each incoming
12824          * ack has only four possible dispositions:
12825          *
12826          * A) It moves the cum-ack forward
12827          * B) It is behind the cum-ack.
12828          * C) It is a window-update ack.
12829          * D) It is a dup-ack.
12830          *
12831          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
12832          * in the incoming mbuf. We also need to still pay attention
12833          * to nxt_pkt since there may be another packet after this
12834          * one.
12835          */
12836 #ifdef TCP_ACCOUNTING
12837         uint64_t ts_val;
12838         uint64_t rdstc;
12839 #endif
12840         int segsiz;
12841         struct timespec ts;
12842         struct tcp_rack *rack;
12843         struct tcp_ackent *ae;
12844         uint32_t tiwin, us_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
12845         int cnt, i, did_out, ourfinisacked = 0;
12846         int win_up_req = 0;
12847         struct tcpopt to_holder, *to = NULL;
12848         int nsegs = 0;
12849         int under_pacing = 1;
12850         int recovery = 0;
12851         int idx;
12852 #ifdef TCP_ACCOUNTING
12853         sched_pin();
12854 #endif
12855         rack = (struct tcp_rack *)tp->t_fb_ptr;
12856         if (rack->gp_ready &&
12857             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
12858                 under_pacing = 0;
12859         else
12860                 under_pacing = 1;
12861
12862         if (rack->r_state != tp->t_state)
12863                 rack_set_state(tp, rack);
12864         to = &to_holder;
12865         to->to_flags = 0;
12866         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
12867                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
12868         cnt = m->m_len / sizeof(struct tcp_ackent);
12869         idx = cnt / 5;
12870         if (idx >= MAX_NUM_OF_CNTS)
12871                 idx = MAX_NUM_OF_CNTS - 1;
12872         counter_u64_add(rack_proc_comp_ack[idx], 1);
12873         counter_u64_add(rack_multi_single_eq, cnt);
12874         high_seq = tp->snd_una;
12875         the_win = tp->snd_wnd;
12876         win_seq = tp->snd_wl1;
12877         win_upd_ack = tp->snd_wl2;
12878         cts = us_cts = tcp_tv_to_usectick(tv);
12879         segsiz = ctf_fixed_maxseg(tp);
12880         if ((rack->rc_gp_dyn_mul) &&
12881             (rack->use_fixed_rate == 0) &&
12882             (rack->rc_always_pace)) {
12883                 /* Check in on probertt */
12884                 rack_check_probe_rtt(rack, us_cts);
12885         }
12886         for (i = 0; i < cnt; i++) {
12887 #ifdef TCP_ACCOUNTING
12888                 ts_val = get_cyclecount();
12889 #endif
12890                 rack_clear_rate_sample(rack);
12891                 ae = ((mtod(m, struct tcp_ackent *)) + i);
12892                 /* Setup the window */
12893                 tiwin = ae->win << tp->snd_scale;
12894                 /* figure out the type of ack */
12895                 if (SEQ_LT(ae->ack, high_seq)) {
12896                         /* Case B*/
12897                         ae->ack_val_set = ACK_BEHIND;
12898                 } else if (SEQ_GT(ae->ack, high_seq)) {
12899                         /* Case A */
12900                         ae->ack_val_set = ACK_CUMACK;
12901                 } else if (tiwin == the_win) {
12902                         /* Case D */
12903                         ae->ack_val_set = ACK_DUPACK;
12904                 } else {
12905                         /* Case C */
12906                         ae->ack_val_set = ACK_RWND;
12907                 }
12908                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
12909                 /* Validate timestamp */
12910                 if (ae->flags & HAS_TSTMP) {
12911                         /* Setup for a timestamp */
12912                         to->to_flags = TOF_TS;
12913                         ae->ts_echo -= tp->ts_offset;
12914                         to->to_tsecr = ae->ts_echo;
12915                         to->to_tsval = ae->ts_value;
12916                         /*
12917                          * If echoed timestamp is later than the current time, fall back to
12918                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
12919                          * were used when this connection was established.
12920                          */
12921                         if (TSTMP_GT(ae->ts_echo, cts))
12922                                 ae->ts_echo = 0;
12923                         if (tp->ts_recent &&
12924                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
12925                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
12926 #ifdef TCP_ACCOUNTING
12927                                         rdstc = get_cyclecount();
12928                                         if (rdstc > ts_val) {
12929                                                 counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
12930                                                                 (rdstc - ts_val));
12931                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12932                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
12933                                                 }
12934                                         }
12935 #endif
12936                                         continue;
12937                                 }
12938                         }
12939                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
12940                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
12941                                 tp->ts_recent_age = tcp_ts_getticks();
12942                                 tp->ts_recent = ae->ts_value;
12943                         }
12944                 } else {
12945                         /* Setup for a no options */
12946                         to->to_flags = 0;
12947                 }
12948                 /* Update the rcv time and perform idle reduction possibly */
12949                 if  (tp->t_idle_reduce &&
12950                      (tp->snd_max == tp->snd_una) &&
12951                      ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
12952                         counter_u64_add(rack_input_idle_reduces, 1);
12953                         rack_cc_after_idle(rack, tp);
12954                 }
12955                 tp->t_rcvtime = ticks;
12956                 /* Now what about ECN? */
12957                 if (tp->t_flags2 & TF2_ECN_PERMIT) {
12958                         if (ae->flags & TH_CWR) {
12959                                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
12960                                 tp->t_flags |= TF_ACKNOW;
12961                         }
12962                         switch (ae->codepoint & IPTOS_ECN_MASK) {
12963                         case IPTOS_ECN_CE:
12964                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
12965                                 KMOD_TCPSTAT_INC(tcps_ecn_ce);
12966                                 break;
12967                         case IPTOS_ECN_ECT0:
12968                                 KMOD_TCPSTAT_INC(tcps_ecn_ect0);
12969                                 break;
12970                         case IPTOS_ECN_ECT1:
12971                                 KMOD_TCPSTAT_INC(tcps_ecn_ect1);
12972                                 break;
12973                         }
12974
12975                         /* Process a packet differently from RFC3168. */
12976                         cc_ecnpkt_handler_flags(tp, ae->flags, ae->codepoint);
12977                         /* Congestion experienced. */
12978                         if (ae->flags & TH_ECE) {
12979                                 rack_cong_signal(tp,  CC_ECN, ae->ack);
12980                         }
12981                 }
12982 #ifdef TCP_ACCOUNTING
12983                 /* Count for the specific type of ack in */
12984                 counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
12985                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12986                         tp->tcp_cnt_counters[ae->ack_val_set]++;
12987                 }
12988 #endif
12989                 /*
12990                  * Note how we could move up these in the determination
12991                  * above, but we don't so that way the timestamp checks (and ECN)
12992                  * is done first before we do any processing on the ACK.
12993                  * The non-compressed path through the code has this
12994                  * weakness (noted by @jtl) that it actually does some
12995                  * processing before verifying the timestamp information.
12996                  * We don't take that path here which is why we set
12997                  * the ack_val_set first, do the timestamp and ecn
12998                  * processing, and then look at what we have setup.
12999                  */
13000                 if (ae->ack_val_set == ACK_BEHIND) {
13001                         /*
13002                          * Case B flag reordering, if window is not closed
13003                          * or it could be a keep-alive or persists
13004                          */
13005                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13006                                 counter_u64_add(rack_reorder_seen, 1);
13007                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13008                         }
13009                 } else if (ae->ack_val_set == ACK_DUPACK) {
13010                         /* Case D */
13011
13012                         rack_strike_dupack(rack);
13013                 } else if (ae->ack_val_set == ACK_RWND) {
13014                         /* Case C */
13015
13016                         win_up_req = 1;
13017                         win_upd_ack = ae->ack;
13018                         win_seq = ae->seq;
13019                         the_win = tiwin;
13020                 } else {
13021                         /* Case A */
13022
13023                         if (SEQ_GT(ae->ack, tp->snd_max)) {
13024                                 /*
13025                                  * We just send an ack since the incoming
13026                                  * ack is beyond the largest seq we sent.
13027                                  */
13028                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
13029                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13030                                         if (tp->t_flags && TF_ACKNOW)
13031                                                 rack->r_wanted_output = 1;
13032                                 }
13033                         } else {
13034                                 nsegs++;
13035                                 /* If the window changed setup to update */
13036                                 if (tiwin != tp->snd_wnd) {
13037                                         win_up_req = 1;
13038                                         win_upd_ack = ae->ack;
13039                                         win_seq = ae->seq;
13040                                         the_win = tiwin;
13041                                 }
13042 #ifdef TCP_ACCOUNTING
13043                                 /* Account for the acks */
13044                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13045                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13046                                 }
13047                                 counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13048                                                 (((ae->ack - high_seq) + segsiz - 1) / segsiz));
13049 #endif
13050                                 high_seq = ae->ack;
13051                                 /* Setup our act_rcv_time */
13052                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13053                                         ts.tv_sec = ae->timestamp / 1000000000;
13054                                         ts.tv_nsec = ae->timestamp % 1000000000;
13055                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13056                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13057                                 } else {
13058                                         rack->r_ctl.act_rcv_time = *tv;
13059                                 }
13060                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13061                         }
13062                 }
13063                 /* And lets be sure to commit the rtt measurements for this ack */
13064                 tcp_rack_xmit_timer_commit(rack, tp);
13065 #ifdef TCP_ACCOUNTING
13066                 rdstc = get_cyclecount();
13067                 if (rdstc > ts_val) {
13068                         counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13069                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13070                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13071                                 if (ae->ack_val_set == ACK_CUMACK)
13072                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13073                         }
13074                 }
13075 #endif
13076         }
13077 #ifdef TCP_ACCOUNTING
13078         ts_val = get_cyclecount();
13079 #endif
13080         acked_amount = acked = (high_seq - tp->snd_una);
13081         if (win_up_req) {
13082                 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13083         }
13084         if (acked) {
13085                 if (rack->sack_attack_disable == 0)
13086                         rack_do_decay(rack);
13087                 if (acked >= segsiz) {
13088                         /*
13089                          * You only get credit for
13090                          * MSS and greater (and you get extra
13091                          * credit for larger cum-ack moves).
13092                          */
13093                         int ac;
13094
13095                         ac = acked / segsiz;
13096                         rack->r_ctl.ack_count += ac;
13097                         counter_u64_add(rack_ack_total, ac);
13098                 }
13099                 if (rack->r_ctl.ack_count > 0xfff00000) {
13100                         /*
13101                          * reduce the number to keep us under
13102                          * a uint32_t.
13103                          */
13104                         rack->r_ctl.ack_count /= 2;
13105                         rack->r_ctl.sack_count /= 2;
13106                 }
13107                 if (tp->t_flags & TF_NEEDSYN) {
13108                         /*
13109                          * T/TCP: Connection was half-synchronized, and our SYN has
13110                          * been ACK'd (so connection is now fully synchronized).  Go
13111                          * to non-starred state, increment snd_una for ACK of SYN,
13112                          * and check if we can do window scaling.
13113                          */
13114                         tp->t_flags &= ~TF_NEEDSYN;
13115                         tp->snd_una++;
13116                         acked_amount = acked = (high_seq - tp->snd_una);
13117                 }
13118                 if (acked > sbavail(&so->so_snd))
13119                         acked_amount = sbavail(&so->so_snd);
13120 #ifdef NETFLIX_EXP_DETECTION
13121                 /*
13122                  * We only care on a cum-ack move if we are in a sack-disabled
13123                  * state. We have already added in to the ack_count, and we never
13124                  * would disable on a cum-ack move, so we only care to do the
13125                  * detection if it may "undo" it, i.e. we were in disabled already.
13126                  */
13127                 if (rack->sack_attack_disable)
13128                         rack_do_detection(tp, rack, acked_amount, segsiz);
13129 #endif
13130                 if (IN_FASTRECOVERY(tp->t_flags) &&
13131                     (rack->rack_no_prr == 0))
13132                         rack_update_prr(tp, rack, acked_amount, high_seq);
13133                 if (IN_RECOVERY(tp->t_flags)) {
13134                         if (SEQ_LT(high_seq, tp->snd_recover) &&
13135                             (SEQ_LT(high_seq, tp->snd_max))) {
13136                                 tcp_rack_partialack(tp);
13137                         } else {
13138                                 rack_post_recovery(tp, high_seq);
13139                                 recovery = 1;
13140                         }
13141                 }
13142                 /* Handle the rack-log-ack part (sendmap) */
13143                 if ((sbused(&so->so_snd) == 0) &&
13144                     (acked > acked_amount) &&
13145                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
13146                     (tp->t_flags & TF_SENTFIN)) {
13147                         /*
13148                          * We must be sure our fin
13149                          * was sent and acked (we can be
13150                          * in FIN_WAIT_1 without having
13151                          * sent the fin).
13152                          */
13153                         ourfinisacked = 1;
13154                         /*
13155                          * Lets make sure snd_una is updated
13156                          * since most likely acked_amount = 0 (it
13157                          * should be).
13158                          */
13159                         tp->snd_una = high_seq;
13160                 }
13161                 /* Did we make a RTO error? */
13162                 if ((tp->t_flags & TF_PREVVALID) &&
13163                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13164                         tp->t_flags &= ~TF_PREVVALID;
13165                         if (tp->t_rxtshift == 1 &&
13166                             (int)(ticks - tp->t_badrxtwin) < 0)
13167                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq);
13168                 }
13169                 /* Handle the data in the socket buffer */
13170                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13171                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13172                 if (acked_amount > 0) {
13173                         struct mbuf *mfree;
13174
13175                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13176                         SOCKBUF_LOCK(&so->so_snd);
13177                         mfree = sbcut_locked(&so->so_snd, acked);
13178                         tp->snd_una = high_seq;
13179                         /* Note we want to hold the sb lock through the sendmap adjust */
13180                         rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13181                         /* Wake up the socket if we have room to write more */
13182                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13183                         sowwakeup_locked(so);
13184                         m_freem(mfree);
13185                 }
13186                 /* update progress */
13187                 tp->t_acktime = ticks;
13188                 rack_log_progress_event(rack, tp, tp->t_acktime,
13189                                         PROGRESS_UPDATE, __LINE__);
13190                 /* Clear out shifts and such */
13191                 tp->t_rxtshift = 0;
13192                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13193                                    rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13194                 rack->rc_tlp_in_progress = 0;
13195                 rack->r_ctl.rc_tlp_cnt_out = 0;
13196                 /* Send recover and snd_nxt must be dragged along */
13197                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13198                         tp->snd_recover = tp->snd_una;
13199                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13200                         tp->snd_nxt = tp->snd_una;
13201                 /*
13202                  * If the RXT timer is running we want to
13203                  * stop it, so we can restart a TLP (or new RXT).
13204                  */
13205                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13206                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13207 #ifdef NETFLIX_HTTP_LOGGING
13208                 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13209 #endif
13210                 tp->snd_wl2 = high_seq;
13211                 tp->t_dupacks = 0;
13212                 if (under_pacing &&
13213                     (rack->use_fixed_rate == 0) &&
13214                     (rack->in_probe_rtt == 0) &&
13215                     rack->rc_gp_dyn_mul &&
13216                     rack->rc_always_pace) {
13217                         /* Check if we are dragging bottom */
13218                         rack_check_bottom_drag(tp, rack, so, acked);
13219                 }
13220                 if (tp->snd_una == tp->snd_max) {
13221                         tp->t_flags &= ~TF_PREVVALID;
13222                         rack->r_ctl.retran_during_recovery = 0;
13223                         rack->r_ctl.dsack_byte_cnt = 0;
13224                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13225                         if (rack->r_ctl.rc_went_idle_time == 0)
13226                                 rack->r_ctl.rc_went_idle_time = 1;
13227                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13228                         if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13229                                 tp->t_acktime = 0;
13230                         /* Set so we might enter persists... */
13231                         rack->r_wanted_output = 1;
13232                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13233                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13234                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13235                             (sbavail(&so->so_snd) == 0) &&
13236                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13237                                 /*
13238                                  * The socket was gone and the
13239                                  * peer sent data (not now in the past), time to
13240                                  * reset him.
13241                                  */
13242                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13243                                 /* tcp_close will kill the inp pre-log the Reset */
13244                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13245 #ifdef TCP_ACCOUNTING
13246                                 rdstc = get_cyclecount();
13247                                 if (rdstc > ts_val) {
13248                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13249                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13250                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13251                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13252                                         }
13253                                 }
13254 #endif
13255                                 m_freem(m);
13256                                 tp = tcp_close(tp);
13257                                 if (tp == NULL) {
13258 #ifdef TCP_ACCOUNTING
13259                                         sched_unpin();
13260 #endif
13261                                         return (1);
13262                                 }
13263                                 /*
13264                                  * We would normally do drop-with-reset which would
13265                                  * send back a reset. We can't since we don't have
13266                                  * all the needed bits. Instead lets arrange for
13267                                  * a call to tcp_output(). That way since we
13268                                  * are in the closed state we will generate a reset.
13269                                  *
13270                                  * Note if tcp_accounting is on we don't unpin since
13271                                  * we do that after the goto label.
13272                                  */
13273                                 goto send_out_a_rst;
13274                         }
13275                         if ((sbused(&so->so_snd) == 0) &&
13276                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
13277                             (tp->t_flags & TF_SENTFIN)) {
13278                                 /*
13279                                  * If we can't receive any more data, then closing user can
13280                                  * proceed. Starting the timer is contrary to the
13281                                  * specification, but if we don't get a FIN we'll hang
13282                                  * forever.
13283                                  *
13284                                  */
13285                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13286                                         soisdisconnected(so);
13287                                         tcp_timer_activate(tp, TT_2MSL,
13288                                                            (tcp_fast_finwait2_recycle ?
13289                                                             tcp_finwait2_timeout :
13290                                                             TP_MAXIDLE(tp)));
13291                                 }
13292                                 if (ourfinisacked == 0) {
13293                                         /*
13294                                          * We don't change to fin-wait-2 if we have our fin acked
13295                                          * which means we are probably in TCPS_CLOSING.
13296                                          */
13297                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13298                                 }
13299                         }
13300                 }
13301                 /* Wake up the socket if we have room to write more */
13302                 if (sbavail(&so->so_snd)) {
13303                         rack->r_wanted_output = 1;
13304                         if (ctf_progress_timeout_check(tp, true)) {
13305                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13306                                                         tp, tick, PROGRESS_DROP, __LINE__);
13307                                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
13308                                 /*
13309                                  * We cheat here and don't send a RST, we should send one
13310                                  * when the pacer drops the connection.
13311                                  */
13312 #ifdef TCP_ACCOUNTING
13313                                 rdstc = get_cyclecount();
13314                                 if (rdstc > ts_val) {
13315                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13316                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13317                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13318                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13319                                         }
13320                                 }
13321                                 sched_unpin();
13322 #endif
13323                                 INP_WUNLOCK(rack->rc_inp);
13324                                 m_freem(m);
13325                                 return (1);
13326                         }
13327                 }
13328                 if (ourfinisacked) {
13329                         switch(tp->t_state) {
13330                         case TCPS_CLOSING:
13331 #ifdef TCP_ACCOUNTING
13332                                 rdstc = get_cyclecount();
13333                                 if (rdstc > ts_val) {
13334                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13335                                                         (rdstc - ts_val));
13336                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13337                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13338                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13339                                         }
13340                                 }
13341                                 sched_unpin();
13342 #endif
13343                                 tcp_twstart(tp);
13344                                 m_freem(m);
13345                                 return (1);
13346                                 break;
13347                         case TCPS_LAST_ACK:
13348 #ifdef TCP_ACCOUNTING
13349                                 rdstc = get_cyclecount();
13350                                 if (rdstc > ts_val) {
13351                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13352                                                         (rdstc - ts_val));
13353                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13354                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13355                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13356                                         }
13357                                 }
13358                                 sched_unpin();
13359 #endif
13360                                 tp = tcp_close(tp);
13361                                 ctf_do_drop(m, tp);
13362                                 return (1);
13363                                 break;
13364                         case TCPS_FIN_WAIT_1:
13365 #ifdef TCP_ACCOUNTING
13366                                 rdstc = get_cyclecount();
13367                                 if (rdstc > ts_val) {
13368                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13369                                                         (rdstc - ts_val));
13370                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13371                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13372                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13373                                         }
13374                                 }
13375 #endif
13376                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13377                                         soisdisconnected(so);
13378                                         tcp_timer_activate(tp, TT_2MSL,
13379                                                            (tcp_fast_finwait2_recycle ?
13380                                                             tcp_finwait2_timeout :
13381                                                             TP_MAXIDLE(tp)));
13382                                 }
13383                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13384                                 break;
13385                         default:
13386                                 break;
13387                         }
13388                 }
13389                 if (rack->r_fast_output) {
13390                         /*
13391                          * We re doing fast output.. can we expand that?
13392                          */
13393                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13394                 }
13395 #ifdef TCP_ACCOUNTING
13396                 rdstc = get_cyclecount();
13397                 if (rdstc > ts_val) {
13398                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13399                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13400                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13401                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13402                         }
13403                 }
13404
13405         } else if (win_up_req) {
13406                 rdstc = get_cyclecount();
13407                 if (rdstc > ts_val) {
13408                         counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13409                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13410                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13411                         }
13412                 }
13413 #endif
13414         }
13415         /* Now is there a next packet, if so we are done */
13416         m_freem(m);
13417         did_out = 0;
13418         if (nxt_pkt) {
13419 #ifdef TCP_ACCOUNTING
13420                 sched_unpin();
13421 #endif
13422                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13423                 return (0);
13424         }
13425         rack_handle_might_revert(tp, rack);
13426         ctf_calc_rwin(so, tp);
13427         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13428         send_out_a_rst:
13429                 (void)tp->t_fb->tfb_tcp_output(tp);
13430                 did_out = 1;
13431         }
13432         rack_free_trim(rack);
13433 #ifdef TCP_ACCOUNTING
13434         sched_unpin();
13435 #endif
13436         rack_timer_audit(tp, rack, &so->so_snd);
13437         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13438         return (0);
13439 }
13440
13441
13442 static int
13443 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13444     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13445     int32_t nxt_pkt, struct timeval *tv)
13446 {
13447 #ifdef TCP_ACCOUNTING
13448         uint64_t ts_val;
13449 #endif
13450         int32_t thflags, retval, did_out = 0;
13451         int32_t way_out = 0;
13452         uint32_t cts;
13453         uint32_t tiwin;
13454         struct timespec ts;
13455         struct tcpopt to;
13456         struct tcp_rack *rack;
13457         struct rack_sendmap *rsm;
13458         int32_t prev_state = 0;
13459 #ifdef TCP_ACCOUNTING
13460         int ack_val_set = 0xf;
13461 #endif
13462         int nsegs;
13463         uint32_t us_cts;
13464         /*
13465          * tv passed from common code is from either M_TSTMP_LRO or
13466          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13467          */
13468         if (m->m_flags & M_ACKCMP) {
13469                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13470         }
13471         if (m->m_flags & M_ACKCMP) {
13472                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13473         }
13474         nsegs = m->m_pkthdr.lro_nsegs;
13475         counter_u64_add(rack_proc_non_comp_ack, 1);
13476         thflags = th->th_flags;
13477 #ifdef TCP_ACCOUNTING
13478         sched_pin();
13479         if (thflags & TH_ACK)
13480                 ts_val = get_cyclecount();
13481 #endif
13482         cts = tcp_tv_to_usectick(tv);
13483         rack = (struct tcp_rack *)tp->t_fb_ptr;
13484
13485         if ((m->m_flags & M_TSTMP) ||
13486             (m->m_flags & M_TSTMP_LRO)) {
13487                 mbuf_tstmp2timespec(m, &ts);
13488                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13489                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13490         } else
13491                 rack->r_ctl.act_rcv_time = *tv;
13492         kern_prefetch(rack, &prev_state);
13493         prev_state = 0;
13494         /*
13495          * Unscale the window into a 32-bit value. For the SYN_SENT state
13496          * the scale is zero.
13497          */
13498         tiwin = th->th_win << tp->snd_scale;
13499         /*
13500          * Parse options on any incoming segment.
13501          */
13502         memset(&to, 0, sizeof(to));
13503         tcp_dooptions(&to, (u_char *)(th + 1),
13504             (th->th_off << 2) - sizeof(struct tcphdr),
13505             (thflags & TH_SYN) ? TO_SYN : 0);
13506 #ifdef TCP_ACCOUNTING
13507         if (thflags & TH_ACK) {
13508                 /*
13509                  * We have a tradeoff here. We can either do what we are
13510                  * doing i.e. pinning to this CPU and then doing the accounting
13511                  * <or> we could do a critical enter, setup the rdtsc and cpu
13512                  * as in below, and then validate we are on the same CPU on
13513                  * exit. I have choosen to not do the critical enter since
13514                  * that often will gain you a context switch, and instead lock
13515                  * us (line above this if) to the same CPU with sched_pin(). This
13516                  * means we may be context switched out for a higher priority
13517                  * interupt but we won't be moved to another CPU.
13518                  *
13519                  * If this occurs (which it won't very often since we most likely
13520                  * are running this code in interupt context and only a higher
13521                  * priority will bump us ... clock?) we will falsely add in
13522                  * to the time the interupt processing time plus the ack processing
13523                  * time. This is ok since its a rare event.
13524                  */
13525                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
13526                                                     ctf_fixed_maxseg(tp));
13527         }
13528 #endif
13529         NET_EPOCH_ASSERT();
13530         INP_WLOCK_ASSERT(tp->t_inpcb);
13531         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
13532             __func__));
13533         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
13534             __func__));
13535         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13536                 union tcp_log_stackspecific log;
13537                 struct timeval ltv;
13538 #ifdef NETFLIX_HTTP_LOGGING
13539                 struct http_sendfile_track *http_req;
13540
13541                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
13542                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
13543                 } else {
13544                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
13545                 }
13546 #endif
13547                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13548                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13549                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13550                 if (rack->rack_no_prr == 0)
13551                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13552                 else
13553                         log.u_bbr.flex1 = 0;
13554                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13555                 log.u_bbr.use_lt_bw <<= 1;
13556                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13557                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13558                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13559                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
13560                 log.u_bbr.flex3 = m->m_flags;
13561                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13562                 log.u_bbr.lost = thflags;
13563                 log.u_bbr.pacing_gain = 0x1;
13564 #ifdef TCP_ACCOUNTING
13565                 log.u_bbr.cwnd_gain = ack_val_set;
13566 #endif
13567                 log.u_bbr.flex7 = 2;
13568                 if (m->m_flags & M_TSTMP) {
13569                         /* Record the hardware timestamp if present */
13570                         mbuf_tstmp2timespec(m, &ts);
13571                         ltv.tv_sec = ts.tv_sec;
13572                         ltv.tv_usec = ts.tv_nsec / 1000;
13573                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13574                 } else if (m->m_flags & M_TSTMP_LRO) {
13575                         /* Record the LRO the arrival timestamp */
13576                         mbuf_tstmp2timespec(m, &ts);
13577                         ltv.tv_sec = ts.tv_sec;
13578                         ltv.tv_usec = ts.tv_nsec / 1000;
13579                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13580                 }
13581                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13582                 /* Log the rcv time */
13583                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
13584 #ifdef NETFLIX_HTTP_LOGGING
13585                 log.u_bbr.applimited = tp->t_http_closed;
13586                 log.u_bbr.applimited <<= 8;
13587                 log.u_bbr.applimited |= tp->t_http_open;
13588                 log.u_bbr.applimited <<= 8;
13589                 log.u_bbr.applimited |= tp->t_http_req;
13590                 if (http_req) {
13591                         /* Copy out any client req info */
13592                         /* seconds */
13593                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13594                         /* useconds */
13595                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13596                         log.u_bbr.rttProp = http_req->timestamp;
13597                         log.u_bbr.cur_del_rate = http_req->start;
13598                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13599                                 log.u_bbr.flex8 |= 1;
13600                         } else {
13601                                 log.u_bbr.flex8 |= 2;
13602                                 log.u_bbr.bw_inuse = http_req->end;
13603                         }
13604                         log.u_bbr.flex6 = http_req->start_seq;
13605                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13606                                 log.u_bbr.flex8 |= 4;
13607                                 log.u_bbr.epoch = http_req->end_seq;
13608                         }
13609                 }
13610 #endif
13611                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
13612                     tlen, &log, true, &ltv);
13613         }
13614         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
13615                 way_out = 4;
13616                 retval = 0;
13617                 m_freem(m);
13618                 goto done_with_input;
13619         }
13620         /*
13621          * If a segment with the ACK-bit set arrives in the SYN-SENT state
13622          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
13623          */
13624         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
13625             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
13626                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13627                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13628 #ifdef TCP_ACCOUNTING
13629                 sched_unpin();
13630 #endif
13631                 return (1);
13632         }
13633
13634         /*
13635          * Parse options on any incoming segment.
13636          */
13637         tcp_dooptions(&to, (u_char *)(th + 1),
13638             (th->th_off << 2) - sizeof(struct tcphdr),
13639             (thflags & TH_SYN) ? TO_SYN : 0);
13640
13641         /*
13642          * If timestamps were negotiated during SYN/ACK and a
13643          * segment without a timestamp is received, silently drop
13644          * the segment, unless it is a RST segment or missing timestamps are
13645          * tolerated.
13646          * See section 3.2 of RFC 7323.
13647          */
13648         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
13649             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
13650                 way_out = 5;
13651                 retval = 0;
13652                 m_freem(m);
13653                 goto done_with_input;
13654         }
13655
13656         /*
13657          * Segment received on connection. Reset idle time and keep-alive
13658          * timer. XXX: This should be done after segment validation to
13659          * ignore broken/spoofed segs.
13660          */
13661         if  (tp->t_idle_reduce &&
13662              (tp->snd_max == tp->snd_una) &&
13663              ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13664                 counter_u64_add(rack_input_idle_reduces, 1);
13665                 rack_cc_after_idle(rack, tp);
13666         }
13667         tp->t_rcvtime = ticks;
13668 #ifdef STATS
13669         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
13670 #endif
13671         if (tiwin > rack->r_ctl.rc_high_rwnd)
13672                 rack->r_ctl.rc_high_rwnd = tiwin;
13673         /*
13674          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
13675          * this to occur after we've validated the segment.
13676          */
13677         if (tp->t_flags2 & TF2_ECN_PERMIT) {
13678                 if (thflags & TH_CWR) {
13679                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13680                         tp->t_flags |= TF_ACKNOW;
13681                 }
13682                 switch (iptos & IPTOS_ECN_MASK) {
13683                 case IPTOS_ECN_CE:
13684                         tp->t_flags2 |= TF2_ECN_SND_ECE;
13685                         KMOD_TCPSTAT_INC(tcps_ecn_ce);
13686                         break;
13687                 case IPTOS_ECN_ECT0:
13688                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13689                         break;
13690                 case IPTOS_ECN_ECT1:
13691                         KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13692                         break;
13693                 }
13694
13695                 /* Process a packet differently from RFC3168. */
13696                 cc_ecnpkt_handler(tp, th, iptos);
13697
13698                 /* Congestion experienced. */
13699                 if (thflags & TH_ECE) {
13700                         rack_cong_signal(tp, CC_ECN, th->th_ack);
13701                 }
13702         }
13703
13704         /*
13705          * If echoed timestamp is later than the current time, fall back to
13706          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13707          * were used when this connection was established.
13708          */
13709         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
13710                 to.to_tsecr -= tp->ts_offset;
13711                 if (TSTMP_GT(to.to_tsecr, cts))
13712                         to.to_tsecr = 0;
13713         }
13714
13715         /*
13716          * If its the first time in we need to take care of options and
13717          * verify we can do SACK for rack!
13718          */
13719         if (rack->r_state == 0) {
13720                 /* Should be init'd by rack_init() */
13721                 KASSERT(rack->rc_inp != NULL,
13722                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
13723                 if (rack->rc_inp == NULL) {
13724                         rack->rc_inp = tp->t_inpcb;
13725                 }
13726
13727                 /*
13728                  * Process options only when we get SYN/ACK back. The SYN
13729                  * case for incoming connections is handled in tcp_syncache.
13730                  * According to RFC1323 the window field in a SYN (i.e., a
13731                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
13732                  * this is traditional behavior, may need to be cleaned up.
13733                  */
13734                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
13735                         /* Handle parallel SYN for ECN */
13736                         if (!(thflags & TH_ACK) &&
13737                             ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
13738                             ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
13739                                 tp->t_flags2 |= TF2_ECN_PERMIT;
13740                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
13741                                 TCPSTAT_INC(tcps_ecn_shs);
13742                         }
13743                         if ((to.to_flags & TOF_SCALE) &&
13744                             (tp->t_flags & TF_REQ_SCALE)) {
13745                                 tp->t_flags |= TF_RCVD_SCALE;
13746                                 tp->snd_scale = to.to_wscale;
13747                         } else
13748                                 tp->t_flags &= ~TF_REQ_SCALE;
13749                         /*
13750                          * Initial send window.  It will be updated with the
13751                          * next incoming segment to the scaled value.
13752                          */
13753                         tp->snd_wnd = th->th_win;
13754                         rack_validate_fo_sendwin_up(tp, rack);
13755                         if ((to.to_flags & TOF_TS) &&
13756                             (tp->t_flags & TF_REQ_TSTMP)) {
13757                                 tp->t_flags |= TF_RCVD_TSTMP;
13758                                 tp->ts_recent = to.to_tsval;
13759                                 tp->ts_recent_age = cts;
13760                         } else
13761                                 tp->t_flags &= ~TF_REQ_TSTMP;
13762                         if (to.to_flags & TOF_MSS) {
13763                                 tcp_mss(tp, to.to_mss);
13764                         }
13765                         if ((tp->t_flags & TF_SACK_PERMIT) &&
13766                             (to.to_flags & TOF_SACKPERM) == 0)
13767                                 tp->t_flags &= ~TF_SACK_PERMIT;
13768                         if (IS_FASTOPEN(tp->t_flags)) {
13769                                 if (to.to_flags & TOF_FASTOPEN) {
13770                                         uint16_t mss;
13771
13772                                         if (to.to_flags & TOF_MSS)
13773                                                 mss = to.to_mss;
13774                                         else
13775                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
13776                                                         mss = TCP6_MSS;
13777                                                 else
13778                                                         mss = TCP_MSS;
13779                                         tcp_fastopen_update_cache(tp, mss,
13780                                             to.to_tfo_len, to.to_tfo_cookie);
13781                                 } else
13782                                         tcp_fastopen_disable_path(tp);
13783                         }
13784                 }
13785                 /*
13786                  * At this point we are at the initial call. Here we decide
13787                  * if we are doing RACK or not. We do this by seeing if
13788                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
13789                  * The code now does do dup-ack counting so if you don't
13790                  * switch back you won't get rack & TLP, but you will still
13791                  * get this stack.
13792                  */
13793
13794                 if ((rack_sack_not_required == 0) &&
13795                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
13796                         tcp_switch_back_to_default(tp);
13797                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
13798                             tlen, iptos);
13799 #ifdef TCP_ACCOUNTING
13800                         sched_unpin();
13801 #endif
13802                         return (1);
13803                 }
13804                 tcp_set_hpts(tp->t_inpcb);
13805                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
13806         }
13807         if (thflags & TH_FIN)
13808                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
13809         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13810         if ((rack->rc_gp_dyn_mul) &&
13811             (rack->use_fixed_rate == 0) &&
13812             (rack->rc_always_pace)) {
13813                 /* Check in on probertt */
13814                 rack_check_probe_rtt(rack, us_cts);
13815         }
13816         if (rack->forced_ack) {
13817                 uint32_t us_rtt;
13818
13819                 /*
13820                  * A persist or keep-alive was forced out, update our
13821                  * min rtt time. Note we do not worry about lost
13822                  * retransmissions since KEEP-ALIVES and persists
13823                  * are usually way long on times of sending (though
13824                  * if we were really paranoid or worried we could
13825                  * at least use timestamps if available to validate).
13826                  */
13827                 rack->forced_ack = 0;
13828                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13829                 if (us_rtt == 0)
13830                         us_rtt = 1;
13831                 rack_log_rtt_upd(tp, rack, us_rtt, 0, NULL, 3);
13832                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13833         }
13834         /*
13835          * This is the one exception case where we set the rack state
13836          * always. All other times (timers etc) we must have a rack-state
13837          * set (so we assure we have done the checks above for SACK).
13838          */
13839         rack->r_ctl.rc_rcvtime = cts;
13840         if (rack->r_state != tp->t_state)
13841                 rack_set_state(tp, rack);
13842         if (SEQ_GT(th->th_ack, tp->snd_una) &&
13843             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
13844                 kern_prefetch(rsm, &prev_state);
13845         prev_state = rack->r_state;
13846         rack_clear_rate_sample(rack);
13847         retval = (*rack->r_substate) (m, th, so,
13848             tp, &to, drop_hdrlen,
13849             tlen, tiwin, thflags, nxt_pkt, iptos);
13850 #ifdef INVARIANTS
13851         if ((retval == 0) &&
13852             (tp->t_inpcb == NULL)) {
13853                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
13854                     retval, tp, prev_state);
13855         }
13856 #endif
13857         if (retval == 0) {
13858                 /*
13859                  * If retval is 1 the tcb is unlocked and most likely the tp
13860                  * is gone.
13861                  */
13862                 INP_WLOCK_ASSERT(tp->t_inpcb);
13863                 if ((rack->rc_gp_dyn_mul) &&
13864                     (rack->rc_always_pace) &&
13865                     (rack->use_fixed_rate == 0) &&
13866                     rack->in_probe_rtt &&
13867                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
13868                         /*
13869                          * If we are going for target, lets recheck before
13870                          * we output.
13871                          */
13872                         rack_check_probe_rtt(rack, us_cts);
13873                 }
13874                 if (rack->set_pacing_done_a_iw == 0) {
13875                         /* How much has been acked? */
13876                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
13877                                 /* We have enough to set in the pacing segment size */
13878                                 rack->set_pacing_done_a_iw = 1;
13879                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13880                         }
13881                 }
13882                 tcp_rack_xmit_timer_commit(rack, tp);
13883 #ifdef TCP_ACCOUNTING
13884                 /*
13885                  * If we set the ack_val_se to what ack processing we are doing
13886                  * we also want to track how many cycles we burned. Note
13887                  * the bits after tcp_output we let be "free". This is because
13888                  * we are also tracking the tcp_output times as well. Note the
13889                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
13890                  * 0xf cannot be returned and is what we initialize it too to
13891                  * indicate we are not doing the tabulations.
13892                  */
13893                 if (ack_val_set != 0xf) {
13894                         uint64_t crtsc;
13895
13896                         crtsc = get_cyclecount();
13897                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13898                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13899                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
13900                         }
13901                 }
13902 #endif
13903                 if (nxt_pkt == 0) {
13904                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13905 do_output_now:
13906                                 did_out = 1;
13907                                 (void)tp->t_fb->tfb_tcp_output(tp);
13908                         }
13909                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
13910                         rack_free_trim(rack);
13911                 }
13912                 if ((nxt_pkt == 0) &&
13913                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
13914                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
13915                      (tp->t_flags & TF_DELACK) ||
13916                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13917                       (tp->t_state <= TCPS_CLOSING)))) {
13918                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
13919                         if ((tp->snd_max == tp->snd_una) &&
13920                             ((tp->t_flags & TF_DELACK) == 0) &&
13921                             (rack->rc_inp->inp_in_hpts) &&
13922                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13923                                 /* keep alive not needed if we are hptsi output yet */
13924                                 ;
13925                         } else {
13926                                 int late = 0;
13927                                 if (rack->rc_inp->inp_in_hpts) {
13928                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13929                                                 us_cts = tcp_get_usecs(NULL);
13930                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13931                                                         rack->r_early = 1;
13932                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13933                                                 } else
13934                                                         late = 1;
13935                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13936                                         }
13937                                         tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
13938                                 }
13939                                 if (late && (did_out == 0)) {
13940                                         /*
13941                                          * We are late in the sending
13942                                          * and we did not call the output
13943                                          * (this probably should not happen).
13944                                          */
13945                                         goto do_output_now;
13946                                 }
13947                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13948                         }
13949                         way_out = 1;
13950                 } else if (nxt_pkt == 0) {
13951                         /* Do we have the correct timer running? */
13952                         rack_timer_audit(tp, rack, &so->so_snd);
13953                         way_out = 2;
13954                 }
13955         done_with_input:
13956                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
13957                 if (did_out)
13958                         rack->r_wanted_output = 0;
13959 #ifdef INVARIANTS
13960                 if (tp->t_inpcb == NULL) {
13961                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
13962                               did_out,
13963                               retval, tp, prev_state);
13964                 }
13965 #endif
13966 #ifdef TCP_ACCOUNTING
13967         } else {
13968                 /*
13969                  * Track the time (see above).
13970                  */
13971                 if (ack_val_set != 0xf) {
13972                         uint64_t crtsc;
13973
13974                         crtsc = get_cyclecount();
13975                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13976                         /*
13977                          * Note we *DO NOT* increment the per-tcb counters since
13978                          * in the else the TP may be gone!!
13979                          */
13980                 }
13981 #endif
13982         }
13983 #ifdef TCP_ACCOUNTING
13984         sched_unpin();
13985 #endif
13986         return (retval);
13987 }
13988
13989 void
13990 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
13991     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
13992 {
13993         struct timeval tv;
13994
13995         /* First lets see if we have old packets */
13996         if (tp->t_in_pkt) {
13997                 if (ctf_do_queued_segments(so, tp, 1)) {
13998                         m_freem(m);
13999                         return;
14000                 }
14001         }
14002         if (m->m_flags & M_TSTMP_LRO) {
14003                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
14004                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
14005         } else {
14006                 /* Should not be should we kassert instead? */
14007                 tcp_get_usecs(&tv);
14008         }
14009         if (rack_do_segment_nounlock(m, th, so, tp,
14010                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14011                 INP_WUNLOCK(tp->t_inpcb);
14012         }
14013 }
14014
14015 struct rack_sendmap *
14016 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14017 {
14018         struct rack_sendmap *rsm = NULL;
14019         int32_t idx;
14020         uint32_t srtt = 0, thresh = 0, ts_low = 0;
14021
14022         /* Return the next guy to be re-transmitted */
14023         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14024                 return (NULL);
14025         }
14026         if (tp->t_flags & TF_SENTFIN) {
14027                 /* retran the end FIN? */
14028                 return (NULL);
14029         }
14030         /* ok lets look at this one */
14031         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14032         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14033                 goto check_it;
14034         }
14035         rsm = rack_find_lowest_rsm(rack);
14036         if (rsm == NULL) {
14037                 return (NULL);
14038         }
14039 check_it:
14040         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14041             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14042                 /*
14043                  * No sack so we automatically do the 3 strikes and
14044                  * retransmit (no rack timer would be started).
14045                  */
14046
14047                 return (rsm);
14048         }
14049         if (rsm->r_flags & RACK_ACKED) {
14050                 return (NULL);
14051         }
14052         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14053             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14054                 /* Its not yet ready */
14055                 return (NULL);
14056         }
14057         srtt = rack_grab_rtt(tp, rack);
14058         idx = rsm->r_rtr_cnt - 1;
14059         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14060         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14061         if ((tsused == ts_low) ||
14062             (TSTMP_LT(tsused, ts_low))) {
14063                 /* No time since sending */
14064                 return (NULL);
14065         }
14066         if ((tsused - ts_low) < thresh) {
14067                 /* It has not been long enough yet */
14068                 return (NULL);
14069         }
14070         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14071             ((rsm->r_flags & RACK_SACK_PASSED) &&
14072              (rack->sack_attack_disable == 0))) {
14073                 /*
14074                  * We have passed the dup-ack threshold <or>
14075                  * a SACK has indicated this is missing.
14076                  * Note that if you are a declared attacker
14077                  * it is only the dup-ack threshold that
14078                  * will cause retransmits.
14079                  */
14080                 /* log retransmit reason */
14081                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14082                 rack->r_fast_output = 0;
14083                 return (rsm);
14084         }
14085         return (NULL);
14086 }
14087
14088 static void
14089 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14090                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14091                            int line, struct rack_sendmap *rsm)
14092 {
14093         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14094                 union tcp_log_stackspecific log;
14095                 struct timeval tv;
14096
14097                 memset(&log, 0, sizeof(log));
14098                 log.u_bbr.flex1 = slot;
14099                 log.u_bbr.flex2 = len;
14100                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14101                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14102                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14103                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14104                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14105                 log.u_bbr.use_lt_bw <<= 1;
14106                 log.u_bbr.use_lt_bw |= rack->r_late;
14107                 log.u_bbr.use_lt_bw <<= 1;
14108                 log.u_bbr.use_lt_bw |= rack->r_early;
14109                 log.u_bbr.use_lt_bw <<= 1;
14110                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14111                 log.u_bbr.use_lt_bw <<= 1;
14112                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14113                 log.u_bbr.use_lt_bw <<= 1;
14114                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14115                 log.u_bbr.use_lt_bw <<= 1;
14116                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14117                 log.u_bbr.use_lt_bw <<= 1;
14118                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14119                 log.u_bbr.pkt_epoch = line;
14120                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14121                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14122                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14123                 log.u_bbr.bw_inuse = bw_est;
14124                 log.u_bbr.delRate = bw;
14125                 if (rack->r_ctl.gp_bw == 0)
14126                         log.u_bbr.cur_del_rate = 0;
14127                 else
14128                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14129                 log.u_bbr.rttProp = len_time;
14130                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14131                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14132                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14133                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14134                         /* We are in slow start */
14135                         log.u_bbr.flex7 = 1;
14136                 } else {
14137                         /* we are on congestion avoidance */
14138                         log.u_bbr.flex7 = 0;
14139                 }
14140                 log.u_bbr.flex8 = method;
14141                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14142                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14143                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14144                 log.u_bbr.cwnd_gain <<= 1;
14145                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14146                 log.u_bbr.cwnd_gain <<= 1;
14147                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14148                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14149                     &rack->rc_inp->inp_socket->so_rcv,
14150                     &rack->rc_inp->inp_socket->so_snd,
14151                     BBR_LOG_HPTSI_CALC, 0,
14152                     0, &log, false, &tv);
14153         }
14154 }
14155
14156 static uint32_t
14157 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14158 {
14159         uint32_t new_tso, user_max;
14160
14161         user_max = rack->rc_user_set_max_segs * mss;
14162         if (rack->rc_force_max_seg) {
14163                 return (user_max);
14164         }
14165         if (rack->use_fixed_rate &&
14166             ((rack->r_ctl.crte == NULL) ||
14167              (bw != rack->r_ctl.crte->rate))) {
14168                 /* Use the user mss since we are not exactly matched */
14169                 return (user_max);
14170         }
14171         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14172         if (new_tso > user_max)
14173                 new_tso = user_max;
14174         return (new_tso);
14175 }
14176
14177 static int32_t
14178 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)
14179 {
14180         uint64_t lentim, fill_bw;
14181
14182         /* Lets first see if we are full, if so continue with normal rate */
14183         rack->r_via_fill_cw = 0;
14184         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14185                 return (slot);
14186         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14187                 return (slot);
14188         if (rack->r_ctl.rc_last_us_rtt == 0)
14189                 return (slot);
14190         if (rack->rc_pace_fill_if_rttin_range &&
14191             (rack->r_ctl.rc_last_us_rtt >=
14192              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14193                 /* The rtt is huge, N * smallest, lets not fill */
14194                 return (slot);
14195         }
14196         /*
14197          * first lets calculate the b/w based on the last us-rtt
14198          * and the sndwnd.
14199          */
14200         fill_bw = rack->r_ctl.cwnd_to_use;
14201         /* Take the rwnd if its smaller */
14202         if (fill_bw > rack->rc_tp->snd_wnd)
14203                 fill_bw = rack->rc_tp->snd_wnd;
14204         if (rack->r_fill_less_agg) {
14205                 /*
14206                  * Now take away the inflight (this will reduce our
14207                  * aggressiveness and yeah, if we get that much out in 1RTT
14208                  * we will have had acks come back and still be behind).
14209                  */
14210                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14211         }
14212         /* Now lets make it into a b/w */
14213         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14214         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14215         /* We are below the min b/w */
14216         if (non_paced)
14217                 *rate_wanted = fill_bw;
14218         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14219                 return (slot);
14220         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14221                 fill_bw = rack->r_ctl.bw_rate_cap;
14222         rack->r_via_fill_cw = 1;
14223         if (rack->r_rack_hw_rate_caps &&
14224             (rack->r_ctl.crte != NULL)) {
14225                 uint64_t high_rate;
14226
14227                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14228                 if (fill_bw > high_rate) {
14229                         /* We are capping bw at the highest rate table entry */
14230                         if (*rate_wanted > high_rate) {
14231                                 /* The original rate was also capped */
14232                                 rack->r_via_fill_cw = 0;
14233                         }
14234                         rack_log_hdwr_pacing(rack,
14235                                              fill_bw, high_rate, __LINE__,
14236                                              0, 3);
14237                         fill_bw = high_rate;
14238                         if (capped)
14239                                 *capped = 1;
14240                 }
14241         } else if ((rack->r_ctl.crte == NULL) &&
14242                    (rack->rack_hdrw_pacing == 0) &&
14243                    (rack->rack_hdw_pace_ena) &&
14244                    rack->r_rack_hw_rate_caps &&
14245                    (rack->rack_attempt_hdwr_pace == 0) &&
14246                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14247                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14248                 /*
14249                  * Ok we may have a first attempt that is greater than our top rate
14250                  * lets check.
14251                  */
14252                 uint64_t high_rate;
14253
14254                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14255                 if (high_rate) {
14256                         if (fill_bw > high_rate) {
14257                                 fill_bw = high_rate;
14258                                 if (capped)
14259                                         *capped = 1;
14260                         }
14261                 }
14262         }
14263         /*
14264          * Ok fill_bw holds our mythical b/w to fill the cwnd
14265          * in a rtt, what does that time wise equate too?
14266          */
14267         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14268         lentim /= fill_bw;
14269         *rate_wanted = fill_bw;
14270         if (non_paced || (lentim < slot)) {
14271                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14272                                            0, lentim, 12, __LINE__, NULL);
14273                 return ((int32_t)lentim);
14274         } else
14275                 return (slot);
14276 }
14277
14278 static int32_t
14279 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14280 {
14281         struct rack_sendmap *lrsm;
14282         int32_t slot = 0;
14283         int can_start_hw_pacing = 1;
14284         int err;
14285
14286         if (rack->rc_always_pace == 0) {
14287                 /*
14288                  * We use the most optimistic possible cwnd/srtt for
14289                  * sending calculations. This will make our
14290                  * calculation anticipate getting more through
14291                  * quicker then possible. But thats ok we don't want
14292                  * the peer to have a gap in data sending.
14293                  */
14294                 uint32_t srtt, cwnd, tr_perms = 0;
14295                 int32_t reduce = 0;
14296
14297         old_method:
14298                 /*
14299                  * We keep no precise pacing with the old method
14300                  * instead we use the pacer to mitigate bursts.
14301                  */
14302                 if (rack->r_ctl.rc_rack_min_rtt)
14303                         srtt = rack->r_ctl.rc_rack_min_rtt;
14304                 else
14305                         srtt = max(tp->t_srtt, 1);
14306                 if (rack->r_ctl.rc_rack_largest_cwnd)
14307                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14308                 else
14309                         cwnd = rack->r_ctl.cwnd_to_use;
14310                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14311                 tr_perms = (cwnd * 1000) / srtt;
14312                 if (tr_perms == 0) {
14313                         tr_perms = ctf_fixed_maxseg(tp);
14314                 }
14315                 /*
14316                  * Calculate how long this will take to drain, if
14317                  * the calculation comes out to zero, thats ok we
14318                  * will use send_a_lot to possibly spin around for
14319                  * more increasing tot_len_this_send to the point
14320                  * that its going to require a pace, or we hit the
14321                  * cwnd. Which in that case we are just waiting for
14322                  * a ACK.
14323                  */
14324                 slot = len / tr_perms;
14325                 /* Now do we reduce the time so we don't run dry? */
14326                 if (slot && rack_slot_reduction) {
14327                         reduce = (slot / rack_slot_reduction);
14328                         if (reduce < slot) {
14329                                 slot -= reduce;
14330                         } else
14331                                 slot = 0;
14332                 }
14333                 slot *= HPTS_USEC_IN_MSEC;
14334                 if (rsm == NULL) {
14335                         /*
14336                          * We always consider ourselves app limited with old style
14337                          * that are not retransmits. This could be the initial
14338                          * measurement, but thats ok its all setup and specially
14339                          * handled. If another send leaks out, then that too will
14340                          * be mark app-limited.
14341                          */
14342                         lrsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14343                         if (lrsm && ((lrsm->r_flags & RACK_APP_LIMITED) == 0)) {
14344                                 rack->r_ctl.rc_first_appl = lrsm;
14345                                 lrsm->r_flags |= RACK_APP_LIMITED;
14346                                 rack->r_ctl.rc_app_limited_cnt++;
14347                         }
14348                 }
14349                 if (rack->rc_pace_to_cwnd) {
14350                         uint64_t rate_wanted = 0;
14351
14352                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14353                         rack->rc_ack_can_sendout_data = 1;
14354                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL);
14355                 } else
14356                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL);
14357         } else {
14358                 uint64_t bw_est, res, lentim, rate_wanted;
14359                 uint32_t orig_val, srtt, segs, oh;
14360                 int capped = 0;
14361                 int prev_fill;
14362
14363                 if ((rack->r_rr_config == 1) && rsm) {
14364                         return (rack->r_ctl.rc_min_to);
14365                 }
14366                 if (rack->use_fixed_rate) {
14367                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14368                 } else if ((rack->r_ctl.init_rate == 0) &&
14369 #ifdef NETFLIX_PEAKRATE
14370                            (rack->rc_tp->t_maxpeakrate == 0) &&
14371 #endif
14372                            (rack->r_ctl.gp_bw == 0)) {
14373                         /* no way to yet do an estimate */
14374                         bw_est = rate_wanted = 0;
14375                 } else {
14376                         bw_est = rack_get_bw(rack);
14377                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14378                 }
14379                 if ((bw_est == 0) || (rate_wanted == 0) ||
14380                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14381                         /*
14382                          * No way yet to make a b/w estimate or
14383                          * our raise is set incorrectly.
14384                          */
14385                         goto old_method;
14386                 }
14387                 /* We need to account for all the overheads */
14388                 segs = (len + segsiz - 1) / segsiz;
14389                 /*
14390                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
14391                  * and how much data we put in each packet. Yes this
14392                  * means we may be off if we are larger than 1500 bytes
14393                  * or smaller. But this just makes us more conservative.
14394                  */
14395                 if (rack_hw_rate_min &&
14396                     (bw_est < rack_hw_rate_min))
14397                         can_start_hw_pacing = 0;
14398                 if (ETHERNET_SEGMENT_SIZE > segsiz)
14399                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
14400                 else
14401                         oh = 0;
14402                 segs *= oh;
14403                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14404                 res = lentim / rate_wanted;
14405                 slot = (uint32_t)res;
14406                 orig_val = rack->r_ctl.rc_pace_max_segs;
14407                 if (rack->r_ctl.crte == NULL) {
14408                         /*
14409                          * Only do this if we are not hardware pacing
14410                          * since if we are doing hw-pacing below we will
14411                          * set make a call after setting up or changing
14412                          * the rate.
14413                          */
14414                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14415                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14416                         /*
14417                          * We lost our rate somehow, this can happen
14418                          * if the interface changed underneath us.
14419                          */
14420                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14421                         rack->r_ctl.crte = NULL;
14422                         /* Lets re-allow attempting to setup pacing */
14423                         rack->rack_hdrw_pacing = 0;
14424                         rack->rack_attempt_hdwr_pace = 0;
14425                         rack_log_hdwr_pacing(rack,
14426                                              rate_wanted, bw_est, __LINE__,
14427                                              0, 6);
14428                 }
14429                 /* Did we change the TSO size, if so log it */
14430                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14431                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL);
14432                 prev_fill = rack->r_via_fill_cw;
14433                 if ((rack->rc_pace_to_cwnd) &&
14434                     (capped == 0) &&
14435                     (rack->use_fixed_rate == 0) &&
14436                     (rack->in_probe_rtt == 0) &&
14437                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14438                         /*
14439                          * We want to pace at our rate *or* faster to
14440                          * fill the cwnd to the max if its not full.
14441                          */
14442                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14443                 }
14444                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14445                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14446                         if ((rack->rack_hdw_pace_ena) &&
14447                             (can_start_hw_pacing > 0) &&
14448                             (rack->rack_hdrw_pacing == 0) &&
14449                             (rack->rack_attempt_hdwr_pace == 0)) {
14450                                 /*
14451                                  * Lets attempt to turn on hardware pacing
14452                                  * if we can.
14453                                  */
14454                                 rack->rack_attempt_hdwr_pace = 1;
14455                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14456                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
14457                                                                        rate_wanted,
14458                                                                        RS_PACING_GEQ,
14459                                                                        &err, &rack->r_ctl.crte_prev_rate);
14460                                 if (rack->r_ctl.crte) {
14461                                         rack->rack_hdrw_pacing = 1;
14462                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14463                                                                                                  0, rack->r_ctl.crte,
14464                                                                                                  NULL);
14465                                         rack_log_hdwr_pacing(rack,
14466                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14467                                                              err, 0);
14468                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14469                                 } else {
14470                                         counter_u64_add(rack_hw_pace_init_fail, 1);
14471                                 }
14472                         } else if (rack->rack_hdrw_pacing &&
14473                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14474                                 /* Do we need to adjust our rate? */
14475                                 const struct tcp_hwrate_limit_table *nrte;
14476
14477                                 if (rack->r_up_only &&
14478                                     (rate_wanted < rack->r_ctl.crte->rate)) {
14479                                         /**
14480                                          * We have four possible states here
14481                                          * having to do with the previous time
14482                                          * and this time.
14483                                          *   previous  |  this-time
14484                                          * A)     0      |     0   -- fill_cw not in the picture
14485                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
14486                                          * C)     1      |     1   -- all rates from fill_cw
14487                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
14488                                          *
14489                                          * For case A, C and D we don't allow a drop. But for
14490                                          * case B where we now our on our steady rate we do
14491                                          * allow a drop.
14492                                          *
14493                                          */
14494                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14495                                                 goto done_w_hdwr;
14496                                 }
14497                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
14498                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14499                                         if (rack_hw_rate_to_low &&
14500                                             (bw_est < rack_hw_rate_to_low)) {
14501                                                 /*
14502                                                  * The pacing rate is too low for hardware, but
14503                                                  * do allow hardware pacing to be restarted.
14504                                                  */
14505                                                 rack_log_hdwr_pacing(rack,
14506                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
14507                                                              0, 5);
14508                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14509                                                 rack->r_ctl.crte = NULL;
14510                                                 rack->rack_attempt_hdwr_pace = 0;
14511                                                 rack->rack_hdrw_pacing = 0;
14512                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14513                                                 goto done_w_hdwr;
14514                                         }
14515                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
14516                                                                    rack->rc_tp,
14517                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
14518                                                                    rate_wanted,
14519                                                                    RS_PACING_GEQ,
14520                                                                    &err, &rack->r_ctl.crte_prev_rate);
14521                                         if (nrte == NULL) {
14522                                                 /* Lost the rate */
14523                                                 rack->rack_hdrw_pacing = 0;
14524                                                 rack->r_ctl.crte = NULL;
14525                                                 rack_log_hdwr_pacing(rack,
14526                                                                      rate_wanted, 0, __LINE__,
14527                                                                      err, 1);
14528                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14529                                                 counter_u64_add(rack_hw_pace_lost, 1);
14530                                         } else if (nrte != rack->r_ctl.crte) {
14531                                                 rack->r_ctl.crte = nrte;
14532                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
14533                                                                                                          segsiz, 0,
14534                                                                                                          rack->r_ctl.crte,
14535                                                                                                          NULL);
14536                                                 rack_log_hdwr_pacing(rack,
14537                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14538                                                                      err, 2);
14539                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
14540                                         }
14541                                 } else {
14542                                         /* We just need to adjust the segment size */
14543                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14544                                         rack_log_hdwr_pacing(rack,
14545                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14546                                                              0, 4);
14547                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14548                                 }
14549                         }
14550                 }
14551                 if ((rack->r_ctl.crte != NULL) &&
14552                     (rack->r_ctl.crte->rate == rate_wanted)) {
14553                         /*
14554                          * We need to add a extra if the rates
14555                          * are exactly matched. The idea is
14556                          * we want the software to make sure the
14557                          * queue is empty before adding more, this
14558                          * gives us N MSS extra pace times where
14559                          * N is our sysctl
14560                          */
14561                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
14562                 }
14563 done_w_hdwr:
14564                 if (rack_limit_time_with_srtt &&
14565                     (rack->use_fixed_rate == 0) &&
14566 #ifdef NETFLIX_PEAKRATE
14567                     (rack->rc_tp->t_maxpeakrate == 0) &&
14568 #endif
14569                     (rack->rack_hdrw_pacing == 0)) {
14570                         /*
14571                          * Sanity check, we do not allow the pacing delay
14572                          * to be longer than the SRTT of the path. If it is
14573                          * a slow path, then adding a packet should increase
14574                          * the RTT and compensate for this i.e. the srtt will
14575                          * be greater so the allowed pacing time will be greater.
14576                          *
14577                          * Note this restriction is not for where a peak rate
14578                          * is set, we are doing fixed pacing or hardware pacing.
14579                          */
14580                         if (rack->rc_tp->t_srtt)
14581                                 srtt = rack->rc_tp->t_srtt;
14582                         else
14583                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
14584                         if (srtt < slot) {
14585                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL);
14586                                 slot = srtt;
14587                         }
14588                 }
14589                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm);
14590         }
14591         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
14592                 /*
14593                  * If this rate is seeing enobufs when it
14594                  * goes to send then either the nic is out
14595                  * of gas or we are mis-estimating the time
14596                  * somehow and not letting the queue empty
14597                  * completely. Lets add to the pacing time.
14598                  */
14599                 int hw_boost_delay;
14600
14601                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
14602                 if (hw_boost_delay > rack_enobuf_hw_max)
14603                         hw_boost_delay = rack_enobuf_hw_max;
14604                 else if (hw_boost_delay < rack_enobuf_hw_min)
14605                         hw_boost_delay = rack_enobuf_hw_min;
14606                 slot += hw_boost_delay;
14607         }
14608         if (slot)
14609                 counter_u64_add(rack_calc_nonzero, 1);
14610         else
14611                 counter_u64_add(rack_calc_zero, 1);
14612         return (slot);
14613 }
14614
14615 static void
14616 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
14617     tcp_seq startseq, uint32_t sb_offset)
14618 {
14619         struct rack_sendmap *my_rsm = NULL;
14620         struct rack_sendmap fe;
14621
14622         if (tp->t_state < TCPS_ESTABLISHED) {
14623                 /*
14624                  * We don't start any measurements if we are
14625                  * not at least established.
14626                  */
14627                 return;
14628         }
14629         tp->t_flags |= TF_GPUTINPROG;
14630         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
14631         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
14632         tp->gput_seq = startseq;
14633         rack->app_limited_needs_set = 0;
14634         if (rack->in_probe_rtt)
14635                 rack->measure_saw_probe_rtt = 1;
14636         else if ((rack->measure_saw_probe_rtt) &&
14637                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
14638                 rack->measure_saw_probe_rtt = 0;
14639         if (rack->rc_gp_filled)
14640                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14641         else {
14642                 /* Special case initial measurement */
14643                 struct timeval tv;
14644
14645                 tp->gput_ts = tcp_get_usecs(&tv);
14646                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14647         }
14648         /*
14649          * We take a guess out into the future,
14650          * if we have no measurement and no
14651          * initial rate, we measure the first
14652          * initial-windows worth of data to
14653          * speed up getting some GP measurement and
14654          * thus start pacing.
14655          */
14656         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
14657                 rack->app_limited_needs_set = 1;
14658                 tp->gput_ack = startseq + max(rc_init_window(rack),
14659                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
14660                 rack_log_pacing_delay_calc(rack,
14661                                            tp->gput_seq,
14662                                            tp->gput_ack,
14663                                            0,
14664                                            tp->gput_ts,
14665                                            rack->r_ctl.rc_app_limited_cnt,
14666                                            9,
14667                                            __LINE__, NULL);
14668                 return;
14669         }
14670         if (sb_offset) {
14671                 /*
14672                  * We are out somewhere in the sb
14673                  * can we use the already outstanding data?
14674                  */
14675
14676                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
14677                         /*
14678                          * Yes first one is good and in this case
14679                          * the tp->gput_ts is correctly set based on
14680                          * the last ack that arrived (no need to
14681                          * set things up when an ack comes in).
14682                          */
14683                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14684                         if ((my_rsm == NULL) ||
14685                             (my_rsm->r_rtr_cnt != 1)) {
14686                                 /* retransmission? */
14687                                 goto use_latest;
14688                         }
14689                 } else {
14690                         if (rack->r_ctl.rc_first_appl == NULL) {
14691                                 /*
14692                                  * If rc_first_appl is NULL
14693                                  * then the cnt should be 0.
14694                                  * This is probably an error, maybe
14695                                  * a KASSERT would be approprate.
14696                                  */
14697                                 goto use_latest;
14698                         }
14699                         /*
14700                          * If we have a marker pointer to the last one that is
14701                          * app limited we can use that, but we need to set
14702                          * things up so that when it gets ack'ed we record
14703                          * the ack time (if its not already acked).
14704                          */
14705                         rack->app_limited_needs_set = 1;
14706                         /*
14707                          * We want to get to the rsm that is either
14708                          * next with space i.e. over 1 MSS or the one
14709                          * after that (after the app-limited).
14710                          */
14711                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14712                                          rack->r_ctl.rc_first_appl);
14713                         if (my_rsm) {
14714                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
14715                                         /* Have to use the next one */
14716                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14717                                                          my_rsm);
14718                                 else {
14719                                         /* Use after the first MSS of it is acked */
14720                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
14721                                         goto start_set;
14722                                 }
14723                         }
14724                         if ((my_rsm == NULL) ||
14725                             (my_rsm->r_rtr_cnt != 1)) {
14726                                 /*
14727                                  * Either its a retransmit or
14728                                  * the last is the app-limited one.
14729                                  */
14730                                 goto use_latest;
14731                         }
14732                 }
14733                 tp->gput_seq = my_rsm->r_start;
14734 start_set:
14735                 if (my_rsm->r_flags & RACK_ACKED) {
14736                         /*
14737                          * This one has been acked use the arrival ack time
14738                          */
14739                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14740                         rack->app_limited_needs_set = 0;
14741                 }
14742                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14743                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
14744                 rack_log_pacing_delay_calc(rack,
14745                                            tp->gput_seq,
14746                                            tp->gput_ack,
14747                                            (uint64_t)my_rsm,
14748                                            tp->gput_ts,
14749                                            rack->r_ctl.rc_app_limited_cnt,
14750                                            9,
14751                                            __LINE__, NULL);
14752                 return;
14753         }
14754
14755 use_latest:
14756         /*
14757          * We don't know how long we may have been
14758          * idle or if this is the first-send. Lets
14759          * setup the flag so we will trim off
14760          * the first ack'd data so we get a true
14761          * measurement.
14762          */
14763         rack->app_limited_needs_set = 1;
14764         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
14765         /* Find this guy so we can pull the send time */
14766         fe.r_start = startseq;
14767         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
14768         if (my_rsm) {
14769                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14770                 if (my_rsm->r_flags & RACK_ACKED) {
14771                         /*
14772                          * Unlikely since its probably what was
14773                          * just transmitted (but I am paranoid).
14774                          */
14775                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14776                         rack->app_limited_needs_set = 0;
14777                 }
14778                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
14779                         /* This also is unlikely */
14780                         tp->gput_seq = my_rsm->r_start;
14781                 }
14782         } else {
14783                 /*
14784                  * TSNH unless we have some send-map limit,
14785                  * and even at that it should not be hitting
14786                  * that limit (we should have stopped sending).
14787                  */
14788                 struct timeval tv;
14789
14790                 microuptime(&tv);
14791                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14792         }
14793         rack_log_pacing_delay_calc(rack,
14794                                    tp->gput_seq,
14795                                    tp->gput_ack,
14796                                    (uint64_t)my_rsm,
14797                                    tp->gput_ts,
14798                                    rack->r_ctl.rc_app_limited_cnt,
14799                                    9, __LINE__, NULL);
14800 }
14801
14802 static inline uint32_t
14803 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
14804     uint32_t avail, int32_t sb_offset)
14805 {
14806         uint32_t len;
14807         uint32_t sendwin;
14808
14809         if (tp->snd_wnd > cwnd_to_use)
14810                 sendwin = cwnd_to_use;
14811         else
14812                 sendwin = tp->snd_wnd;
14813         if (ctf_outstanding(tp) >= tp->snd_wnd) {
14814                 /* We never want to go over our peers rcv-window */
14815                 len = 0;
14816         } else {
14817                 uint32_t flight;
14818
14819                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
14820                 if (flight >= sendwin) {
14821                         /*
14822                          * We have in flight what we are allowed by cwnd (if
14823                          * it was rwnd blocking it would have hit above out
14824                          * >= tp->snd_wnd).
14825                          */
14826                         return (0);
14827                 }
14828                 len = sendwin - flight;
14829                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
14830                         /* We would send too much (beyond the rwnd) */
14831                         len = tp->snd_wnd - ctf_outstanding(tp);
14832                 }
14833                 if ((len + sb_offset) > avail) {
14834                         /*
14835                          * We don't have that much in the SB, how much is
14836                          * there?
14837                          */
14838                         len = avail - sb_offset;
14839                 }
14840         }
14841         return (len);
14842 }
14843
14844 static void
14845 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
14846              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
14847              int rsm_is_null, int optlen, int line, uint16_t mode)
14848 {
14849         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14850                 union tcp_log_stackspecific log;
14851                 struct timeval tv;
14852
14853                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14854                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
14855                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
14856                 log.u_bbr.flex1 = error;
14857                 log.u_bbr.flex2 = flags;
14858                 log.u_bbr.flex3 = rsm_is_null;
14859                 log.u_bbr.flex4 = ipoptlen;
14860                 log.u_bbr.flex5 = tp->rcv_numsacks;
14861                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
14862                 log.u_bbr.flex7 = optlen;
14863                 log.u_bbr.flex8 = rack->r_fsb_inited;
14864                 log.u_bbr.applimited = rack->r_fast_output;
14865                 log.u_bbr.bw_inuse = rack_get_bw(rack);
14866                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
14867                 log.u_bbr.cwnd_gain = mode;
14868                 log.u_bbr.pkts_out = orig_len;
14869                 log.u_bbr.lt_epoch = len;
14870                 log.u_bbr.delivered = line;
14871                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14872                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14873                 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
14874                                len, &log, false, NULL, NULL, 0, &tv);
14875         }
14876 }
14877
14878
14879 static struct mbuf *
14880 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
14881                    struct rack_fast_send_blk *fsb,
14882                    int32_t seglimit, int32_t segsize, int hw_tls)
14883 {
14884 #ifdef KERN_TLS
14885         struct ktls_session *tls, *ntls;
14886         struct mbuf *start;
14887 #endif
14888         struct mbuf *m, *n, **np, *smb;
14889         struct mbuf *top;
14890         int32_t off, soff;
14891         int32_t len = *plen;
14892         int32_t fragsize;
14893         int32_t len_cp = 0;
14894         uint32_t mlen, frags;
14895
14896         soff = off = the_off;
14897         smb = m = the_m;
14898         np = &top;
14899         top = NULL;
14900 #ifdef KERN_TLS
14901         if (hw_tls && (m->m_flags & M_EXTPG))
14902                 tls = m->m_epg_tls;
14903         else
14904                 tls = NULL;
14905         start = m;
14906 #endif
14907         while (len > 0) {
14908                 if (m == NULL) {
14909                         *plen = len_cp;
14910                         break;
14911                 }
14912 #ifdef KERN_TLS
14913                 if (hw_tls) {
14914                         if (m->m_flags & M_EXTPG)
14915                                 ntls = m->m_epg_tls;
14916                         else
14917                                 ntls = NULL;
14918
14919                         /*
14920                          * Avoid mixing TLS records with handshake
14921                          * data or TLS records from different
14922                          * sessions.
14923                          */
14924                         if (tls != ntls) {
14925                                 MPASS(m != start);
14926                                 *plen = len_cp;
14927                                 break;
14928                         }
14929                 }
14930 #endif
14931                 mlen = min(len, m->m_len - off);
14932                 if (seglimit) {
14933                         /*
14934                          * For M_EXTPG mbufs, add 3 segments
14935                          * + 1 in case we are crossing page boundaries
14936                          * + 2 in case the TLS hdr/trailer are used
14937                          * It is cheaper to just add the segments
14938                          * than it is to take the cache miss to look
14939                          * at the mbuf ext_pgs state in detail.
14940                          */
14941                         if (m->m_flags & M_EXTPG) {
14942                                 fragsize = min(segsize, PAGE_SIZE);
14943                                 frags = 3;
14944                         } else {
14945                                 fragsize = segsize;
14946                                 frags = 0;
14947                         }
14948
14949                         /* Break if we really can't fit anymore. */
14950                         if ((frags + 1) >= seglimit) {
14951                                 *plen = len_cp;
14952                                 break;
14953                         }
14954
14955                         /*
14956                          * Reduce size if you can't copy the whole
14957                          * mbuf. If we can't copy the whole mbuf, also
14958                          * adjust len so the loop will end after this
14959                          * mbuf.
14960                          */
14961                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
14962                                 mlen = (seglimit - frags - 1) * fragsize;
14963                                 len = mlen;
14964                                 *plen = len_cp + len;
14965                         }
14966                         frags += howmany(mlen, fragsize);
14967                         if (frags == 0)
14968                                 frags++;
14969                         seglimit -= frags;
14970                         KASSERT(seglimit > 0,
14971                             ("%s: seglimit went too low", __func__));
14972                 }
14973                 n = m_get(M_NOWAIT, m->m_type);
14974                 *np = n;
14975                 if (n == NULL)
14976                         goto nospace;
14977                 n->m_len = mlen;
14978                 soff += mlen;
14979                 len_cp += n->m_len;
14980                 if (m->m_flags & (M_EXT|M_EXTPG)) {
14981                         n->m_data = m->m_data + off;
14982                         mb_dupcl(n, m);
14983                 } else {
14984                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
14985                             (u_int)n->m_len);
14986                 }
14987                 len -= n->m_len;
14988                 off = 0;
14989                 m = m->m_next;
14990                 np = &n->m_next;
14991                 if (len || (soff == smb->m_len)) {
14992                         /*
14993                          * We have more so we move forward  or
14994                          * we have consumed the entire mbuf and
14995                          * len has fell to 0.
14996                          */
14997                         soff = 0;
14998                         smb = m;
14999                 }
15000
15001         }
15002         if (fsb != NULL) {
15003                 fsb->m = smb;
15004                 fsb->off = soff;
15005                 if (smb) {
15006                         /*
15007                          * Save off the size of the mbuf. We do
15008                          * this so that we can recognize when it
15009                          * has been trimmed by sbcut() as acks
15010                          * come in.
15011                          */
15012                         fsb->o_m_len = smb->m_len;
15013                 } else {
15014                         /*
15015                          * This is the case where the next mbuf went to NULL. This
15016                          * means with this copy we have sent everything in the sb.
15017                          * In theory we could clear the fast_output flag, but lets
15018                          * not since its possible that we could get more added
15019                          * and acks that call the extend function which would let
15020                          * us send more.
15021                          */
15022                         fsb->o_m_len = 0;
15023                 }
15024         }
15025         return (top);
15026 nospace:
15027         if (top)
15028                 m_freem(top);
15029         return (NULL);
15030
15031 }
15032
15033 /*
15034  * This is a copy of m_copym(), taking the TSO segment size/limit
15035  * constraints into account, and advancing the sndptr as it goes.
15036  */
15037 static struct mbuf *
15038 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15039                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15040 {
15041         struct mbuf *m, *n;
15042         int32_t soff;
15043
15044         soff = rack->r_ctl.fsb.off;
15045         m = rack->r_ctl.fsb.m;
15046         if (rack->r_ctl.fsb.o_m_len != m->m_len) {
15047                 /*
15048                  * The mbuf had the front of it chopped off by an ack
15049                  * we need to adjust the soff/off by that difference.
15050                  */
15051                 uint32_t delta;
15052
15053                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15054                 soff -= delta;
15055         }
15056         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15057         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15058         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15059                                  __FUNCTION__,
15060                                  rack, *plen, m, m->m_len));
15061         /* Save off the right location before we copy and advance */
15062         *s_soff = soff;
15063         *s_mb = rack->r_ctl.fsb.m;
15064         n = rack_fo_base_copym(m, soff, plen,
15065                                &rack->r_ctl.fsb,
15066                                seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15067         return (n);
15068 }
15069
15070 static int
15071 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15072                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len)
15073 {
15074         /*
15075          * Enter the fast retransmit path. We are given that a sched_pin is
15076          * in place (if accounting is compliled in) and the cycle count taken
15077          * at the entry is in the ts_val. The concept her is that the rsm
15078          * now holds the mbuf offsets and such so we can directly transmit
15079          * without a lot of overhead, the len field is already set for
15080          * us to prohibit us from sending too much (usually its 1MSS).
15081          */
15082         struct ip *ip = NULL;
15083         struct udphdr *udp = NULL;
15084         struct tcphdr *th = NULL;
15085         struct mbuf *m = NULL;
15086         struct inpcb *inp;
15087         uint8_t *cpto;
15088         struct tcp_log_buffer *lgb;
15089 #ifdef TCP_ACCOUNTING
15090         uint64_t crtsc;
15091         int cnt_thru = 1;
15092 #endif
15093         int doing_tlp = 0;
15094         struct tcpopt to;
15095         u_char opt[TCP_MAXOLEN];
15096         uint32_t hdrlen, optlen;
15097         int32_t slot, segsiz, max_val, tso = 0, error, flags, ulen = 0;
15098         uint32_t us_cts;
15099         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15100         uint32_t if_hw_tsomaxsegsize;
15101
15102 #ifdef INET6
15103         struct ip6_hdr *ip6 = NULL;
15104
15105         if (rack->r_is_v6) {
15106                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15107                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15108         } else
15109 #endif                          /* INET6 */
15110         {
15111                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15112                 hdrlen = sizeof(struct tcpiphdr);
15113         }
15114         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15115                 goto failed;
15116         }
15117         if (rsm->r_flags & RACK_TLP)
15118                 doing_tlp = 1;
15119         startseq = rsm->r_start;
15120         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15121         inp = rack->rc_inp;
15122         to.to_flags = 0;
15123         flags = tcp_outflags[tp->t_state];
15124         if (flags & (TH_SYN|TH_RST)) {
15125                 goto failed;
15126         }
15127         if (rsm->r_flags & RACK_HAS_FIN) {
15128                 /* We can't send a FIN here */
15129                 goto failed;
15130         }
15131         if (flags & TH_FIN) {
15132                 /* We never send a FIN */
15133                 flags &= ~TH_FIN;
15134         }
15135         if (tp->t_flags & TF_RCVD_TSTMP) {
15136                 to.to_tsval = ms_cts + tp->ts_offset;
15137                 to.to_tsecr = tp->ts_recent;
15138                 to.to_flags = TOF_TS;
15139         }
15140         optlen = tcp_addoptions(&to, opt);
15141         hdrlen += optlen;
15142         udp = rack->r_ctl.fsb.udp;
15143         if (udp)
15144                 hdrlen += sizeof(struct udphdr);
15145         if (rack->r_ctl.rc_pace_max_segs)
15146                 max_val = rack->r_ctl.rc_pace_max_segs;
15147         else if (rack->rc_user_set_max_segs)
15148                 max_val = rack->rc_user_set_max_segs * segsiz;
15149         else
15150                 max_val = len;
15151         if ((tp->t_flags & TF_TSO) &&
15152             V_tcp_do_tso &&
15153             (len > segsiz) &&
15154             (tp->t_port == 0))
15155                 tso = 1;
15156 #ifdef INET6
15157         if (MHLEN < hdrlen + max_linkhdr)
15158                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15159         else
15160 #endif
15161                 m = m_gethdr(M_NOWAIT, MT_DATA);
15162         if (m == NULL)
15163                 goto failed;
15164         m->m_data += max_linkhdr;
15165         m->m_len = hdrlen;
15166         th = rack->r_ctl.fsb.th;
15167         /* Establish the len to send */
15168         if (len > max_val)
15169                 len = max_val;
15170         if ((tso) && (len + optlen > tp->t_maxseg)) {
15171                 uint32_t if_hw_tsomax;
15172                 int32_t max_len;
15173
15174                 /* extract TSO information */
15175                 if_hw_tsomax = tp->t_tsomax;
15176                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15177                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15178                 /*
15179                  * Check if we should limit by maximum payload
15180                  * length:
15181                  */
15182                 if (if_hw_tsomax != 0) {
15183                         /* compute maximum TSO length */
15184                         max_len = (if_hw_tsomax - hdrlen -
15185                                    max_linkhdr);
15186                         if (max_len <= 0) {
15187                                 goto failed;
15188                         } else if (len > max_len) {
15189                                 len = max_len;
15190                         }
15191                 }
15192                 if (len <= segsiz) {
15193                         /*
15194                          * In case there are too many small fragments don't
15195                          * use TSO:
15196                          */
15197                         tso = 0;
15198                 }
15199         } else {
15200                 tso = 0;
15201         }
15202         if ((tso == 0) && (len > segsiz))
15203                 len = segsiz;
15204         us_cts = tcp_get_usecs(tv);
15205         if ((len == 0) ||
15206             (len <= MHLEN - hdrlen - max_linkhdr)) {
15207                 goto failed;
15208         }
15209         th->th_seq = htonl(rsm->r_start);
15210         th->th_ack = htonl(tp->rcv_nxt);
15211         /*
15212          * The PUSH bit should only be applied
15213          * if the full retransmission is made. If
15214          * we are sending less than this is the
15215          * left hand edge and should not have
15216          * the PUSH bit.
15217          */
15218         if ((rsm->r_flags & RACK_HAD_PUSH) &&
15219             (len == (rsm->r_end - rsm->r_start)))
15220                 flags |= TH_PUSH;
15221         th->th_flags = flags;
15222         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15223         if (th->th_win == 0) {
15224                 tp->t_sndzerowin++;
15225                 tp->t_flags |= TF_RXWIN0SENT;
15226         } else
15227                 tp->t_flags &= ~TF_RXWIN0SENT;
15228         if (rsm->r_flags & RACK_TLP) {
15229                 /*
15230                  * TLP should not count in retran count, but
15231                  * in its own bin
15232                  */
15233                 counter_u64_add(rack_tlp_retran, 1);
15234                 counter_u64_add(rack_tlp_retran_bytes, len);
15235         } else {
15236                 tp->t_sndrexmitpack++;
15237                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15238                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15239         }
15240 #ifdef STATS
15241         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15242                                  len);
15243 #endif
15244         if (rsm->m == NULL)
15245                 goto failed;
15246         if (rsm->orig_m_len != rsm->m->m_len) {
15247                 /* Fix up the orig_m_len and possibly the mbuf offset */
15248                 rack_adjust_orig_mlen(rsm);
15249         }
15250         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15251         if (len <= segsiz) {
15252                 /*
15253                  * Must have ran out of mbufs for the copy
15254                  * shorten it to no longer need tso. Lets
15255                  * not put on sendalot since we are low on
15256                  * mbufs.
15257                  */
15258                 tso = 0;
15259         }
15260         if ((m->m_next == NULL) || (len <= 0)){
15261                 goto failed;
15262         }
15263         if (udp) {
15264                 if (rack->r_is_v6)
15265                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15266                 else
15267                         ulen = hdrlen + len - sizeof(struct ip);
15268                 udp->uh_ulen = htons(ulen);
15269         }
15270         m->m_pkthdr.rcvif = (struct ifnet *)0;
15271         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15272 #ifdef INET6
15273         if (rack->r_is_v6) {
15274                 if (tp->t_port) {
15275                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15276                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15277                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15278                         th->th_sum = htons(0);
15279                         UDPSTAT_INC(udps_opackets);
15280                 } else {
15281                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15282                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15283                         th->th_sum = in6_cksum_pseudo(ip6,
15284                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15285                                                       0);
15286                 }
15287         }
15288 #endif
15289 #if defined(INET6) && defined(INET)
15290         else
15291 #endif
15292 #ifdef INET
15293         {
15294                 if (tp->t_port) {
15295                         m->m_pkthdr.csum_flags = CSUM_UDP;
15296                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15297                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15298                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15299                         th->th_sum = htons(0);
15300                         UDPSTAT_INC(udps_opackets);
15301                 } else {
15302                         m->m_pkthdr.csum_flags = CSUM_TCP;
15303                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15304                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15305                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15306                                                                         IPPROTO_TCP + len + optlen));
15307                 }
15308                 /* IP version must be set here for ipv4/ipv6 checking later */
15309                 KASSERT(ip->ip_v == IPVERSION,
15310                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15311         }
15312 #endif
15313         if (tso) {
15314                 KASSERT(len > tp->t_maxseg - optlen,
15315                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15316                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15317                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15318         }
15319 #ifdef INET6
15320         if (rack->r_is_v6) {
15321                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15322                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15323                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15324                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15325                 else
15326                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15327         }
15328 #endif
15329 #if defined(INET) && defined(INET6)
15330         else
15331 #endif
15332 #ifdef INET
15333         {
15334                 ip->ip_len = htons(m->m_pkthdr.len);
15335                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15336                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15337                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15338                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15339                                 ip->ip_off |= htons(IP_DF);
15340                         }
15341                 } else {
15342                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15343                 }
15344         }
15345 #endif
15346         /* Time to copy in our header */
15347         cpto = mtod(m, uint8_t *);
15348         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15349         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15350         if (optlen) {
15351                 bcopy(opt, th + 1, optlen);
15352                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15353         } else {
15354                 th->th_off = sizeof(struct tcphdr) >> 2;
15355         }
15356         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15357                 union tcp_log_stackspecific log;
15358
15359                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15360                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15361                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15362                 if (rack->rack_no_prr)
15363                         log.u_bbr.flex1 = 0;
15364                 else
15365                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15366                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15367                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15368                 log.u_bbr.flex4 = max_val;
15369                 log.u_bbr.flex5 = 0;
15370                 /* Save off the early/late values */
15371                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15372                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15373                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15374                 log.u_bbr.flex8 = 1;
15375                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15376                 log.u_bbr.flex7 = 55;
15377                 log.u_bbr.pkts_out = tp->t_maxseg;
15378                 log.u_bbr.timeStamp = cts;
15379                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15380                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15381                 log.u_bbr.delivered = 0;
15382                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15383                                      len, &log, false, NULL, NULL, 0, tv);
15384         } else
15385                 lgb = NULL;
15386 #ifdef INET6
15387         if (rack->r_is_v6) {
15388                 error = ip6_output(m, NULL,
15389                                    &inp->inp_route6,
15390                                    0, NULL, NULL, inp);
15391         }
15392 #endif
15393 #if defined(INET) && defined(INET6)
15394         else
15395 #endif
15396 #ifdef INET
15397         {
15398                 error = ip_output(m, NULL,
15399                                   &inp->inp_route,
15400                                   0, 0, inp);
15401         }
15402 #endif
15403         m = NULL;
15404         if (lgb) {
15405                 lgb->tlb_errno = error;
15406                 lgb = NULL;
15407         }
15408         if (error) {
15409                 goto failed;
15410         }
15411         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15412                         rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
15413         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15414                 rack->rc_tlp_in_progress = 1;
15415                 rack->r_ctl.rc_tlp_cnt_out++;
15416         }
15417         if (error == 0)
15418                 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
15419         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15420         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15421         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15422                 rack->r_ctl.retran_during_recovery += len;
15423         {
15424                 int idx;
15425
15426                 idx = (len / segsiz) + 3;
15427                 if (idx >= TCP_MSS_ACCT_ATIMER)
15428                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15429                 else
15430                         counter_u64_add(rack_out_size[idx], 1);
15431         }
15432         if (tp->t_rtttime == 0) {
15433                 tp->t_rtttime = ticks;
15434                 tp->t_rtseq = startseq;
15435                 KMOD_TCPSTAT_INC(tcps_segstimed);
15436         }
15437         counter_u64_add(rack_fto_rsm_send, 1);
15438         if (error && (error == ENOBUFS)) {
15439                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15440                 if (rack->rc_enobuf < 0x7f)
15441                         rack->rc_enobuf++;
15442                 if (slot < (10 * HPTS_USEC_IN_MSEC))
15443                         slot = 10 * HPTS_USEC_IN_MSEC;
15444         } else
15445                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15446         if ((slot == 0) ||
15447             (rack->rc_always_pace == 0) ||
15448             (rack->r_rr_config == 1)) {
15449                 /*
15450                  * We have no pacing set or we
15451                  * are using old-style rack or
15452                  * we are overriden to use the old 1ms pacing.
15453                  */
15454                 slot = rack->r_ctl.rc_min_to;
15455         }
15456         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
15457         if (rack->r_must_retran) {
15458                 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
15459                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
15460                         /*
15461                          * We have retransmitted all we need.
15462                          */
15463                         rack->r_must_retran = 0;
15464                         rack->r_ctl.rc_out_at_rto = 0;
15465                 }
15466         }
15467 #ifdef TCP_ACCOUNTING
15468         crtsc = get_cyclecount();
15469         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15470                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15471         }
15472         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15473         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15474                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15475         }
15476         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15477         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15478                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
15479         }
15480         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
15481         sched_unpin();
15482 #endif
15483         return (0);
15484 failed:
15485         if (m)
15486                 m_free(m);
15487         return (-1);
15488 }
15489
15490 static void
15491 rack_sndbuf_autoscale(struct tcp_rack *rack)
15492 {
15493         /*
15494          * Automatic sizing of send socket buffer.  Often the send buffer
15495          * size is not optimally adjusted to the actual network conditions
15496          * at hand (delay bandwidth product).  Setting the buffer size too
15497          * small limits throughput on links with high bandwidth and high
15498          * delay (eg. trans-continental/oceanic links).  Setting the
15499          * buffer size too big consumes too much real kernel memory,
15500          * especially with many connections on busy servers.
15501          *
15502          * The criteria to step up the send buffer one notch are:
15503          *  1. receive window of remote host is larger than send buffer
15504          *     (with a fudge factor of 5/4th);
15505          *  2. send buffer is filled to 7/8th with data (so we actually
15506          *     have data to make use of it);
15507          *  3. send buffer fill has not hit maximal automatic size;
15508          *  4. our send window (slow start and cogestion controlled) is
15509          *     larger than sent but unacknowledged data in send buffer.
15510          *
15511          * Note that the rack version moves things much faster since
15512          * we want to avoid hitting cache lines in the rack_fast_output()
15513          * path so this is called much less often and thus moves
15514          * the SB forward by a percentage.
15515          */
15516         struct socket *so;
15517         struct tcpcb *tp;
15518         uint32_t sendwin, scaleup;
15519
15520         tp = rack->rc_tp;
15521         so = rack->rc_inp->inp_socket;
15522         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
15523         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
15524                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
15525                     sbused(&so->so_snd) >=
15526                     (so->so_snd.sb_hiwat / 8 * 7) &&
15527                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
15528                     sendwin >= (sbused(&so->so_snd) -
15529                     (tp->snd_nxt - tp->snd_una))) {
15530                         if (rack_autosndbuf_inc)
15531                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
15532                         else
15533                                 scaleup = V_tcp_autosndbuf_inc;
15534                         if (scaleup < V_tcp_autosndbuf_inc)
15535                                 scaleup = V_tcp_autosndbuf_inc;
15536                         scaleup += so->so_snd.sb_hiwat;
15537                         if (scaleup > V_tcp_autosndbuf_max)
15538                                 scaleup = V_tcp_autosndbuf_max;
15539                         if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
15540                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
15541                 }
15542         }
15543 }
15544
15545 static int
15546 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
15547                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
15548 {
15549         /*
15550          * Enter to do fast output. We are given that the sched_pin is
15551          * in place (if accounting is compiled in) and the cycle count taken
15552          * at entry is in place in ts_val. The idea here is that
15553          * we know how many more bytes needs to be sent (presumably either
15554          * during pacing or to fill the cwnd and that was greater than
15555          * the max-burst). We have how much to send and all the info we
15556          * need to just send.
15557          */
15558         struct ip *ip = NULL;
15559         struct udphdr *udp = NULL;
15560         struct tcphdr *th = NULL;
15561         struct mbuf *m, *s_mb;
15562         struct inpcb *inp;
15563         uint8_t *cpto;
15564         struct tcp_log_buffer *lgb;
15565 #ifdef TCP_ACCOUNTING
15566         uint64_t crtsc;
15567 #endif
15568         struct tcpopt to;
15569         u_char opt[TCP_MAXOLEN];
15570         uint32_t hdrlen, optlen;
15571         int cnt_thru = 1;
15572         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, flags, ulen = 0;
15573         uint32_t us_cts, s_soff;
15574         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15575         uint32_t if_hw_tsomaxsegsize;
15576         uint16_t add_flag = RACK_SENT_FP;
15577 #ifdef INET6
15578         struct ip6_hdr *ip6 = NULL;
15579
15580         if (rack->r_is_v6) {
15581                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15582                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15583         } else
15584 #endif                          /* INET6 */
15585         {
15586                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15587                 hdrlen = sizeof(struct tcpiphdr);
15588         }
15589         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15590                 m = NULL;
15591                 goto failed;
15592         }
15593         startseq = tp->snd_max;
15594         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15595         inp = rack->rc_inp;
15596         len = rack->r_ctl.fsb.left_to_send;
15597         to.to_flags = 0;
15598         flags = rack->r_ctl.fsb.tcp_flags;
15599         if (tp->t_flags & TF_RCVD_TSTMP) {
15600                 to.to_tsval = ms_cts + tp->ts_offset;
15601                 to.to_tsecr = tp->ts_recent;
15602                 to.to_flags = TOF_TS;
15603         }
15604         optlen = tcp_addoptions(&to, opt);
15605         hdrlen += optlen;
15606         udp = rack->r_ctl.fsb.udp;
15607         if (udp)
15608                 hdrlen += sizeof(struct udphdr);
15609         if (rack->r_ctl.rc_pace_max_segs)
15610                 max_val = rack->r_ctl.rc_pace_max_segs;
15611         else if (rack->rc_user_set_max_segs)
15612                 max_val = rack->rc_user_set_max_segs * segsiz;
15613         else
15614                 max_val = len;
15615         if ((tp->t_flags & TF_TSO) &&
15616             V_tcp_do_tso &&
15617             (len > segsiz) &&
15618             (tp->t_port == 0))
15619                 tso = 1;
15620 again:
15621 #ifdef INET6
15622         if (MHLEN < hdrlen + max_linkhdr)
15623                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15624         else
15625 #endif
15626                 m = m_gethdr(M_NOWAIT, MT_DATA);
15627         if (m == NULL)
15628                 goto failed;
15629         m->m_data += max_linkhdr;
15630         m->m_len = hdrlen;
15631         th = rack->r_ctl.fsb.th;
15632         /* Establish the len to send */
15633         if (len > max_val)
15634                 len = max_val;
15635         if ((tso) && (len + optlen > tp->t_maxseg)) {
15636                 uint32_t if_hw_tsomax;
15637                 int32_t max_len;
15638
15639                 /* extract TSO information */
15640                 if_hw_tsomax = tp->t_tsomax;
15641                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15642                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15643                 /*
15644                  * Check if we should limit by maximum payload
15645                  * length:
15646                  */
15647                 if (if_hw_tsomax != 0) {
15648                         /* compute maximum TSO length */
15649                         max_len = (if_hw_tsomax - hdrlen -
15650                                    max_linkhdr);
15651                         if (max_len <= 0) {
15652                                 goto failed;
15653                         } else if (len > max_len) {
15654                                 len = max_len;
15655                         }
15656                 }
15657                 if (len <= segsiz) {
15658                         /*
15659                          * In case there are too many small fragments don't
15660                          * use TSO:
15661                          */
15662                         tso = 0;
15663                 }
15664         } else {
15665                 tso = 0;
15666         }
15667         if ((tso == 0) && (len > segsiz))
15668                 len = segsiz;
15669         us_cts = tcp_get_usecs(tv);
15670         if ((len == 0) ||
15671             (len <= MHLEN - hdrlen - max_linkhdr)) {
15672                 goto failed;
15673         }
15674         sb_offset = tp->snd_max - tp->snd_una;
15675         th->th_seq = htonl(tp->snd_max);
15676         th->th_ack = htonl(tp->rcv_nxt);
15677         th->th_flags = flags;
15678         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15679         if (th->th_win == 0) {
15680                 tp->t_sndzerowin++;
15681                 tp->t_flags |= TF_RXWIN0SENT;
15682         } else
15683                 tp->t_flags &= ~TF_RXWIN0SENT;
15684         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
15685         KMOD_TCPSTAT_INC(tcps_sndpack);
15686         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
15687 #ifdef STATS
15688         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
15689                                  len);
15690 #endif
15691         if (rack->r_ctl.fsb.m == NULL)
15692                 goto failed;
15693
15694         /* s_mb and s_soff are saved for rack_log_output */
15695         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
15696                                     &s_mb, &s_soff);
15697         if (len <= segsiz) {
15698                 /*
15699                  * Must have ran out of mbufs for the copy
15700                  * shorten it to no longer need tso. Lets
15701                  * not put on sendalot since we are low on
15702                  * mbufs.
15703                  */
15704                 tso = 0;
15705         }
15706         if (rack->r_ctl.fsb.rfo_apply_push &&
15707             (len == rack->r_ctl.fsb.left_to_send)) {
15708                 th->th_flags |= TH_PUSH;
15709                 add_flag |= RACK_HAD_PUSH;
15710         }
15711         if ((m->m_next == NULL) || (len <= 0)){
15712                 goto failed;
15713         }
15714         if (udp) {
15715                 if (rack->r_is_v6)
15716                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15717                 else
15718                         ulen = hdrlen + len - sizeof(struct ip);
15719                 udp->uh_ulen = htons(ulen);
15720         }
15721         m->m_pkthdr.rcvif = (struct ifnet *)0;
15722         if (tp->t_state == TCPS_ESTABLISHED &&
15723             (tp->t_flags2 & TF2_ECN_PERMIT)) {
15724                 /*
15725                  * If the peer has ECN, mark data packets with ECN capable
15726                  * transmission (ECT). Ignore pure ack packets,
15727                  * retransmissions.
15728                  */
15729                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) {
15730 #ifdef INET6
15731                         if (rack->r_is_v6)
15732                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
15733                         else
15734 #endif
15735                                 ip->ip_tos |= IPTOS_ECN_ECT0;
15736                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
15737                         /*
15738                          * Reply with proper ECN notifications.
15739                          * Only set CWR on new data segments.
15740                          */
15741                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
15742                                 flags |= TH_CWR;
15743                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
15744                         }
15745                 }
15746                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
15747                         flags |= TH_ECE;
15748         }
15749         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15750 #ifdef INET6
15751         if (rack->r_is_v6) {
15752                 if (tp->t_port) {
15753                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15754                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15755                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15756                         th->th_sum = htons(0);
15757                         UDPSTAT_INC(udps_opackets);
15758                 } else {
15759                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15760                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15761                         th->th_sum = in6_cksum_pseudo(ip6,
15762                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15763                                                       0);
15764                 }
15765         }
15766 #endif
15767 #if defined(INET6) && defined(INET)
15768         else
15769 #endif
15770 #ifdef INET
15771         {
15772                 if (tp->t_port) {
15773                         m->m_pkthdr.csum_flags = CSUM_UDP;
15774                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15775                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15776                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15777                         th->th_sum = htons(0);
15778                         UDPSTAT_INC(udps_opackets);
15779                 } else {
15780                         m->m_pkthdr.csum_flags = CSUM_TCP;
15781                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15782                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15783                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15784                                                                         IPPROTO_TCP + len + optlen));
15785                 }
15786                 /* IP version must be set here for ipv4/ipv6 checking later */
15787                 KASSERT(ip->ip_v == IPVERSION,
15788                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15789         }
15790 #endif
15791         if (tso) {
15792                 KASSERT(len > tp->t_maxseg - optlen,
15793                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15794                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15795                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15796         }
15797 #ifdef INET6
15798         if (rack->r_is_v6) {
15799                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15800                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15801                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15802                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15803                 else
15804                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15805         }
15806 #endif
15807 #if defined(INET) && defined(INET6)
15808         else
15809 #endif
15810 #ifdef INET
15811         {
15812                 ip->ip_len = htons(m->m_pkthdr.len);
15813                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15814                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15815                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15816                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15817                                 ip->ip_off |= htons(IP_DF);
15818                         }
15819                 } else {
15820                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15821                 }
15822         }
15823 #endif
15824         /* Time to copy in our header */
15825         cpto = mtod(m, uint8_t *);
15826         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15827         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15828         if (optlen) {
15829                 bcopy(opt, th + 1, optlen);
15830                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15831         } else {
15832                 th->th_off = sizeof(struct tcphdr) >> 2;
15833         }
15834         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15835                 union tcp_log_stackspecific log;
15836
15837                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15838                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15839                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15840                 if (rack->rack_no_prr)
15841                         log.u_bbr.flex1 = 0;
15842                 else
15843                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15844                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15845                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15846                 log.u_bbr.flex4 = max_val;
15847                 log.u_bbr.flex5 = 0;
15848                 /* Save off the early/late values */
15849                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15850                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15851                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15852                 log.u_bbr.flex8 = 0;
15853                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15854                 log.u_bbr.flex7 = 44;
15855                 log.u_bbr.pkts_out = tp->t_maxseg;
15856                 log.u_bbr.timeStamp = cts;
15857                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15858                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15859                 log.u_bbr.delivered = 0;
15860                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15861                                      len, &log, false, NULL, NULL, 0, tv);
15862         } else
15863                 lgb = NULL;
15864 #ifdef INET6
15865         if (rack->r_is_v6) {
15866                 error = ip6_output(m, NULL,
15867                                    &inp->inp_route6,
15868                                    0, NULL, NULL, inp);
15869         }
15870 #endif
15871 #if defined(INET) && defined(INET6)
15872         else
15873 #endif
15874 #ifdef INET
15875         {
15876                 error = ip_output(m, NULL,
15877                                   &inp->inp_route,
15878                                   0, 0, inp);
15879         }
15880 #endif
15881         if (lgb) {
15882                 lgb->tlb_errno = error;
15883                 lgb = NULL;
15884         }
15885         if (error) {
15886                 *send_err = error;
15887                 m = NULL;
15888                 goto failed;
15889         }
15890         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
15891                         NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
15892         m = NULL;
15893         if (tp->snd_una == tp->snd_max) {
15894                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
15895                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
15896                 tp->t_acktime = ticks;
15897         }
15898         if (error == 0)
15899                 tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
15900
15901         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15902         tot_len += len;
15903         if ((tp->t_flags & TF_GPUTINPROG) == 0)
15904                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
15905         tp->snd_max += len;
15906         tp->snd_nxt = tp->snd_max;
15907         {
15908                 int idx;
15909
15910                 idx = (len / segsiz) + 3;
15911                 if (idx >= TCP_MSS_ACCT_ATIMER)
15912                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15913                 else
15914                         counter_u64_add(rack_out_size[idx], 1);
15915         }
15916         if (len <= rack->r_ctl.fsb.left_to_send)
15917                 rack->r_ctl.fsb.left_to_send -= len;
15918         else
15919                 rack->r_ctl.fsb.left_to_send = 0;
15920         if (rack->r_ctl.fsb.left_to_send < segsiz) {
15921                 rack->r_fast_output = 0;
15922                 rack->r_ctl.fsb.left_to_send = 0;
15923                 /* At the end of fast_output scale up the sb */
15924                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
15925                 rack_sndbuf_autoscale(rack);
15926                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
15927         }
15928         if (tp->t_rtttime == 0) {
15929                 tp->t_rtttime = ticks;
15930                 tp->t_rtseq = startseq;
15931                 KMOD_TCPSTAT_INC(tcps_segstimed);
15932         }
15933         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
15934             (max_val > len) &&
15935             (tso == 0)) {
15936                 max_val -= len;
15937                 len = segsiz;
15938                 th = rack->r_ctl.fsb.th;
15939                 cnt_thru++;
15940                 goto again;
15941         }
15942         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15943         counter_u64_add(rack_fto_send, 1);
15944         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
15945         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
15946 #ifdef TCP_ACCOUNTING
15947         crtsc = get_cyclecount();
15948         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15949                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15950         }
15951         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15952         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15953                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15954         }
15955         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15956         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15957                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
15958         }
15959         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
15960         sched_unpin();
15961 #endif
15962         return (0);
15963 failed:
15964         if (m)
15965                 m_free(m);
15966         rack->r_fast_output = 0;
15967         return (-1);
15968 }
15969
15970 static int
15971 rack_output(struct tcpcb *tp)
15972 {
15973         struct socket *so;
15974         uint32_t recwin;
15975         uint32_t sb_offset, s_moff = 0;
15976         int32_t len, flags, error = 0;
15977         struct mbuf *m, *s_mb = NULL;
15978         struct mbuf *mb;
15979         uint32_t if_hw_tsomaxsegcount = 0;
15980         uint32_t if_hw_tsomaxsegsize;
15981         int32_t segsiz, minseg;
15982         long tot_len_this_send = 0;
15983 #ifdef INET
15984         struct ip *ip = NULL;
15985 #endif
15986 #ifdef TCPDEBUG
15987         struct ipovly *ipov = NULL;
15988 #endif
15989         struct udphdr *udp = NULL;
15990         struct tcp_rack *rack;
15991         struct tcphdr *th;
15992         uint8_t pass = 0;
15993         uint8_t mark = 0;
15994         uint8_t wanted_cookie = 0;
15995         u_char opt[TCP_MAXOLEN];
15996         unsigned ipoptlen, optlen, hdrlen, ulen=0;
15997         uint32_t rack_seq;
15998
15999 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16000         unsigned ipsec_optlen = 0;
16001
16002 #endif
16003         int32_t idle, sendalot;
16004         int32_t sub_from_prr = 0;
16005         volatile int32_t sack_rxmit;
16006         struct rack_sendmap *rsm = NULL;
16007         int32_t tso, mtu;
16008         struct tcpopt to;
16009         int32_t slot = 0;
16010         int32_t sup_rack = 0;
16011         uint32_t cts, ms_cts, delayed, early;
16012         uint16_t add_flag = RACK_SENT_SP;
16013         uint8_t hpts_calling,  doing_tlp = 0;
16014         uint32_t cwnd_to_use, pace_max_seg;
16015         int32_t do_a_prefetch = 0;
16016         int32_t prefetch_rsm = 0;
16017         int32_t orig_len = 0;
16018         struct timeval tv;
16019         int32_t prefetch_so_done = 0;
16020         struct tcp_log_buffer *lgb;
16021         struct inpcb *inp;
16022         struct sockbuf *sb;
16023         uint64_t ts_val = 0;
16024 #ifdef TCP_ACCOUNTING
16025         uint64_t crtsc;
16026 #endif
16027 #ifdef INET6
16028         struct ip6_hdr *ip6 = NULL;
16029         int32_t isipv6;
16030 #endif
16031         uint8_t filled_all = 0;
16032         bool hw_tls = false;
16033
16034         /* setup and take the cache hits here */
16035         rack = (struct tcp_rack *)tp->t_fb_ptr;
16036 #ifdef TCP_ACCOUNTING
16037         sched_pin();
16038         ts_val = get_cyclecount();
16039 #endif
16040         hpts_calling = rack->rc_inp->inp_hpts_calls;
16041         NET_EPOCH_ASSERT();
16042         INP_WLOCK_ASSERT(rack->rc_inp);
16043 #ifdef TCP_OFFLOAD
16044         if (tp->t_flags & TF_TOE) {
16045 #ifdef TCP_ACCOUNTING
16046                 sched_unpin();
16047 #endif
16048                 return (tcp_offload_output(tp));
16049         }
16050 #endif
16051         /*
16052          * For TFO connections in SYN_RECEIVED, only allow the initial
16053          * SYN|ACK and those sent by the retransmit timer.
16054          */
16055         if (IS_FASTOPEN(tp->t_flags) &&
16056             (tp->t_state == TCPS_SYN_RECEIVED) &&
16057             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16058             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16059 #ifdef TCP_ACCOUNTING
16060                 sched_unpin();
16061 #endif
16062                 return (0);
16063         }
16064 #ifdef INET6
16065         if (rack->r_state) {
16066                 /* Use the cache line loaded if possible */
16067                 isipv6 = rack->r_is_v6;
16068         } else {
16069                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16070         }
16071 #endif
16072         early = 0;
16073         cts = tcp_get_usecs(&tv);
16074         ms_cts = tcp_tv_to_mssectick(&tv);
16075         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16076             rack->rc_inp->inp_in_hpts) {
16077                 /*
16078                  * We are on the hpts for some timer but not hptsi output.
16079                  * Remove from the hpts unconditionally.
16080                  */
16081                 rack_timer_cancel(tp, rack, cts, __LINE__);
16082         }
16083         /* Are we pacing and late? */
16084         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16085             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16086                 /* We are delayed */
16087                 delayed = cts - rack->r_ctl.rc_last_output_to;
16088         } else {
16089                 delayed = 0;
16090         }
16091         /* Do the timers, which may override the pacer */
16092         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16093                 if (rack_process_timers(tp, rack, cts, hpts_calling)) {
16094                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16095 #ifdef TCP_ACCOUNTING
16096                         sched_unpin();
16097 #endif
16098                         return (0);
16099                 }
16100         }
16101         if (rack->rc_in_persist) {
16102                 if (rack->rc_inp->inp_in_hpts == 0) {
16103                         /* Timer is not running */
16104                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16105                 }
16106 #ifdef TCP_ACCOUNTING
16107                 sched_unpin();
16108 #endif
16109                 return (0);
16110         }
16111         if ((rack->r_timer_override) ||
16112             (rack->rc_ack_can_sendout_data) ||
16113             (delayed) ||
16114             (tp->t_state < TCPS_ESTABLISHED)) {
16115                 rack->rc_ack_can_sendout_data = 0;
16116                 if (rack->rc_inp->inp_in_hpts)
16117                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
16118         } else if (rack->rc_inp->inp_in_hpts) {
16119                 /*
16120                  * On the hpts you can't pass even if ACKNOW is on, we will
16121                  * when the hpts fires.
16122                  */
16123 #ifdef TCP_ACCOUNTING
16124                 crtsc = get_cyclecount();
16125                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16126                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16127                 }
16128                 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16129                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16130                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16131                 }
16132                 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16133                 sched_unpin();
16134 #endif
16135                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16136                 return (0);
16137         }
16138         rack->rc_inp->inp_hpts_calls = 0;
16139         /* Finish out both pacing early and late accounting */
16140         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16141             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16142                 early = rack->r_ctl.rc_last_output_to - cts;
16143         } else
16144                 early = 0;
16145         if (delayed) {
16146                 rack->r_ctl.rc_agg_delayed += delayed;
16147                 rack->r_late = 1;
16148         } else if (early) {
16149                 rack->r_ctl.rc_agg_early += early;
16150                 rack->r_early = 1;
16151         }
16152         /* Now that early/late accounting is done turn off the flag */
16153         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16154         rack->r_wanted_output = 0;
16155         rack->r_timer_override = 0;
16156         if ((tp->t_state != rack->r_state) &&
16157             TCPS_HAVEESTABLISHED(tp->t_state)) {
16158                 rack_set_state(tp, rack);
16159         }
16160         if ((rack->r_fast_output) &&
16161             (tp->rcv_numsacks == 0)) {
16162                 int ret;
16163
16164                 error = 0;
16165                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16166                 if (ret >= 0)
16167                         return(ret);
16168                 else if (error) {
16169                         inp = rack->rc_inp;
16170                         so = inp->inp_socket;
16171                         sb = &so->so_snd;
16172                         goto nomore;
16173                 }
16174         }
16175         inp = rack->rc_inp;
16176         /*
16177          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16178          * only allow the initial SYN or SYN|ACK and those sent
16179          * by the retransmit timer.
16180          */
16181         if (IS_FASTOPEN(tp->t_flags) &&
16182             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16183              (tp->t_state == TCPS_SYN_SENT)) &&
16184             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16185             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16186                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16187                 so = inp->inp_socket;
16188                 sb = &so->so_snd;
16189                 goto just_return_nolock;
16190         }
16191         /*
16192          * Determine length of data that should be transmitted, and flags
16193          * that will be used. If there is some data or critical controls
16194          * (SYN, RST) to send, then transmit; otherwise, investigate
16195          * further.
16196          */
16197         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16198         if (tp->t_idle_reduce) {
16199                 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16200                         rack_cc_after_idle(rack, tp);
16201         }
16202         tp->t_flags &= ~TF_LASTIDLE;
16203         if (idle) {
16204                 if (tp->t_flags & TF_MORETOCOME) {
16205                         tp->t_flags |= TF_LASTIDLE;
16206                         idle = 0;
16207                 }
16208         }
16209         if ((tp->snd_una == tp->snd_max) &&
16210             rack->r_ctl.rc_went_idle_time &&
16211             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16212                 idle = cts - rack->r_ctl.rc_went_idle_time;
16213                 if (idle > rack_min_probertt_hold) {
16214                         /* Count as a probe rtt */
16215                         if (rack->in_probe_rtt == 0) {
16216                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16217                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16218                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16219                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16220                         } else {
16221                                 rack_exit_probertt(rack, cts);
16222                         }
16223                 }
16224                 idle = 0;
16225         }
16226         if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16227                 rack_init_fsb_block(tp, rack);
16228 again:
16229         /*
16230          * If we've recently taken a timeout, snd_max will be greater than
16231          * snd_nxt.  There may be SACK information that allows us to avoid
16232          * resending already delivered data.  Adjust snd_nxt accordingly.
16233          */
16234         sendalot = 0;
16235         cts = tcp_get_usecs(&tv);
16236         ms_cts = tcp_tv_to_mssectick(&tv);
16237         tso = 0;
16238         mtu = 0;
16239         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16240         minseg = segsiz;
16241         if (rack->r_ctl.rc_pace_max_segs == 0)
16242                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16243         else
16244                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16245         sb_offset = tp->snd_max - tp->snd_una;
16246         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16247         flags = tcp_outflags[tp->t_state];
16248         while (rack->rc_free_cnt < rack_free_cache) {
16249                 rsm = rack_alloc(rack);
16250                 if (rsm == NULL) {
16251                         if (inp->inp_hpts_calls)
16252                                 /* Retry in a ms */
16253                                 slot = (1 * HPTS_USEC_IN_MSEC);
16254                         so = inp->inp_socket;
16255                         sb = &so->so_snd;
16256                         goto just_return_nolock;
16257                 }
16258                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16259                 rack->rc_free_cnt++;
16260                 rsm = NULL;
16261         }
16262         if (inp->inp_hpts_calls)
16263                 inp->inp_hpts_calls = 0;
16264         sack_rxmit = 0;
16265         len = 0;
16266         rsm = NULL;
16267         if (flags & TH_RST) {
16268                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16269                 so = inp->inp_socket;
16270                 sb = &so->so_snd;
16271                 goto send;
16272         }
16273         if (rack->r_ctl.rc_resend) {
16274                 /* Retransmit timer */
16275                 rsm = rack->r_ctl.rc_resend;
16276                 rack->r_ctl.rc_resend = NULL;
16277                 rsm->r_flags &= ~RACK_TLP;
16278                 len = rsm->r_end - rsm->r_start;
16279                 sack_rxmit = 1;
16280                 sendalot = 0;
16281                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16282                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16283                          __func__, __LINE__,
16284                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16285                 sb_offset = rsm->r_start - tp->snd_una;
16286                 if (len >= segsiz)
16287                         len = segsiz;
16288         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16289                 /* We have a retransmit that takes precedence */
16290                 rsm->r_flags &= ~RACK_TLP;
16291                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16292                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16293                         /* Enter recovery if not induced by a time-out */
16294                         rack->r_ctl.rc_rsm_start = rsm->r_start;
16295                         rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
16296                         rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
16297                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
16298                 }
16299 #ifdef INVARIANTS
16300                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16301                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16302                               tp, rack, rsm, rsm->r_start, tp->snd_una);
16303                 }
16304 #endif
16305                 len = rsm->r_end - rsm->r_start;
16306                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16307                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16308                          __func__, __LINE__,
16309                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16310                 sb_offset = rsm->r_start - tp->snd_una;
16311                 sendalot = 0;
16312                 if (len >= segsiz)
16313                         len = segsiz;
16314                 if (len > 0) {
16315                         sack_rxmit = 1;
16316                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16317                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16318                             min(len, segsiz));
16319                         counter_u64_add(rack_rtm_prr_retran, 1);
16320                 }
16321         } else if (rack->r_ctl.rc_tlpsend) {
16322                 /* Tail loss probe */
16323                 long cwin;
16324                 long tlen;
16325
16326                 doing_tlp = 1;
16327                 /*
16328                  * Check if we can do a TLP with a RACK'd packet
16329                  * this can happen if we are not doing the rack
16330                  * cheat and we skipped to a TLP and it
16331                  * went off.
16332                  */
16333                 rsm = rack->r_ctl.rc_tlpsend;
16334                 rsm->r_flags |= RACK_TLP;
16335
16336                 rack->r_ctl.rc_tlpsend = NULL;
16337                 sack_rxmit = 1;
16338                 tlen = rsm->r_end - rsm->r_start;
16339                 if (tlen > segsiz)
16340                         tlen = segsiz;
16341                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16342                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16343                          __func__, __LINE__,
16344                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16345                 sb_offset = rsm->r_start - tp->snd_una;
16346                 cwin = min(tp->snd_wnd, tlen);
16347                 len = cwin;
16348         }
16349         if (rack->r_must_retran &&
16350             (rsm == NULL)) {
16351                 /*
16352                  * Non-Sack and we had a RTO or MTU change, we
16353                  * need to retransmit until we reach
16354                  * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
16355                  */
16356                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
16357                         int sendwin, flight;
16358
16359                         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16360                         flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16361                         if (flight >= sendwin) {
16362                                 so = inp->inp_socket;
16363                                 sb = &so->so_snd;
16364                                 goto just_return_nolock;
16365                         }
16366                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16367                         KASSERT(rsm != NULL, ("rsm is NULL rack:%p r_must_retran set", rack));
16368                         if (rsm == NULL) {
16369                                 /* TSNH */
16370                                 rack->r_must_retran = 0;
16371                                 rack->r_ctl.rc_out_at_rto = 0;
16372                                 rack->r_must_retran = 0;
16373                                 so = inp->inp_socket;
16374                                 sb = &so->so_snd;
16375                                 goto just_return_nolock;
16376                         }
16377                         sack_rxmit = 1;
16378                         len = rsm->r_end - rsm->r_start;
16379                         sendalot = 0;
16380                         sb_offset = rsm->r_start - tp->snd_una;
16381                         if (len >= segsiz)
16382                                 len = segsiz;
16383                 } else {
16384                         /* We must be done if there is nothing outstanding */
16385                         rack->r_must_retran = 0;
16386                         rack->r_ctl.rc_out_at_rto = 0;
16387                 }
16388         }
16389         /*
16390          * Enforce a connection sendmap count limit if set
16391          * as long as we are not retransmiting.
16392          */
16393         if ((rsm == NULL) &&
16394             (rack->do_detection == 0) &&
16395             (V_tcp_map_entries_limit > 0) &&
16396             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16397                 counter_u64_add(rack_to_alloc_limited, 1);
16398                 if (!rack->alloc_limit_reported) {
16399                         rack->alloc_limit_reported = 1;
16400                         counter_u64_add(rack_alloc_limited_conns, 1);
16401                 }
16402                 so = inp->inp_socket;
16403                 sb = &so->so_snd;
16404                 goto just_return_nolock;
16405         }
16406         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16407                 /* we are retransmitting the fin */
16408                 len--;
16409                 if (len) {
16410                         /*
16411                          * When retransmitting data do *not* include the
16412                          * FIN. This could happen from a TLP probe.
16413                          */
16414                         flags &= ~TH_FIN;
16415                 }
16416         }
16417 #ifdef INVARIANTS
16418         /* For debugging */
16419         rack->r_ctl.rc_rsm_at_retran = rsm;
16420 #endif
16421         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16422             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16423                 int ret;
16424
16425                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len);
16426                 if (ret == 0)
16427                         return (0);
16428         }
16429         so = inp->inp_socket;
16430         sb = &so->so_snd;
16431         if (do_a_prefetch == 0) {
16432                 kern_prefetch(sb, &do_a_prefetch);
16433                 do_a_prefetch = 1;
16434         }
16435 #ifdef NETFLIX_SHARED_CWND
16436         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
16437             rack->rack_enable_scwnd) {
16438                 /* We are doing cwnd sharing */
16439                 if (rack->gp_ready &&
16440                     (rack->rack_attempted_scwnd == 0) &&
16441                     (rack->r_ctl.rc_scw == NULL) &&
16442                     tp->t_lib) {
16443                         /* The pcbid is in, lets make an attempt */
16444                         counter_u64_add(rack_try_scwnd, 1);
16445                         rack->rack_attempted_scwnd = 1;
16446                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
16447                                                                    &rack->r_ctl.rc_scw_index,
16448                                                                    segsiz);
16449                 }
16450                 if (rack->r_ctl.rc_scw &&
16451                     (rack->rack_scwnd_is_idle == 1) &&
16452                     sbavail(&so->so_snd)) {
16453                         /* we are no longer out of data */
16454                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16455                         rack->rack_scwnd_is_idle = 0;
16456                 }
16457                 if (rack->r_ctl.rc_scw) {
16458                         /* First lets update and get the cwnd */
16459                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
16460                                                                     rack->r_ctl.rc_scw_index,
16461                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
16462                 }
16463         }
16464 #endif
16465         /*
16466          * Get standard flags, and add SYN or FIN if requested by 'hidden'
16467          * state flags.
16468          */
16469         if (tp->t_flags & TF_NEEDFIN)
16470                 flags |= TH_FIN;
16471         if (tp->t_flags & TF_NEEDSYN)
16472                 flags |= TH_SYN;
16473         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
16474                 void *end_rsm;
16475                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
16476                 if (end_rsm)
16477                         kern_prefetch(end_rsm, &prefetch_rsm);
16478                 prefetch_rsm = 1;
16479         }
16480         SOCKBUF_LOCK(sb);
16481         /*
16482          * If snd_nxt == snd_max and we have transmitted a FIN, the
16483          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
16484          * negative length.  This can also occur when TCP opens up its
16485          * congestion window while receiving additional duplicate acks after
16486          * fast-retransmit because TCP will reset snd_nxt to snd_max after
16487          * the fast-retransmit.
16488          *
16489          * In the normal retransmit-FIN-only case, however, snd_nxt will be
16490          * set to snd_una, the sb_offset will be 0, and the length may wind
16491          * up 0.
16492          *
16493          * If sack_rxmit is true we are retransmitting from the scoreboard
16494          * in which case len is already set.
16495          */
16496         if ((sack_rxmit == 0) &&
16497             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
16498                 uint32_t avail;
16499
16500                 avail = sbavail(sb);
16501                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
16502                         sb_offset = tp->snd_nxt - tp->snd_una;
16503                 else
16504                         sb_offset = 0;
16505                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
16506                         if (rack->r_ctl.rc_tlp_new_data) {
16507                                 /* TLP is forcing out new data */
16508                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
16509                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
16510                                 }
16511                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
16512                                         if (tp->snd_wnd > sb_offset)
16513                                                 len = tp->snd_wnd - sb_offset;
16514                                         else
16515                                                 len = 0;
16516                                 } else {
16517                                         len = rack->r_ctl.rc_tlp_new_data;
16518                                 }
16519                                 rack->r_ctl.rc_tlp_new_data = 0;
16520                                 doing_tlp = 1;
16521                         }  else {
16522                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
16523                         }
16524                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
16525                                 /*
16526                                  * For prr=off, we need to send only 1 MSS
16527                                  * at a time. We do this because another sack could
16528                                  * be arriving that causes us to send retransmits and
16529                                  * we don't want to be on a long pace due to a larger send
16530                                  * that keeps us from sending out the retransmit.
16531                                  */
16532                                 len = segsiz;
16533                         }
16534                 } else {
16535                         uint32_t outstanding;
16536                         /*
16537                          * We are inside of a Fast recovery episode, this
16538                          * is caused by a SACK or 3 dup acks. At this point
16539                          * we have sent all the retransmissions and we rely
16540                          * on PRR to dictate what we will send in the form of
16541                          * new data.
16542                          */
16543
16544                         outstanding = tp->snd_max - tp->snd_una;
16545                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
16546                                 if (tp->snd_wnd > outstanding) {
16547                                         len = tp->snd_wnd - outstanding;
16548                                         /* Check to see if we have the data */
16549                                         if ((sb_offset + len) > avail) {
16550                                                 /* It does not all fit */
16551                                                 if (avail > sb_offset)
16552                                                         len = avail - sb_offset;
16553                                                 else
16554                                                         len = 0;
16555                                         }
16556                                 } else {
16557                                         len = 0;
16558                                 }
16559                         } else if (avail > sb_offset) {
16560                                 len = avail - sb_offset;
16561                         } else {
16562                                 len = 0;
16563                         }
16564                         if (len > 0) {
16565                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
16566                                         len = rack->r_ctl.rc_prr_sndcnt;
16567                                 }
16568                                 if (len > 0) {
16569                                         sub_from_prr = 1;
16570                                         counter_u64_add(rack_rtm_prr_newdata, 1);
16571                                 }
16572                         }
16573                         if (len > segsiz) {
16574                                 /*
16575                                  * We should never send more than a MSS when
16576                                  * retransmitting or sending new data in prr
16577                                  * mode unless the override flag is on. Most
16578                                  * likely the PRR algorithm is not going to
16579                                  * let us send a lot as well :-)
16580                                  */
16581                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
16582                                         len = segsiz;
16583                                 }
16584                         } else if (len < segsiz) {
16585                                 /*
16586                                  * Do we send any? The idea here is if the
16587                                  * send empty's the socket buffer we want to
16588                                  * do it. However if not then lets just wait
16589                                  * for our prr_sndcnt to get bigger.
16590                                  */
16591                                 long leftinsb;
16592
16593                                 leftinsb = sbavail(sb) - sb_offset;
16594                                 if (leftinsb > len) {
16595                                         /* This send does not empty the sb */
16596                                         len = 0;
16597                                 }
16598                         }
16599                 }
16600         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
16601                 /*
16602                  * If you have not established
16603                  * and are not doing FAST OPEN
16604                  * no data please.
16605                  */
16606                 if ((sack_rxmit == 0) &&
16607                     (!IS_FASTOPEN(tp->t_flags))){
16608                         len = 0;
16609                         sb_offset = 0;
16610                 }
16611         }
16612         if (prefetch_so_done == 0) {
16613                 kern_prefetch(so, &prefetch_so_done);
16614                 prefetch_so_done = 1;
16615         }
16616         /*
16617          * Lop off SYN bit if it has already been sent.  However, if this is
16618          * SYN-SENT state and if segment contains data and if we don't know
16619          * that foreign host supports TAO, suppress sending segment.
16620          */
16621         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
16622             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
16623                 /*
16624                  * When sending additional segments following a TFO SYN|ACK,
16625                  * do not include the SYN bit.
16626                  */
16627                 if (IS_FASTOPEN(tp->t_flags) &&
16628                     (tp->t_state == TCPS_SYN_RECEIVED))
16629                         flags &= ~TH_SYN;
16630         }
16631         /*
16632          * Be careful not to send data and/or FIN on SYN segments. This
16633          * measure is needed to prevent interoperability problems with not
16634          * fully conformant TCP implementations.
16635          */
16636         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
16637                 len = 0;
16638                 flags &= ~TH_FIN;
16639         }
16640         /*
16641          * On TFO sockets, ensure no data is sent in the following cases:
16642          *
16643          *  - When retransmitting SYN|ACK on a passively-created socket
16644          *
16645          *  - When retransmitting SYN on an actively created socket
16646          *
16647          *  - When sending a zero-length cookie (cookie request) on an
16648          *    actively created socket
16649          *
16650          *  - When the socket is in the CLOSED state (RST is being sent)
16651          */
16652         if (IS_FASTOPEN(tp->t_flags) &&
16653             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
16654              ((tp->t_state == TCPS_SYN_SENT) &&
16655               (tp->t_tfo_client_cookie_len == 0)) ||
16656              (flags & TH_RST))) {
16657                 sack_rxmit = 0;
16658                 len = 0;
16659         }
16660         /* Without fast-open there should never be data sent on a SYN */
16661         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
16662                 tp->snd_nxt = tp->iss;
16663                 len = 0;
16664         }
16665         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
16666                 /* We only send 1 MSS if we have a DSACK block */
16667                 add_flag |= RACK_SENT_W_DSACK;
16668                 len = segsiz;
16669         }
16670         orig_len = len;
16671         if (len <= 0) {
16672                 /*
16673                  * If FIN has been sent but not acked, but we haven't been
16674                  * called to retransmit, len will be < 0.  Otherwise, window
16675                  * shrank after we sent into it.  If window shrank to 0,
16676                  * cancel pending retransmit, pull snd_nxt back to (closed)
16677                  * window, and set the persist timer if it isn't already
16678                  * going.  If the window didn't close completely, just wait
16679                  * for an ACK.
16680                  *
16681                  * We also do a general check here to ensure that we will
16682                  * set the persist timer when we have data to send, but a
16683                  * 0-byte window. This makes sure the persist timer is set
16684                  * even if the packet hits one of the "goto send" lines
16685                  * below.
16686                  */
16687                 len = 0;
16688                 if ((tp->snd_wnd == 0) &&
16689                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16690                     (tp->snd_una == tp->snd_max) &&
16691                     (sb_offset < (int)sbavail(sb))) {
16692                         rack_enter_persist(tp, rack, cts);
16693                 }
16694         } else if ((rsm == NULL) &&
16695                    (doing_tlp == 0) &&
16696                    (len < pace_max_seg)) {
16697                 /*
16698                  * We are not sending a maximum sized segment for
16699                  * some reason. Should we not send anything (think
16700                  * sws or persists)?
16701                  */
16702                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16703                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16704                     (len < minseg) &&
16705                     (len < (int)(sbavail(sb) - sb_offset))) {
16706                         /*
16707                          * Here the rwnd is less than
16708                          * the minimum pacing size, this is not a retransmit,
16709                          * we are established and
16710                          * the send is not the last in the socket buffer
16711                          * we send nothing, and we may enter persists
16712                          * if nothing is outstanding.
16713                          */
16714                         len = 0;
16715                         if (tp->snd_max == tp->snd_una) {
16716                                 /*
16717                                  * Nothing out we can
16718                                  * go into persists.
16719                                  */
16720                                 rack_enter_persist(tp, rack, cts);
16721                         }
16722                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
16723                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16724                            (len < (int)(sbavail(sb) - sb_offset)) &&
16725                            (len < minseg)) {
16726                         /*
16727                          * Here we are not retransmitting, and
16728                          * the cwnd is not so small that we could
16729                          * not send at least a min size (rxt timer
16730                          * not having gone off), We have 2 segments or
16731                          * more already in flight, its not the tail end
16732                          * of the socket buffer  and the cwnd is blocking
16733                          * us from sending out a minimum pacing segment size.
16734                          * Lets not send anything.
16735                          */
16736                         len = 0;
16737                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
16738                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16739                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16740                            (len < (int)(sbavail(sb) - sb_offset)) &&
16741                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
16742                         /*
16743                          * Here we have a send window but we have
16744                          * filled it up and we can't send another pacing segment.
16745                          * We also have in flight more than 2 segments
16746                          * and we are not completing the sb i.e. we allow
16747                          * the last bytes of the sb to go out even if
16748                          * its not a full pacing segment.
16749                          */
16750                         len = 0;
16751                 } else if ((rack->r_ctl.crte != NULL) &&
16752                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
16753                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
16754                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
16755                            (len < (int)(sbavail(sb) - sb_offset))) {
16756                         /*
16757                          * Here we are doing hardware pacing, this is not a TLP,
16758                          * we are not sending a pace max segment size, there is rwnd
16759                          * room to send at least N pace_max_seg, the cwnd is greater
16760                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
16761                          * more segments in flight and its not the tail of the socket buffer.
16762                          *
16763                          * We don't want to send instead we need to get more ack's in to
16764                          * allow us to send a full pacing segment. Normally, if we are pacing
16765                          * about the right speed, we should have finished our pacing
16766                          * send as most of the acks have come back if we are at the
16767                          * right rate. This is a bit fuzzy since return path delay
16768                          * can delay the acks, which is why we want to make sure we
16769                          * have cwnd space to have a bit more than a max pace segments in flight.
16770                          *
16771                          * If we have not gotten our acks back we are pacing at too high a
16772                          * rate delaying will not hurt and will bring our GP estimate down by
16773                          * injecting the delay. If we don't do this we will send
16774                          * 2 MSS out in response to the acks being clocked in which
16775                          * defeats the point of hw-pacing (i.e. to help us get
16776                          * larger TSO's out).
16777                          */
16778                         len = 0;
16779
16780                 }
16781
16782         }
16783         /* len will be >= 0 after this point. */
16784         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
16785         rack_sndbuf_autoscale(rack);
16786         /*
16787          * Decide if we can use TCP Segmentation Offloading (if supported by
16788          * hardware).
16789          *
16790          * TSO may only be used if we are in a pure bulk sending state.  The
16791          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
16792          * options prevent using TSO.  With TSO the TCP header is the same
16793          * (except for the sequence number) for all generated packets.  This
16794          * makes it impossible to transmit any options which vary per
16795          * generated segment or packet.
16796          *
16797          * IPv4 handling has a clear separation of ip options and ip header
16798          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
16799          * the right thing below to provide length of just ip options and thus
16800          * checking for ipoptlen is enough to decide if ip options are present.
16801          */
16802         ipoptlen = 0;
16803 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16804         /*
16805          * Pre-calculate here as we save another lookup into the darknesses
16806          * of IPsec that way and can actually decide if TSO is ok.
16807          */
16808 #ifdef INET6
16809         if (isipv6 && IPSEC_ENABLED(ipv6))
16810                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
16811 #ifdef INET
16812         else
16813 #endif
16814 #endif                          /* INET6 */
16815 #ifdef INET
16816                 if (IPSEC_ENABLED(ipv4))
16817                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
16818 #endif                          /* INET */
16819 #endif
16820
16821 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16822         ipoptlen += ipsec_optlen;
16823 #endif
16824         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
16825             (tp->t_port == 0) &&
16826             ((tp->t_flags & TF_SIGNATURE) == 0) &&
16827             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
16828             ipoptlen == 0)
16829                 tso = 1;
16830         {
16831                 uint32_t outstanding;
16832
16833                 outstanding = tp->snd_max - tp->snd_una;
16834                 if (tp->t_flags & TF_SENTFIN) {
16835                         /*
16836                          * If we sent a fin, snd_max is 1 higher than
16837                          * snd_una
16838                          */
16839                         outstanding--;
16840                 }
16841                 if (sack_rxmit) {
16842                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
16843                                 flags &= ~TH_FIN;
16844                 } else {
16845                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
16846                                    sbused(sb)))
16847                                 flags &= ~TH_FIN;
16848                 }
16849         }
16850         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
16851             (long)TCP_MAXWIN << tp->rcv_scale);
16852
16853         /*
16854          * Sender silly window avoidance.   We transmit under the following
16855          * conditions when len is non-zero:
16856          *
16857          * - We have a full segment (or more with TSO) - This is the last
16858          * buffer in a write()/send() and we are either idle or running
16859          * NODELAY - we've timed out (e.g. persist timer) - we have more
16860          * then 1/2 the maximum send window's worth of data (receiver may be
16861          * limited the window size) - we need to retransmit
16862          */
16863         if (len) {
16864                 if (len >= segsiz) {
16865                         goto send;
16866                 }
16867                 /*
16868                  * NOTE! on localhost connections an 'ack' from the remote
16869                  * end may occur synchronously with the output and cause us
16870                  * to flush a buffer queued with moretocome.  XXX
16871                  *
16872                  */
16873                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
16874                     (idle || (tp->t_flags & TF_NODELAY)) &&
16875                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
16876                     (tp->t_flags & TF_NOPUSH) == 0) {
16877                         pass = 2;
16878                         goto send;
16879                 }
16880                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
16881                         pass = 22;
16882                         goto send;
16883                 }
16884                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
16885                         pass = 4;
16886                         goto send;
16887                 }
16888                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
16889                         pass = 5;
16890                         goto send;
16891                 }
16892                 if (sack_rxmit) {
16893                         pass = 6;
16894                         goto send;
16895                 }
16896                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
16897                     (ctf_outstanding(tp) < (segsiz * 2))) {
16898                         /*
16899                          * We have less than two MSS outstanding (delayed ack)
16900                          * and our rwnd will not let us send a full sized
16901                          * MSS. Lets go ahead and let this small segment
16902                          * out because we want to try to have at least two
16903                          * packets inflight to not be caught by delayed ack.
16904                          */
16905                         pass = 12;
16906                         goto send;
16907                 }
16908         }
16909         /*
16910          * Sending of standalone window updates.
16911          *
16912          * Window updates are important when we close our window due to a
16913          * full socket buffer and are opening it again after the application
16914          * reads data from it.  Once the window has opened again and the
16915          * remote end starts to send again the ACK clock takes over and
16916          * provides the most current window information.
16917          *
16918          * We must avoid the silly window syndrome whereas every read from
16919          * the receive buffer, no matter how small, causes a window update
16920          * to be sent.  We also should avoid sending a flurry of window
16921          * updates when the socket buffer had queued a lot of data and the
16922          * application is doing small reads.
16923          *
16924          * Prevent a flurry of pointless window updates by only sending an
16925          * update when we can increase the advertized window by more than
16926          * 1/4th of the socket buffer capacity.  When the buffer is getting
16927          * full or is very small be more aggressive and send an update
16928          * whenever we can increase by two mss sized segments. In all other
16929          * situations the ACK's to new incoming data will carry further
16930          * window increases.
16931          *
16932          * Don't send an independent window update if a delayed ACK is
16933          * pending (it will get piggy-backed on it) or the remote side
16934          * already has done a half-close and won't send more data.  Skip
16935          * this if the connection is in T/TCP half-open state.
16936          */
16937         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
16938             !(tp->t_flags & TF_DELACK) &&
16939             !TCPS_HAVERCVDFIN(tp->t_state)) {
16940                 /*
16941                  * "adv" is the amount we could increase the window, taking
16942                  * into account that we are limited by TCP_MAXWIN <<
16943                  * tp->rcv_scale.
16944                  */
16945                 int32_t adv;
16946                 int oldwin;
16947
16948                 adv = recwin;
16949                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
16950                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
16951                         if (adv > oldwin)
16952                             adv -= oldwin;
16953                         else {
16954                                 /* We can't increase the window */
16955                                 adv = 0;
16956                         }
16957                 } else
16958                         oldwin = 0;
16959
16960                 /*
16961                  * If the new window size ends up being the same as or less
16962                  * than the old size when it is scaled, then don't force
16963                  * a window update.
16964                  */
16965                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
16966                         goto dontupdate;
16967
16968                 if (adv >= (int32_t)(2 * segsiz) &&
16969                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
16970                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
16971                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
16972                         pass = 7;
16973                         goto send;
16974                 }
16975                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
16976                         pass = 23;
16977                         goto send;
16978                 }
16979         }
16980 dontupdate:
16981
16982         /*
16983          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
16984          * is also a catch-all for the retransmit timer timeout case.
16985          */
16986         if (tp->t_flags & TF_ACKNOW) {
16987                 pass = 8;
16988                 goto send;
16989         }
16990         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
16991                 pass = 9;
16992                 goto send;
16993         }
16994         /*
16995          * If our state indicates that FIN should be sent and we have not
16996          * yet done so, then we need to send.
16997          */
16998         if ((flags & TH_FIN) &&
16999             (tp->snd_nxt == tp->snd_una)) {
17000                 pass = 11;
17001                 goto send;
17002         }
17003         /*
17004          * No reason to send a segment, just return.
17005          */
17006 just_return:
17007         SOCKBUF_UNLOCK(sb);
17008 just_return_nolock:
17009         {
17010                 int app_limited = CTF_JR_SENT_DATA;
17011
17012                 if (tot_len_this_send > 0) {
17013                         /* Make sure snd_nxt is up to max */
17014                         rack->r_ctl.fsb.recwin = recwin;
17015                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17016                         if ((error == 0) &&
17017                             rack_use_rfo &&
17018                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
17019                             (ipoptlen == 0) &&
17020                             (tp->snd_nxt == tp->snd_max) &&
17021                             (tp->rcv_numsacks == 0) &&
17022                             rack->r_fsb_inited &&
17023                             TCPS_HAVEESTABLISHED(tp->t_state) &&
17024                             (rack->r_must_retran == 0) &&
17025                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
17026                             (len > 0) && (orig_len > 0) &&
17027                             (orig_len > len) &&
17028                             ((orig_len - len) >= segsiz) &&
17029                             ((optlen == 0) ||
17030                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17031                                 /* We can send at least one more MSS using our fsb */
17032
17033                                 rack->r_fast_output = 1;
17034                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17035                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17036                                 rack->r_ctl.fsb.tcp_flags = flags;
17037                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
17038                                 if (hw_tls)
17039                                         rack->r_ctl.fsb.hw_tls = 1;
17040                                 else
17041                                         rack->r_ctl.fsb.hw_tls = 0;
17042                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17043                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
17044                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17045                                          (tp->snd_max - tp->snd_una)));
17046                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
17047                                         rack->r_fast_output = 0;
17048                                 else {
17049                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17050                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
17051                                         else
17052                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
17053                                 }
17054                         } else
17055                                 rack->r_fast_output = 0;
17056
17057
17058                         rack_log_fsb(rack, tp, so, flags,
17059                                      ipoptlen, orig_len, len, 0,
17060                                      1, optlen, __LINE__, 1);
17061                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17062                                 tp->snd_nxt = tp->snd_max;
17063                 } else {
17064                         int end_window = 0;
17065                         uint32_t seq = tp->gput_ack;
17066
17067                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17068                         if (rsm) {
17069                                 /*
17070                                  * Mark the last sent that we just-returned (hinting
17071                                  * that delayed ack may play a role in any rtt measurement).
17072                                  */
17073                                 rsm->r_just_ret = 1;
17074                         }
17075                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17076                         rack->r_ctl.rc_agg_delayed = 0;
17077                         rack->r_early = 0;
17078                         rack->r_late = 0;
17079                         rack->r_ctl.rc_agg_early = 0;
17080                         if ((ctf_outstanding(tp) +
17081                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17082                                  minseg)) >= tp->snd_wnd) {
17083                                 /* We are limited by the rwnd */
17084                                 app_limited = CTF_JR_RWND_LIMITED;
17085                                 if (IN_FASTRECOVERY(tp->t_flags))
17086                                     rack->r_ctl.rc_prr_sndcnt = 0;
17087                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17088                                 /* We are limited by whats available -- app limited */
17089                                 app_limited = CTF_JR_APP_LIMITED;
17090                                 if (IN_FASTRECOVERY(tp->t_flags))
17091                                     rack->r_ctl.rc_prr_sndcnt = 0;
17092                         } else if ((idle == 0) &&
17093                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17094                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17095                                    (len < segsiz)) {
17096                                 /*
17097                                  * No delay is not on and the
17098                                  * user is sending less than 1MSS. This
17099                                  * brings out SWS avoidance so we
17100                                  * don't send. Another app-limited case.
17101                                  */
17102                                 app_limited = CTF_JR_APP_LIMITED;
17103                         } else if (tp->t_flags & TF_NOPUSH) {
17104                                 /*
17105                                  * The user has requested no push of
17106                                  * the last segment and we are
17107                                  * at the last segment. Another app
17108                                  * limited case.
17109                                  */
17110                                 app_limited = CTF_JR_APP_LIMITED;
17111                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17112                                 /* Its the cwnd */
17113                                 app_limited = CTF_JR_CWND_LIMITED;
17114                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17115                                    (rack->rack_no_prr == 0) &&
17116                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17117                                 app_limited = CTF_JR_PRR;
17118                         } else {
17119                                 /* Now why here are we not sending? */
17120 #ifdef NOW
17121 #ifdef INVARIANTS
17122                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17123 #endif
17124 #endif
17125                                 app_limited = CTF_JR_ASSESSING;
17126                         }
17127                         /*
17128                          * App limited in some fashion, for our pacing GP
17129                          * measurements we don't want any gap (even cwnd).
17130                          * Close  down the measurement window.
17131                          */
17132                         if (rack_cwnd_block_ends_measure &&
17133                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17134                              (app_limited == CTF_JR_PRR))) {
17135                                 /*
17136                                  * The reason we are not sending is
17137                                  * the cwnd (or prr). We have been configured
17138                                  * to end the measurement window in
17139                                  * this case.
17140                                  */
17141                                 end_window = 1;
17142                         } else if (rack_rwnd_block_ends_measure &&
17143                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17144                                 /*
17145                                  * We are rwnd limited and have been
17146                                  * configured to end the measurement
17147                                  * window in this case.
17148                                  */
17149                                 end_window = 1;
17150                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17151                                 /*
17152                                  * A true application limited period, we have
17153                                  * ran out of data.
17154                                  */
17155                                 end_window = 1;
17156                         } else if (app_limited == CTF_JR_ASSESSING) {
17157                                 /*
17158                                  * In the assessing case we hit the end of
17159                                  * the if/else and had no known reason
17160                                  * This will panic us under invariants..
17161                                  *
17162                                  * If we get this out in logs we need to
17163                                  * investagate which reason we missed.
17164                                  */
17165                                 end_window = 1;
17166                         }
17167                         if (end_window) {
17168                                 uint8_t log = 0;
17169
17170                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17171                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17172                                         /* Mark the last packet has app limited */
17173                                         tp->gput_ack = tp->snd_max;
17174                                         log = 1;
17175                                 }
17176                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17177                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17178                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17179                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17180                                         else {
17181                                                 /*
17182                                                  * Go out to the end app limited and mark
17183                                                  * this new one as next and move the end_appl up
17184                                                  * to this guy.
17185                                                  */
17186                                                 if (rack->r_ctl.rc_end_appl)
17187                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17188                                                 rack->r_ctl.rc_end_appl = rsm;
17189                                         }
17190                                         rsm->r_flags |= RACK_APP_LIMITED;
17191                                         rack->r_ctl.rc_app_limited_cnt++;
17192                                 }
17193                                 if (log)
17194                                         rack_log_pacing_delay_calc(rack,
17195                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17196                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL);
17197                         }
17198                 }
17199                 if (slot) {
17200                         /* set the rack tcb into the slot N */
17201                         counter_u64_add(rack_paced_segments, 1);
17202                 } else if (tot_len_this_send) {
17203                         counter_u64_add(rack_unpaced_segments, 1);
17204                 }
17205                 /* Check if we need to go into persists or not */
17206                 if ((tp->snd_max == tp->snd_una) &&
17207                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17208                     sbavail(sb) &&
17209                     (sbavail(sb) > tp->snd_wnd) &&
17210                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17211                         /* Yes lets make sure to move to persist before timer-start */
17212                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17213                 }
17214                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17215                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17216         }
17217 #ifdef NETFLIX_SHARED_CWND
17218         if ((sbavail(sb) == 0) &&
17219             rack->r_ctl.rc_scw) {
17220                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17221                 rack->rack_scwnd_is_idle = 1;
17222         }
17223 #endif
17224 #ifdef TCP_ACCOUNTING
17225         if (tot_len_this_send > 0) {
17226                 crtsc = get_cyclecount();
17227                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17228                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
17229                 }
17230                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17231                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17232                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17233                 }
17234                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17235                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17236                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17237                 }
17238                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17239         } else {
17240                 crtsc = get_cyclecount();
17241                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17242                         tp->tcp_cnt_counters[SND_LIMITED]++;
17243                 }
17244                 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17245                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17246                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17247                 }
17248                 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17249         }
17250         sched_unpin();
17251 #endif
17252         return (0);
17253
17254 send:
17255         if (rsm || sack_rxmit)
17256                 counter_u64_add(rack_nfto_resend, 1);
17257         else
17258                 counter_u64_add(rack_non_fto_send, 1);
17259         if ((flags & TH_FIN) &&
17260             sbavail(sb)) {
17261                 /*
17262                  * We do not transmit a FIN
17263                  * with data outstanding. We
17264                  * need to make it so all data
17265                  * is acked first.
17266                  */
17267                 flags &= ~TH_FIN;
17268         }
17269         /* Enforce stack imposed max seg size if we have one */
17270         if (rack->r_ctl.rc_pace_max_segs &&
17271             (len > rack->r_ctl.rc_pace_max_segs)) {
17272                 mark = 1;
17273                 len = rack->r_ctl.rc_pace_max_segs;
17274         }
17275         SOCKBUF_LOCK_ASSERT(sb);
17276         if (len > 0) {
17277                 if (len >= segsiz)
17278                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17279                 else
17280                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17281         }
17282         /*
17283          * Before ESTABLISHED, force sending of initial options unless TCP
17284          * set not to do any options. NOTE: we assume that the IP/TCP header
17285          * plus TCP options always fit in a single mbuf, leaving room for a
17286          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17287          * + optlen <= MCLBYTES
17288          */
17289         optlen = 0;
17290 #ifdef INET6
17291         if (isipv6)
17292                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17293         else
17294 #endif
17295                 hdrlen = sizeof(struct tcpiphdr);
17296
17297         /*
17298          * Compute options for segment. We only have to care about SYN and
17299          * established connection segments.  Options for SYN-ACK segments
17300          * are handled in TCP syncache.
17301          */
17302         to.to_flags = 0;
17303         if ((tp->t_flags & TF_NOOPT) == 0) {
17304                 /* Maximum segment size. */
17305                 if (flags & TH_SYN) {
17306                         tp->snd_nxt = tp->iss;
17307                         to.to_mss = tcp_mssopt(&inp->inp_inc);
17308                         if (tp->t_port)
17309                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
17310                         to.to_flags |= TOF_MSS;
17311
17312                         /*
17313                          * On SYN or SYN|ACK transmits on TFO connections,
17314                          * only include the TFO option if it is not a
17315                          * retransmit, as the presence of the TFO option may
17316                          * have caused the original SYN or SYN|ACK to have
17317                          * been dropped by a middlebox.
17318                          */
17319                         if (IS_FASTOPEN(tp->t_flags) &&
17320                             (tp->t_rxtshift == 0)) {
17321                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
17322                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17323                                         to.to_tfo_cookie =
17324                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
17325                                         to.to_flags |= TOF_FASTOPEN;
17326                                         wanted_cookie = 1;
17327                                 } else if (tp->t_state == TCPS_SYN_SENT) {
17328                                         to.to_tfo_len =
17329                                                 tp->t_tfo_client_cookie_len;
17330                                         to.to_tfo_cookie =
17331                                                 tp->t_tfo_cookie.client;
17332                                         to.to_flags |= TOF_FASTOPEN;
17333                                         wanted_cookie = 1;
17334                                         /*
17335                                          * If we wind up having more data to
17336                                          * send with the SYN than can fit in
17337                                          * one segment, don't send any more
17338                                          * until the SYN|ACK comes back from
17339                                          * the other end.
17340                                          */
17341                                         sendalot = 0;
17342                                 }
17343                         }
17344                 }
17345                 /* Window scaling. */
17346                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17347                         to.to_wscale = tp->request_r_scale;
17348                         to.to_flags |= TOF_SCALE;
17349                 }
17350                 /* Timestamps. */
17351                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
17352                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17353                         to.to_tsval = ms_cts + tp->ts_offset;
17354                         to.to_tsecr = tp->ts_recent;
17355                         to.to_flags |= TOF_TS;
17356                 }
17357                 /* Set receive buffer autosizing timestamp. */
17358                 if (tp->rfbuf_ts == 0 &&
17359                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
17360                         tp->rfbuf_ts = tcp_ts_getticks();
17361                 /* Selective ACK's. */
17362                 if (tp->t_flags & TF_SACK_PERMIT) {
17363                         if (flags & TH_SYN)
17364                                 to.to_flags |= TOF_SACKPERM;
17365                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17366                                  tp->rcv_numsacks > 0) {
17367                                 to.to_flags |= TOF_SACK;
17368                                 to.to_nsacks = tp->rcv_numsacks;
17369                                 to.to_sacks = (u_char *)tp->sackblks;
17370                         }
17371                 }
17372 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17373                 /* TCP-MD5 (RFC2385). */
17374                 if (tp->t_flags & TF_SIGNATURE)
17375                         to.to_flags |= TOF_SIGNATURE;
17376 #endif                          /* TCP_SIGNATURE */
17377
17378                 /* Processing the options. */
17379                 hdrlen += optlen = tcp_addoptions(&to, opt);
17380                 /*
17381                  * If we wanted a TFO option to be added, but it was unable
17382                  * to fit, ensure no data is sent.
17383                  */
17384                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17385                     !(to.to_flags & TOF_FASTOPEN))
17386                         len = 0;
17387         }
17388         if (tp->t_port) {
17389                 if (V_tcp_udp_tunneling_port == 0) {
17390                         /* The port was removed?? */
17391                         SOCKBUF_UNLOCK(&so->so_snd);
17392 #ifdef TCP_ACCOUNTING
17393                         crtsc = get_cyclecount();
17394                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17395                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17396                         }
17397                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
17398                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17399                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17400                         }
17401                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
17402                         sched_unpin();
17403 #endif
17404                         return (EHOSTUNREACH);
17405                 }
17406                 hdrlen += sizeof(struct udphdr);
17407         }
17408 #ifdef INET6
17409         if (isipv6)
17410                 ipoptlen = ip6_optlen(tp->t_inpcb);
17411         else
17412 #endif
17413                 if (tp->t_inpcb->inp_options)
17414                         ipoptlen = tp->t_inpcb->inp_options->m_len -
17415                                 offsetof(struct ipoption, ipopt_list);
17416                 else
17417                         ipoptlen = 0;
17418 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17419         ipoptlen += ipsec_optlen;
17420 #endif
17421
17422         /*
17423          * Adjust data length if insertion of options will bump the packet
17424          * length beyond the t_maxseg length. Clear the FIN bit because we
17425          * cut off the tail of the segment.
17426          */
17427         if (len + optlen + ipoptlen > tp->t_maxseg) {
17428                 if (tso) {
17429                         uint32_t if_hw_tsomax;
17430                         uint32_t moff;
17431                         int32_t max_len;
17432
17433                         /* extract TSO information */
17434                         if_hw_tsomax = tp->t_tsomax;
17435                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
17436                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
17437                         KASSERT(ipoptlen == 0,
17438                                 ("%s: TSO can't do IP options", __func__));
17439
17440                         /*
17441                          * Check if we should limit by maximum payload
17442                          * length:
17443                          */
17444                         if (if_hw_tsomax != 0) {
17445                                 /* compute maximum TSO length */
17446                                 max_len = (if_hw_tsomax - hdrlen -
17447                                            max_linkhdr);
17448                                 if (max_len <= 0) {
17449                                         len = 0;
17450                                 } else if (len > max_len) {
17451                                         sendalot = 1;
17452                                         len = max_len;
17453                                         mark = 2;
17454                                 }
17455                         }
17456                         /*
17457                          * Prevent the last segment from being fractional
17458                          * unless the send sockbuf can be emptied:
17459                          */
17460                         max_len = (tp->t_maxseg - optlen);
17461                         if ((sb_offset + len) < sbavail(sb)) {
17462                                 moff = len % (u_int)max_len;
17463                                 if (moff != 0) {
17464                                         mark = 3;
17465                                         len -= moff;
17466                                 }
17467                         }
17468                         /*
17469                          * In case there are too many small fragments don't
17470                          * use TSO:
17471                          */
17472                         if (len <= segsiz) {
17473                                 mark = 4;
17474                                 tso = 0;
17475                         }
17476                         /*
17477                          * Send the FIN in a separate segment after the bulk
17478                          * sending is done. We don't trust the TSO
17479                          * implementations to clear the FIN flag on all but
17480                          * the last segment.
17481                          */
17482                         if (tp->t_flags & TF_NEEDFIN) {
17483                                 sendalot = 4;
17484                         }
17485                 } else {
17486                         mark = 5;
17487                         if (optlen + ipoptlen >= tp->t_maxseg) {
17488                                 /*
17489                                  * Since we don't have enough space to put
17490                                  * the IP header chain and the TCP header in
17491                                  * one packet as required by RFC 7112, don't
17492                                  * send it. Also ensure that at least one
17493                                  * byte of the payload can be put into the
17494                                  * TCP segment.
17495                                  */
17496                                 SOCKBUF_UNLOCK(&so->so_snd);
17497                                 error = EMSGSIZE;
17498                                 sack_rxmit = 0;
17499                                 goto out;
17500                         }
17501                         len = tp->t_maxseg - optlen - ipoptlen;
17502                         sendalot = 5;
17503                 }
17504         } else {
17505                 tso = 0;
17506                 mark = 6;
17507         }
17508         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
17509                 ("%s: len > IP_MAXPACKET", __func__));
17510 #ifdef DIAGNOSTIC
17511 #ifdef INET6
17512         if (max_linkhdr + hdrlen > MCLBYTES)
17513 #else
17514                 if (max_linkhdr + hdrlen > MHLEN)
17515 #endif
17516                         panic("tcphdr too big");
17517 #endif
17518
17519         /*
17520          * This KASSERT is here to catch edge cases at a well defined place.
17521          * Before, those had triggered (random) panic conditions further
17522          * down.
17523          */
17524         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17525         if ((len == 0) &&
17526             (flags & TH_FIN) &&
17527             (sbused(sb))) {
17528                 /*
17529                  * We have outstanding data, don't send a fin by itself!.
17530                  */
17531                 goto just_return;
17532         }
17533         /*
17534          * Grab a header mbuf, attaching a copy of data to be transmitted,
17535          * and initialize the header from the template for sends on this
17536          * connection.
17537          */
17538         hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
17539         if (len) {
17540                 uint32_t max_val;
17541                 uint32_t moff;
17542
17543                 if (rack->r_ctl.rc_pace_max_segs)
17544                         max_val = rack->r_ctl.rc_pace_max_segs;
17545                 else if (rack->rc_user_set_max_segs)
17546                         max_val = rack->rc_user_set_max_segs * segsiz;
17547                 else
17548                         max_val = len;
17549                 /*
17550                  * We allow a limit on sending with hptsi.
17551                  */
17552                 if (len > max_val) {
17553                         mark = 7;
17554                         len = max_val;
17555                 }
17556 #ifdef INET6
17557                 if (MHLEN < hdrlen + max_linkhdr)
17558                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
17559                 else
17560 #endif
17561                         m = m_gethdr(M_NOWAIT, MT_DATA);
17562
17563                 if (m == NULL) {
17564                         SOCKBUF_UNLOCK(sb);
17565                         error = ENOBUFS;
17566                         sack_rxmit = 0;
17567                         goto out;
17568                 }
17569                 m->m_data += max_linkhdr;
17570                 m->m_len = hdrlen;
17571
17572                 /*
17573                  * Start the m_copy functions from the closest mbuf to the
17574                  * sb_offset in the socket buffer chain.
17575                  */
17576                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
17577                 s_mb = mb;
17578                 s_moff = moff;
17579                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
17580                         m_copydata(mb, moff, (int)len,
17581                                    mtod(m, caddr_t)+hdrlen);
17582                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17583                                 sbsndptr_adv(sb, mb, len);
17584                         m->m_len += len;
17585                 } else {
17586                         struct sockbuf *msb;
17587
17588                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17589                                 msb = NULL;
17590                         else
17591                                 msb = sb;
17592                         m->m_next = tcp_m_copym(
17593                                 mb, moff, &len,
17594                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
17595                                 ((rsm == NULL) ? hw_tls : 0)
17596 #ifdef NETFLIX_COPY_ARGS
17597                                 , &filled_all
17598 #endif
17599                                 );
17600                         if (len <= (tp->t_maxseg - optlen)) {
17601                                 /*
17602                                  * Must have ran out of mbufs for the copy
17603                                  * shorten it to no longer need tso. Lets
17604                                  * not put on sendalot since we are low on
17605                                  * mbufs.
17606                                  */
17607                                 tso = 0;
17608                         }
17609                         if (m->m_next == NULL) {
17610                                 SOCKBUF_UNLOCK(sb);
17611                                 (void)m_free(m);
17612                                 error = ENOBUFS;
17613                                 sack_rxmit = 0;
17614                                 goto out;
17615                         }
17616                 }
17617                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
17618                         if (rsm && (rsm->r_flags & RACK_TLP)) {
17619                                 /*
17620                                  * TLP should not count in retran count, but
17621                                  * in its own bin
17622                                  */
17623                                 counter_u64_add(rack_tlp_retran, 1);
17624                                 counter_u64_add(rack_tlp_retran_bytes, len);
17625                         } else {
17626                                 tp->t_sndrexmitpack++;
17627                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
17628                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
17629                         }
17630 #ifdef STATS
17631                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
17632                                                  len);
17633 #endif
17634                 } else {
17635                         KMOD_TCPSTAT_INC(tcps_sndpack);
17636                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
17637 #ifdef STATS
17638                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
17639                                                  len);
17640 #endif
17641                 }
17642                 /*
17643                  * If we're sending everything we've got, set PUSH. (This
17644                  * will keep happy those implementations which only give
17645                  * data to the user when a buffer fills or a PUSH comes in.)
17646                  */
17647                 if (sb_offset + len == sbused(sb) &&
17648                     sbused(sb) &&
17649                     !(flags & TH_SYN)) {
17650                         flags |= TH_PUSH;
17651                         add_flag |= RACK_HAD_PUSH;
17652                 }
17653
17654                 SOCKBUF_UNLOCK(sb);
17655         } else {
17656                 SOCKBUF_UNLOCK(sb);
17657                 if (tp->t_flags & TF_ACKNOW)
17658                         KMOD_TCPSTAT_INC(tcps_sndacks);
17659                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
17660                         KMOD_TCPSTAT_INC(tcps_sndctrl);
17661                 else
17662                         KMOD_TCPSTAT_INC(tcps_sndwinup);
17663
17664                 m = m_gethdr(M_NOWAIT, MT_DATA);
17665                 if (m == NULL) {
17666                         error = ENOBUFS;
17667                         sack_rxmit = 0;
17668                         goto out;
17669                 }
17670 #ifdef INET6
17671                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
17672                     MHLEN >= hdrlen) {
17673                         M_ALIGN(m, hdrlen);
17674                 } else
17675 #endif
17676                         m->m_data += max_linkhdr;
17677                 m->m_len = hdrlen;
17678         }
17679         SOCKBUF_UNLOCK_ASSERT(sb);
17680         m->m_pkthdr.rcvif = (struct ifnet *)0;
17681 #ifdef MAC
17682         mac_inpcb_create_mbuf(inp, m);
17683 #endif
17684         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
17685 #ifdef INET6
17686                 if (isipv6)
17687                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
17688                 else
17689 #endif                          /* INET6 */
17690                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
17691                 th = rack->r_ctl.fsb.th;
17692                 udp = rack->r_ctl.fsb.udp;
17693                 if (udp) {
17694 #ifdef INET6
17695                         if (isipv6)
17696                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17697                         else
17698 #endif                          /* INET6 */
17699                                 ulen = hdrlen + len - sizeof(struct ip);
17700                         udp->uh_ulen = htons(ulen);
17701                 }
17702         } else {
17703 #ifdef INET6
17704                 if (isipv6) {
17705                         ip6 = mtod(m, struct ip6_hdr *);
17706                         if (tp->t_port) {
17707                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
17708                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17709                                 udp->uh_dport = tp->t_port;
17710                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17711                                 udp->uh_ulen = htons(ulen);
17712                                 th = (struct tcphdr *)(udp + 1);
17713                         } else
17714                                 th = (struct tcphdr *)(ip6 + 1);
17715                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
17716                 } else
17717 #endif                          /* INET6 */
17718                 {
17719                         ip = mtod(m, struct ip *);
17720 #ifdef TCPDEBUG
17721                         ipov = (struct ipovly *)ip;
17722 #endif
17723                         if (tp->t_port) {
17724                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
17725                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17726                                 udp->uh_dport = tp->t_port;
17727                                 ulen = hdrlen + len - sizeof(struct ip);
17728                                 udp->uh_ulen = htons(ulen);
17729                                 th = (struct tcphdr *)(udp + 1);
17730                         } else
17731                                 th = (struct tcphdr *)(ip + 1);
17732                         tcpip_fillheaders(inp, tp->t_port, ip, th);
17733                 }
17734         }
17735         /*
17736          * Fill in fields, remembering maximum advertised window for use in
17737          * delaying messages about window sizes. If resending a FIN, be sure
17738          * not to use a new sequence number.
17739          */
17740         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
17741             tp->snd_nxt == tp->snd_max)
17742                 tp->snd_nxt--;
17743         /*
17744          * If we are starting a connection, send ECN setup SYN packet. If we
17745          * are on a retransmit, we may resend those bits a number of times
17746          * as per RFC 3168.
17747          */
17748         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
17749                 if (tp->t_rxtshift >= 1) {
17750                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
17751                                 flags |= TH_ECE | TH_CWR;
17752                 } else
17753                         flags |= TH_ECE | TH_CWR;
17754         }
17755         /* Handle parallel SYN for ECN */
17756         if ((tp->t_state == TCPS_SYN_RECEIVED) &&
17757             (tp->t_flags2 & TF2_ECN_SND_ECE)) {
17758                 flags |= TH_ECE;
17759                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
17760         }
17761         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17762             (tp->t_flags2 & TF2_ECN_PERMIT)) {
17763                 /*
17764                  * If the peer has ECN, mark data packets with ECN capable
17765                  * transmission (ECT). Ignore pure ack packets,
17766                  * retransmissions.
17767                  */
17768                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
17769                     (sack_rxmit == 0)) {
17770 #ifdef INET6
17771                         if (isipv6)
17772                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
17773                         else
17774 #endif
17775                                 ip->ip_tos |= IPTOS_ECN_ECT0;
17776                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
17777                         /*
17778                          * Reply with proper ECN notifications.
17779                          * Only set CWR on new data segments.
17780                          */
17781                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
17782                                 flags |= TH_CWR;
17783                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
17784                         }
17785                 }
17786                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
17787                         flags |= TH_ECE;
17788         }
17789         /*
17790          * If we are doing retransmissions, then snd_nxt will not reflect
17791          * the first unsent octet.  For ACK only packets, we do not want the
17792          * sequence number of the retransmitted packet, we want the sequence
17793          * number of the next unsent octet.  So, if there is no data (and no
17794          * SYN or FIN), use snd_max instead of snd_nxt when filling in
17795          * ti_seq.  But if we are in persist state, snd_max might reflect
17796          * one byte beyond the right edge of the window, so use snd_nxt in
17797          * that case, since we know we aren't doing a retransmission.
17798          * (retransmit and persist are mutually exclusive...)
17799          */
17800         if (sack_rxmit == 0) {
17801                 if (len || (flags & (TH_SYN | TH_FIN))) {
17802                         th->th_seq = htonl(tp->snd_nxt);
17803                         rack_seq = tp->snd_nxt;
17804                 } else {
17805                         th->th_seq = htonl(tp->snd_max);
17806                         rack_seq = tp->snd_max;
17807                 }
17808         } else {
17809                 th->th_seq = htonl(rsm->r_start);
17810                 rack_seq = rsm->r_start;
17811         }
17812         th->th_ack = htonl(tp->rcv_nxt);
17813         th->th_flags = flags;
17814         /*
17815          * Calculate receive window.  Don't shrink window, but avoid silly
17816          * window syndrome.
17817          * If a RST segment is sent, advertise a window of zero.
17818          */
17819         if (flags & TH_RST) {
17820                 recwin = 0;
17821         } else {
17822                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
17823                     recwin < (long)segsiz) {
17824                         recwin = 0;
17825                 }
17826                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
17827                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
17828                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
17829         }
17830
17831         /*
17832          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
17833          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
17834          * handled in syncache.
17835          */
17836         if (flags & TH_SYN)
17837                 th->th_win = htons((u_short)
17838                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
17839         else {
17840                 /* Avoid shrinking window with window scaling. */
17841                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
17842                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
17843         }
17844         /*
17845          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
17846          * window.  This may cause the remote transmitter to stall.  This
17847          * flag tells soreceive() to disable delayed acknowledgements when
17848          * draining the buffer.  This can occur if the receiver is
17849          * attempting to read more data than can be buffered prior to
17850          * transmitting on the connection.
17851          */
17852         if (th->th_win == 0) {
17853                 tp->t_sndzerowin++;
17854                 tp->t_flags |= TF_RXWIN0SENT;
17855         } else
17856                 tp->t_flags &= ~TF_RXWIN0SENT;
17857         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
17858         /* Now are we using fsb?, if so copy the template data to the mbuf */
17859         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
17860                 uint8_t *cpto;
17861
17862                 cpto = mtod(m, uint8_t *);
17863                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
17864                 /*
17865                  * We have just copied in:
17866                  * IP/IP6
17867                  * <optional udphdr>
17868                  * tcphdr (no options)
17869                  *
17870                  * We need to grab the correct pointers into the mbuf
17871                  * for both the tcp header, and possibly the udp header (if tunneling).
17872                  * We do this by using the offset in the copy buffer and adding it
17873                  * to the mbuf base pointer (cpto).
17874                  */
17875 #ifdef INET6
17876                 if (isipv6)
17877                         ip6 = mtod(m, struct ip6_hdr *);
17878                 else
17879 #endif                          /* INET6 */
17880                         ip = mtod(m, struct ip *);
17881                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
17882                 /* If we have a udp header lets set it into the mbuf as well */
17883                 if (udp)
17884                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
17885         }
17886 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17887         if (to.to_flags & TOF_SIGNATURE) {
17888                 /*
17889                  * Calculate MD5 signature and put it into the place
17890                  * determined before.
17891                  * NOTE: since TCP options buffer doesn't point into
17892                  * mbuf's data, calculate offset and use it.
17893                  */
17894                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
17895                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
17896                         /*
17897                          * Do not send segment if the calculation of MD5
17898                          * digest has failed.
17899                          */
17900                         goto out;
17901                 }
17902         }
17903 #endif
17904         if (optlen) {
17905                 bcopy(opt, th + 1, optlen);
17906                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
17907         }
17908         /*
17909          * Put TCP length in extended header, and then checksum extended
17910          * header and data.
17911          */
17912         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
17913 #ifdef INET6
17914         if (isipv6) {
17915                 /*
17916                  * ip6_plen is not need to be filled now, and will be filled
17917                  * in ip6_output.
17918                  */
17919                 if (tp->t_port) {
17920                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
17921                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17922                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
17923                         th->th_sum = htons(0);
17924                         UDPSTAT_INC(udps_opackets);
17925                 } else {
17926                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
17927                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17928                         th->th_sum = in6_cksum_pseudo(ip6,
17929                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
17930                                                       0);
17931                 }
17932         }
17933 #endif
17934 #if defined(INET6) && defined(INET)
17935         else
17936 #endif
17937 #ifdef INET
17938         {
17939                 if (tp->t_port) {
17940                         m->m_pkthdr.csum_flags = CSUM_UDP;
17941                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17942                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
17943                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
17944                         th->th_sum = htons(0);
17945                         UDPSTAT_INC(udps_opackets);
17946                 } else {
17947                         m->m_pkthdr.csum_flags = CSUM_TCP;
17948                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17949                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
17950                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
17951                                                                         IPPROTO_TCP + len + optlen));
17952                 }
17953                 /* IP version must be set here for ipv4/ipv6 checking later */
17954                 KASSERT(ip->ip_v == IPVERSION,
17955                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
17956         }
17957 #endif
17958         /*
17959          * Enable TSO and specify the size of the segments. The TCP pseudo
17960          * header checksum is always provided. XXX: Fixme: This is currently
17961          * not the case for IPv6.
17962          */
17963         if (tso) {
17964                 KASSERT(len > tp->t_maxseg - optlen,
17965                         ("%s: len <= tso_segsz", __func__));
17966                 m->m_pkthdr.csum_flags |= CSUM_TSO;
17967                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
17968         }
17969         KASSERT(len + hdrlen == m_length(m, NULL),
17970                 ("%s: mbuf chain different than expected: %d + %u != %u",
17971                  __func__, len, hdrlen, m_length(m, NULL)));
17972
17973 #ifdef TCP_HHOOK
17974         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
17975         hhook_run_tcp_est_out(tp, th, &to, len, tso);
17976 #endif
17977         /* We're getting ready to send; log now. */
17978         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
17979                 union tcp_log_stackspecific log;
17980
17981                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17982                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
17983                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
17984                 if (rack->rack_no_prr)
17985                         log.u_bbr.flex1 = 0;
17986                 else
17987                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
17988                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
17989                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
17990                 log.u_bbr.flex4 = orig_len;
17991                 if (filled_all)
17992                         log.u_bbr.flex5 = 0x80000000;
17993                 else
17994                         log.u_bbr.flex5 = 0;
17995                 /* Save off the early/late values */
17996                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
17997                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
17998                 log.u_bbr.bw_inuse = rack_get_bw(rack);
17999                 if (rsm || sack_rxmit) {
18000                         if (doing_tlp)
18001                                 log.u_bbr.flex8 = 2;
18002                         else
18003                                 log.u_bbr.flex8 = 1;
18004                 } else {
18005                         log.u_bbr.flex8 = 0;
18006                 }
18007                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18008                 log.u_bbr.flex7 = mark;
18009                 log.u_bbr.flex7 <<= 8;
18010                 log.u_bbr.flex7 |= pass;
18011                 log.u_bbr.pkts_out = tp->t_maxseg;
18012                 log.u_bbr.timeStamp = cts;
18013                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18014                 log.u_bbr.lt_epoch = cwnd_to_use;
18015                 log.u_bbr.delivered = sendalot;
18016                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18017                                      len, &log, false, NULL, NULL, 0, &tv);
18018         } else
18019                 lgb = NULL;
18020
18021         /*
18022          * Fill in IP length and desired time to live and send to IP level.
18023          * There should be a better way to handle ttl and tos; we could keep
18024          * them in the template, but need a way to checksum without them.
18025          */
18026         /*
18027          * m->m_pkthdr.len should have been set before cksum calcuration,
18028          * because in6_cksum() need it.
18029          */
18030 #ifdef INET6
18031         if (isipv6) {
18032                 /*
18033                  * we separately set hoplimit for every segment, since the
18034                  * user might want to change the value via setsockopt. Also,
18035                  * desired default hop limit might be changed via Neighbor
18036                  * Discovery.
18037                  */
18038                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18039
18040                 /*
18041                  * Set the packet size here for the benefit of DTrace
18042                  * probes. ip6_output() will set it properly; it's supposed
18043                  * to include the option header lengths as well.
18044                  */
18045                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18046
18047                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18048                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18049                 else
18050                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18051
18052                 if (tp->t_state == TCPS_SYN_SENT)
18053                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18054
18055                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18056                 /* TODO: IPv6 IP6TOS_ECT bit on */
18057                 error = ip6_output(m,
18058 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18059                                    inp->in6p_outputopts,
18060 #else
18061                                    NULL,
18062 #endif
18063                                    &inp->inp_route6,
18064                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18065                                    NULL, NULL, inp);
18066
18067                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18068                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18069         }
18070 #endif                          /* INET6 */
18071 #if defined(INET) && defined(INET6)
18072         else
18073 #endif
18074 #ifdef INET
18075         {
18076                 ip->ip_len = htons(m->m_pkthdr.len);
18077 #ifdef INET6
18078                 if (inp->inp_vflag & INP_IPV6PROTO)
18079                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18080 #endif                          /* INET6 */
18081                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18082                 /*
18083                  * If we do path MTU discovery, then we set DF on every
18084                  * packet. This might not be the best thing to do according
18085                  * to RFC3390 Section 2. However the tcp hostcache migitates
18086                  * the problem so it affects only the first tcp connection
18087                  * with a host.
18088                  *
18089                  * NB: Don't set DF on small MTU/MSS to have a safe
18090                  * fallback.
18091                  */
18092                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18093                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18094                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18095                                 ip->ip_off |= htons(IP_DF);
18096                         }
18097                 } else {
18098                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18099                 }
18100
18101                 if (tp->t_state == TCPS_SYN_SENT)
18102                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18103
18104                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18105
18106                 error = ip_output(m,
18107 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18108                                   inp->inp_options,
18109 #else
18110                                   NULL,
18111 #endif
18112                                   &inp->inp_route,
18113                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18114                                   inp);
18115                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18116                         mtu = inp->inp_route.ro_nh->nh_mtu;
18117         }
18118 #endif                          /* INET */
18119
18120 out:
18121         if (lgb) {
18122                 lgb->tlb_errno = error;
18123                 lgb = NULL;
18124         }
18125         /*
18126          * In transmit state, time the transmission and arrange for the
18127          * retransmit.  In persist state, just set snd_max.
18128          */
18129         if (error == 0) {
18130                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18131                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18132                 if (rsm && (doing_tlp == 0)) {
18133                         /* Set we retransmitted */
18134                         rack->rc_gp_saw_rec = 1;
18135                 } else {
18136                         if (cwnd_to_use > tp->snd_ssthresh) {
18137                                 /* Set we sent in CA */
18138                                 rack->rc_gp_saw_ca = 1;
18139                         } else {
18140                                 /* Set we sent in SS */
18141                                 rack->rc_gp_saw_ss = 1;
18142                         }
18143                 }
18144                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18145                     (tp->t_flags & TF_SACK_PERMIT) &&
18146                     tp->rcv_numsacks > 0)
18147                         tcp_clean_dsack_blocks(tp);
18148                 tot_len_this_send += len;
18149                 if (len == 0)
18150                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18151                 else if (len == 1) {
18152                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18153                 } else if (len > 1) {
18154                         int idx;
18155
18156                         idx = (len / segsiz) + 3;
18157                         if (idx >= TCP_MSS_ACCT_ATIMER)
18158                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18159                         else
18160                                 counter_u64_add(rack_out_size[idx], 1);
18161                 }
18162         }
18163         if ((rack->rack_no_prr == 0) &&
18164             sub_from_prr &&
18165             (error == 0)) {
18166                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18167                         rack->r_ctl.rc_prr_sndcnt -= len;
18168                 else
18169                         rack->r_ctl.rc_prr_sndcnt = 0;
18170         }
18171         sub_from_prr = 0;
18172         if (doing_tlp && (rsm == NULL)) {
18173                 /* New send doing a TLP */
18174                 add_flag |= RACK_TLP;
18175         }
18176         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18177                         rack_to_usec_ts(&tv),
18178                         rsm, add_flag, s_mb, s_moff, hw_tls);
18179
18180
18181         if ((error == 0) &&
18182             (len > 0) &&
18183             (tp->snd_una == tp->snd_max))
18184                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18185         {
18186                 tcp_seq startseq = tp->snd_nxt;
18187
18188                 /* Track our lost count */
18189                 if (rsm && (doing_tlp == 0))
18190                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18191                 /*
18192                  * Advance snd_nxt over sequence space of this segment.
18193                  */
18194                 if (error)
18195                         /* We don't log or do anything with errors */
18196                         goto nomore;
18197                 if (doing_tlp == 0) {
18198                         if (rsm == NULL) {
18199                                 /*
18200                                  * Not a retransmission of some
18201                                  * sort, new data is going out so
18202                                  * clear our TLP count and flag.
18203                                  */
18204                                 rack->rc_tlp_in_progress = 0;
18205                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18206                         }
18207                 } else {
18208                         /*
18209                          * We have just sent a TLP, mark that it is true
18210                          * and make sure our in progress is set so we
18211                          * continue to check the count.
18212                          */
18213                         rack->rc_tlp_in_progress = 1;
18214                         rack->r_ctl.rc_tlp_cnt_out++;
18215                 }
18216                 if (flags & (TH_SYN | TH_FIN)) {
18217                         if (flags & TH_SYN)
18218                                 tp->snd_nxt++;
18219                         if (flags & TH_FIN) {
18220                                 tp->snd_nxt++;
18221                                 tp->t_flags |= TF_SENTFIN;
18222                         }
18223                 }
18224                 /* In the ENOBUFS case we do *not* update snd_max */
18225                 if (sack_rxmit)
18226                         goto nomore;
18227
18228                 tp->snd_nxt += len;
18229                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18230                         if (tp->snd_una == tp->snd_max) {
18231                                 /*
18232                                  * Update the time we just added data since
18233                                  * none was outstanding.
18234                                  */
18235                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18236                                 tp->t_acktime = ticks;
18237                         }
18238                         tp->snd_max = tp->snd_nxt;
18239                         /*
18240                          * Time this transmission if not a retransmission and
18241                          * not currently timing anything.
18242                          * This is only relevant in case of switching back to
18243                          * the base stack.
18244                          */
18245                         if (tp->t_rtttime == 0) {
18246                                 tp->t_rtttime = ticks;
18247                                 tp->t_rtseq = startseq;
18248                                 KMOD_TCPSTAT_INC(tcps_segstimed);
18249                         }
18250                         if (len &&
18251                             ((tp->t_flags & TF_GPUTINPROG) == 0))
18252                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18253                 }
18254                 /*
18255                  * If we are doing FO we need to update the mbuf position and subtract
18256                  * this happens when the peer sends us duplicate information and
18257                  * we thus want to send a DSACK.
18258                  *
18259                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18260                  * turned off? If not then we are going to echo multiple DSACK blocks
18261                  * out (with the TSO), which we should not be doing.
18262                  */
18263                 if (rack->r_fast_output && len) {
18264                         if (rack->r_ctl.fsb.left_to_send > len)
18265                                 rack->r_ctl.fsb.left_to_send -= len;
18266                         else
18267                                 rack->r_ctl.fsb.left_to_send = 0;
18268                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18269                                 rack->r_fast_output = 0;
18270                         if (rack->r_fast_output) {
18271                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18272                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18273                         }
18274                 }
18275         }
18276 nomore:
18277         if (error) {
18278                 rack->r_ctl.rc_agg_delayed = 0;
18279                 rack->r_early = 0;
18280                 rack->r_late = 0;
18281                 rack->r_ctl.rc_agg_early = 0;
18282                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
18283                 /*
18284                  * Failures do not advance the seq counter above. For the
18285                  * case of ENOBUFS we will fall out and retry in 1ms with
18286                  * the hpts. Everything else will just have to retransmit
18287                  * with the timer.
18288                  *
18289                  * In any case, we do not want to loop around for another
18290                  * send without a good reason.
18291                  */
18292                 sendalot = 0;
18293                 switch (error) {
18294                 case EPERM:
18295                         tp->t_softerror = error;
18296 #ifdef TCP_ACCOUNTING
18297                         crtsc = get_cyclecount();
18298                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18299                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18300                         }
18301                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18302                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18303                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18304                         }
18305                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18306                         sched_unpin();
18307 #endif
18308                         return (error);
18309                 case ENOBUFS:
18310                         /*
18311                          * Pace us right away to retry in a some
18312                          * time
18313                          */
18314                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18315                         if (rack->rc_enobuf < 0x7f)
18316                                 rack->rc_enobuf++;
18317                         if (slot < (10 * HPTS_USEC_IN_MSEC))
18318                                 slot = 10 * HPTS_USEC_IN_MSEC;
18319                         if (rack->r_ctl.crte != NULL) {
18320                                 counter_u64_add(rack_saw_enobuf_hw, 1);
18321                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
18322                         }
18323                         counter_u64_add(rack_saw_enobuf, 1);
18324                         goto enobufs;
18325                 case EMSGSIZE:
18326                         /*
18327                          * For some reason the interface we used initially
18328                          * to send segments changed to another or lowered
18329                          * its MTU. If TSO was active we either got an
18330                          * interface without TSO capabilits or TSO was
18331                          * turned off. If we obtained mtu from ip_output()
18332                          * then update it and try again.
18333                          */
18334                         if (tso)
18335                                 tp->t_flags &= ~TF_TSO;
18336                         if (mtu != 0) {
18337                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
18338                                 goto again;
18339                         }
18340                         slot = 10 * HPTS_USEC_IN_MSEC;
18341                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18342 #ifdef TCP_ACCOUNTING
18343                         crtsc = get_cyclecount();
18344                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18345                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18346                         }
18347                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18348                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18349                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18350                         }
18351                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18352                         sched_unpin();
18353 #endif
18354                         return (error);
18355                 case ENETUNREACH:
18356                         counter_u64_add(rack_saw_enetunreach, 1);
18357                 case EHOSTDOWN:
18358                 case EHOSTUNREACH:
18359                 case ENETDOWN:
18360                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
18361                                 tp->t_softerror = error;
18362                         }
18363                         /* FALLTHROUGH */
18364                 default:
18365                         slot = 10 * HPTS_USEC_IN_MSEC;
18366                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18367 #ifdef TCP_ACCOUNTING
18368                         crtsc = get_cyclecount();
18369                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18370                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18371                         }
18372                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18373                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18374                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18375                         }
18376                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18377                         sched_unpin();
18378 #endif
18379                         return (error);
18380                 }
18381         } else {
18382                 rack->rc_enobuf = 0;
18383                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18384                         rack->r_ctl.retran_during_recovery += len;
18385         }
18386         KMOD_TCPSTAT_INC(tcps_sndtotal);
18387
18388         /*
18389          * Data sent (as far as we can tell). If this advertises a larger
18390          * window than any other segment, then remember the size of the
18391          * advertised window. Any pending ACK has now been sent.
18392          */
18393         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18394                 tp->rcv_adv = tp->rcv_nxt + recwin;
18395
18396         tp->last_ack_sent = tp->rcv_nxt;
18397         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18398 enobufs:
18399         if (sendalot) {
18400                 /* Do we need to turn off sendalot? */
18401                 if (rack->r_ctl.rc_pace_max_segs &&
18402                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18403                         /* We hit our max. */
18404                         sendalot = 0;
18405                 } else if ((rack->rc_user_set_max_segs) &&
18406                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18407                         /* We hit the user defined max */
18408                         sendalot = 0;
18409                 }
18410         }
18411         if ((error == 0) && (flags & TH_FIN))
18412                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18413         if (flags & TH_RST) {
18414                 /*
18415                  * We don't send again after sending a RST.
18416                  */
18417                 slot = 0;
18418                 sendalot = 0;
18419                 if (error == 0)
18420                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18421         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18422                 /*
18423                  * Get our pacing rate, if an error
18424                  * occurred in sending (ENOBUF) we would
18425                  * hit the else if with slot preset. Other
18426                  * errors return.
18427                  */
18428                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18429         }
18430         if (rsm &&
18431             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18432             rack->use_rack_rr) {
18433                 /* Its a retransmit and we use the rack cheat? */
18434                 if ((slot == 0) ||
18435                     (rack->rc_always_pace == 0) ||
18436                     (rack->r_rr_config == 1)) {
18437                         /*
18438                          * We have no pacing set or we
18439                          * are using old-style rack or
18440                          * we are overriden to use the old 1ms pacing.
18441                          */
18442                         slot = rack->r_ctl.rc_min_to;
18443                 }
18444         }
18445         /* We have sent clear the flag */
18446         rack->r_ent_rec_ns = 0;
18447         if (rack->r_must_retran) {
18448                 if (rsm) {
18449                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
18450                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
18451                                 /*
18452                                  * We have retransmitted all.
18453                                  */
18454                                 rack->r_must_retran = 0;
18455                                 rack->r_ctl.rc_out_at_rto = 0;
18456                         }
18457                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18458                         /*
18459                          * Sending new data will also kill
18460                          * the loop.
18461                          */
18462                         rack->r_must_retran = 0;
18463                         rack->r_ctl.rc_out_at_rto = 0;
18464                 }
18465         }
18466         rack->r_ctl.fsb.recwin = recwin;
18467         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
18468             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18469                 /*
18470                  * We hit an RTO and now have past snd_max at the RTO
18471                  * clear all the WAS flags.
18472                  */
18473                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
18474         }
18475         if (slot) {
18476                 /* set the rack tcb into the slot N */
18477                 counter_u64_add(rack_paced_segments, 1);
18478                 if ((error == 0) &&
18479                     rack_use_rfo &&
18480                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18481                     (rsm == NULL) &&
18482                     (tp->snd_nxt == tp->snd_max) &&
18483                     (ipoptlen == 0) &&
18484                     (tp->rcv_numsacks == 0) &&
18485                     rack->r_fsb_inited &&
18486                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18487                     (rack->r_must_retran == 0) &&
18488                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18489                     (len > 0) && (orig_len > 0) &&
18490                     (orig_len > len) &&
18491                     ((orig_len - len) >= segsiz) &&
18492                     ((optlen == 0) ||
18493                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18494                         /* We can send at least one more MSS using our fsb */
18495
18496                         rack->r_fast_output = 1;
18497                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18498                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18499                         rack->r_ctl.fsb.tcp_flags = flags;
18500                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18501                         if (hw_tls)
18502                                 rack->r_ctl.fsb.hw_tls = 1;
18503                         else
18504                                 rack->r_ctl.fsb.hw_tls = 0;
18505                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18506                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18507                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18508                                  (tp->snd_max - tp->snd_una)));
18509                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18510                                 rack->r_fast_output = 0;
18511                         else {
18512                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18513                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18514                                 else
18515                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18516                         }
18517                 } else
18518                         rack->r_fast_output = 0;
18519                 rack_log_fsb(rack, tp, so, flags,
18520                              ipoptlen, orig_len, len, error,
18521                              (rsm == NULL), optlen, __LINE__, 2);
18522         } else if (sendalot) {
18523                 int ret;
18524
18525                 if (len)
18526                         counter_u64_add(rack_unpaced_segments, 1);
18527                 sack_rxmit = 0;
18528                 if ((error == 0) &&
18529                     rack_use_rfo &&
18530                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18531                     (rsm == NULL) &&
18532                     (ipoptlen == 0) &&
18533                     (tp->rcv_numsacks == 0) &&
18534                     (tp->snd_nxt == tp->snd_max) &&
18535                     (rack->r_must_retran == 0) &&
18536                     rack->r_fsb_inited &&
18537                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18538                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18539                     (len > 0) && (orig_len > 0) &&
18540                     (orig_len > len) &&
18541                     ((orig_len - len) >= segsiz) &&
18542                     ((optlen == 0) ||
18543                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18544                         /* we can use fast_output for more */
18545
18546                         rack->r_fast_output = 1;
18547                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18548                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18549                         rack->r_ctl.fsb.tcp_flags = flags;
18550                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18551                         if (hw_tls)
18552                                 rack->r_ctl.fsb.hw_tls = 1;
18553                         else
18554                                 rack->r_ctl.fsb.hw_tls = 0;
18555                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18556                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18557                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18558                                  (tp->snd_max - tp->snd_una)));
18559                         if (rack->r_ctl.fsb.left_to_send < segsiz) {
18560                                 rack->r_fast_output = 0;
18561                         }
18562                         if (rack->r_fast_output) {
18563                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18564                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18565                                 else
18566                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18567                                 rack_log_fsb(rack, tp, so, flags,
18568                                              ipoptlen, orig_len, len, error,
18569                                              (rsm == NULL), optlen, __LINE__, 3);
18570                                 error = 0;
18571                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
18572                                 if (ret >= 0)
18573                                         return (ret);
18574                                 else if (error)
18575                                         goto nomore;
18576
18577                         }
18578                 }
18579                 goto again;
18580         } else if (len) {
18581                 counter_u64_add(rack_unpaced_segments, 1);
18582         }
18583         /* Assure when we leave that snd_nxt will point to top */
18584         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
18585                 tp->snd_nxt = tp->snd_max;
18586         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
18587 #ifdef TCP_ACCOUNTING
18588         crtsc = get_cyclecount() - ts_val;
18589         if (tot_len_this_send) {
18590                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18591                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
18592                 }
18593                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
18594                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18595                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
18596                 }
18597                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
18598                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18599                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
18600                 }
18601                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
18602         } else {
18603                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18604                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
18605                 }
18606                 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
18607                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18608                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
18609                 }
18610                 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
18611         }
18612         sched_unpin();
18613 #endif
18614         if (error == ENOBUFS)
18615                 error = 0;
18616         return (error);
18617 }
18618
18619 static void
18620 rack_update_seg(struct tcp_rack *rack)
18621 {
18622         uint32_t orig_val;
18623
18624         orig_val = rack->r_ctl.rc_pace_max_segs;
18625         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
18626         if (orig_val != rack->r_ctl.rc_pace_max_segs)
18627                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL);
18628 }
18629
18630 static void
18631 rack_mtu_change(struct tcpcb *tp)
18632 {
18633         /*
18634          * The MSS may have changed
18635          */
18636         struct tcp_rack *rack;
18637
18638         rack = (struct tcp_rack *)tp->t_fb_ptr;
18639         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
18640                 /*
18641                  * The MTU has changed we need to resend everything
18642                  * since all we have sent is lost. We first fix
18643                  * up the mtu though.
18644                  */
18645                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
18646                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
18647                 rack_remxt_tmr(tp);
18648                 rack->r_fast_output = 0;
18649                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
18650                                                 rack->r_ctl.rc_sacked);
18651                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
18652                 rack->r_must_retran = 1;
18653
18654         }
18655         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
18656         /* We don't use snd_nxt to retransmit */
18657         tp->snd_nxt = tp->snd_max;
18658 }
18659
18660 static int
18661 rack_set_profile(struct tcp_rack *rack, int prof)
18662 {
18663         int err = EINVAL;
18664         if (prof == 1) {
18665                 /* pace_always=1 */
18666                 if (rack->rc_always_pace == 0) {
18667                         if (tcp_can_enable_pacing() == 0)
18668                                 return (EBUSY);
18669                 }
18670                 rack->rc_always_pace = 1;
18671                 if (rack->use_fixed_rate || rack->gp_ready)
18672                         rack_set_cc_pacing(rack);
18673                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18674                 rack->rack_attempt_hdwr_pace = 0;
18675                 /* cmpack=1 */
18676                 if (rack_use_cmp_acks)
18677                         rack->r_use_cmp_ack = 1;
18678                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18679                     rack->r_use_cmp_ack)
18680                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18681                 /* scwnd=1 */
18682                 rack->rack_enable_scwnd = 1;
18683                 /* dynamic=100 */
18684                 rack->rc_gp_dyn_mul = 1;
18685                 /* gp_inc_ca */
18686                 rack->r_ctl.rack_per_of_gp_ca = 100;
18687                 /* rrr_conf=3 */
18688                 rack->r_rr_config = 3;
18689                 /* npush=2 */
18690                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18691                 /* fillcw=1 */
18692                 rack->rc_pace_to_cwnd = 1;
18693                 rack->rc_pace_fill_if_rttin_range = 0;
18694                 rack->rtt_limit_mul = 0;
18695                 /* noprr=1 */
18696                 rack->rack_no_prr = 1;
18697                 /* lscwnd=1 */
18698                 rack->r_limit_scw = 1;
18699                 /* gp_inc_rec */
18700                 rack->r_ctl.rack_per_of_gp_rec = 90;
18701                 err = 0;
18702
18703         } else if (prof == 3) {
18704                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
18705                 /* pace_always=1 */
18706                 if (rack->rc_always_pace == 0) {
18707                         if (tcp_can_enable_pacing() == 0)
18708                                 return (EBUSY);
18709                 }
18710                 rack->rc_always_pace = 1;
18711                 if (rack->use_fixed_rate || rack->gp_ready)
18712                         rack_set_cc_pacing(rack);
18713                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18714                 rack->rack_attempt_hdwr_pace = 0;
18715                 /* cmpack=1 */
18716                 if (rack_use_cmp_acks)
18717                         rack->r_use_cmp_ack = 1;
18718                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18719                     rack->r_use_cmp_ack)
18720                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18721                 /* scwnd=1 */
18722                 rack->rack_enable_scwnd = 1;
18723                 /* dynamic=100 */
18724                 rack->rc_gp_dyn_mul = 1;
18725                 /* gp_inc_ca */
18726                 rack->r_ctl.rack_per_of_gp_ca = 100;
18727                 /* rrr_conf=3 */
18728                 rack->r_rr_config = 3;
18729                 /* npush=2 */
18730                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18731                 /* fillcw=2 */
18732                 rack->rc_pace_to_cwnd = 1;
18733                 rack->r_fill_less_agg = 1;
18734                 rack->rc_pace_fill_if_rttin_range = 0;
18735                 rack->rtt_limit_mul = 0;
18736                 /* noprr=1 */
18737                 rack->rack_no_prr = 1;
18738                 /* lscwnd=1 */
18739                 rack->r_limit_scw = 1;
18740                 /* gp_inc_rec */
18741                 rack->r_ctl.rack_per_of_gp_rec = 90;
18742                 err = 0;
18743
18744
18745         } else if (prof == 2) {
18746                 /* cmpack=1 */
18747                 if (rack->rc_always_pace == 0) {
18748                         if (tcp_can_enable_pacing() == 0)
18749                                 return (EBUSY);
18750                 }
18751                 rack->rc_always_pace = 1;
18752                 if (rack->use_fixed_rate || rack->gp_ready)
18753                         rack_set_cc_pacing(rack);
18754                 rack->r_use_cmp_ack = 1;
18755                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18756                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18757                 /* pace_always=1 */
18758                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18759                 /* scwnd=1 */
18760                 rack->rack_enable_scwnd = 1;
18761                 /* dynamic=100 */
18762                 rack->rc_gp_dyn_mul = 1;
18763                 rack->r_ctl.rack_per_of_gp_ca = 100;
18764                 /* rrr_conf=3 */
18765                 rack->r_rr_config = 3;
18766                 /* npush=2 */
18767                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18768                 /* fillcw=1 */
18769                 rack->rc_pace_to_cwnd = 1;
18770                 rack->rc_pace_fill_if_rttin_range = 0;
18771                 rack->rtt_limit_mul = 0;
18772                 /* noprr=1 */
18773                 rack->rack_no_prr = 1;
18774                 /* lscwnd=0 */
18775                 rack->r_limit_scw = 0;
18776                 err = 0;
18777         } else if (prof == 0) {
18778                 /* This changes things back to the default settings */
18779                 err = 0;
18780                 if (rack->rc_always_pace) {
18781                         tcp_decrement_paced_conn();
18782                         rack_undo_cc_pacing(rack);
18783                         rack->rc_always_pace = 0;
18784                 }
18785                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
18786                         rack->rc_always_pace = 1;
18787                         if (rack->use_fixed_rate || rack->gp_ready)
18788                                 rack_set_cc_pacing(rack);
18789                 } else
18790                         rack->rc_always_pace = 0;
18791                 if (rack_use_cmp_acks)
18792                         rack->r_use_cmp_ack = 1;
18793                 else
18794                         rack->r_use_cmp_ack = 0;
18795                 if (rack_disable_prr)
18796                         rack->rack_no_prr = 1;
18797                 else
18798                         rack->rack_no_prr = 0;
18799                 if (rack_gp_no_rec_chg)
18800                         rack->rc_gp_no_rec_chg = 1;
18801                 else
18802                         rack->rc_gp_no_rec_chg = 0;
18803                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
18804                         rack->r_mbuf_queue = 1;
18805                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18806                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18807                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18808                 } else {
18809                         rack->r_mbuf_queue = 0;
18810                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
18811                 }
18812                 if (rack_enable_shared_cwnd)
18813                         rack->rack_enable_scwnd = 1;
18814                 else
18815                         rack->rack_enable_scwnd = 0;
18816                 if (rack_do_dyn_mul) {
18817                         /* When dynamic adjustment is on CA needs to start at 100% */
18818                         rack->rc_gp_dyn_mul = 1;
18819                         if (rack_do_dyn_mul >= 100)
18820                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
18821                 } else {
18822                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
18823                         rack->rc_gp_dyn_mul = 0;
18824                 }
18825                 rack->r_rr_config = 0;
18826                 rack->r_ctl.rc_no_push_at_mrtt = 0;
18827                 rack->rc_pace_to_cwnd = 0;
18828                 rack->rc_pace_fill_if_rttin_range = 0;
18829                 rack->rtt_limit_mul = 0;
18830
18831                 if (rack_enable_hw_pacing)
18832                         rack->rack_hdw_pace_ena = 1;
18833                 else
18834                         rack->rack_hdw_pace_ena = 0;
18835                 if (rack_disable_prr)
18836                         rack->rack_no_prr = 1;
18837                 else
18838                         rack->rack_no_prr = 0;
18839                 if (rack_limits_scwnd)
18840                         rack->r_limit_scw  = 1;
18841                 else
18842                         rack->r_limit_scw  = 0;
18843                 err = 0;
18844         }
18845         return (err);
18846 }
18847
18848 static int
18849 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
18850 {
18851         struct deferred_opt_list *dol;
18852
18853         dol = malloc(sizeof(struct deferred_opt_list),
18854                      M_TCPFSB, M_NOWAIT|M_ZERO);
18855         if (dol == NULL) {
18856                 /*
18857                  * No space yikes -- fail out..
18858                  */
18859                 return (0);
18860         }
18861         dol->optname = sopt_name;
18862         dol->optval = loptval;
18863         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
18864         return (1);
18865 }
18866
18867 static int
18868 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
18869                     uint32_t optval, uint64_t loptval)
18870 {
18871         struct epoch_tracker et;
18872         struct sockopt sopt;
18873         struct cc_newreno_opts opt;
18874         uint64_t val;
18875         int error = 0;
18876         uint16_t ca, ss;
18877
18878         switch (sopt_name) {
18879
18880         case TCP_RACK_PACING_BETA:
18881                 RACK_OPTS_INC(tcp_rack_beta);
18882                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18883                         /* This only works for newreno. */
18884                         error = EINVAL;
18885                         break;
18886                 }
18887                 if (rack->rc_pacing_cc_set) {
18888                         /*
18889                          * Set them into the real CC module
18890                          * whats in the rack pcb is the old values
18891                          * to be used on restoral/
18892                          */
18893                         sopt.sopt_dir = SOPT_SET;
18894                         opt.name = CC_NEWRENO_BETA;
18895                         opt.val = optval;
18896                         if (CC_ALGO(tp)->ctl_output != NULL)
18897                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18898                         else {
18899                                 error = ENOENT;
18900                                 break;
18901                         }
18902                 } else {
18903                         /*
18904                          * Not pacing yet so set it into our local
18905                          * rack pcb storage.
18906                          */
18907                         rack->r_ctl.rc_saved_beta.beta = optval;
18908                 }
18909                 break;
18910         case TCP_RACK_TIMER_SLOP:
18911                 RACK_OPTS_INC(tcp_rack_timer_slop);
18912                 rack->r_ctl.timer_slop = optval;
18913                 if (rack->rc_tp->t_srtt) {
18914                         /*
18915                          * If we have an SRTT lets update t_rxtcur
18916                          * to have the new slop.
18917                          */
18918                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
18919                                            rack_rto_min, rack_rto_max,
18920                                            rack->r_ctl.timer_slop);
18921                 }
18922                 break;
18923         case TCP_RACK_PACING_BETA_ECN:
18924                 RACK_OPTS_INC(tcp_rack_beta_ecn);
18925                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18926                         /* This only works for newreno. */
18927                         error = EINVAL;
18928                         break;
18929                 }
18930                 if (rack->rc_pacing_cc_set) {
18931                         /*
18932                          * Set them into the real CC module
18933                          * whats in the rack pcb is the old values
18934                          * to be used on restoral/
18935                          */
18936                         sopt.sopt_dir = SOPT_SET;
18937                         opt.name = CC_NEWRENO_BETA_ECN;
18938                         opt.val = optval;
18939                         if (CC_ALGO(tp)->ctl_output != NULL)
18940                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18941                         else
18942                                 error = ENOENT;
18943                 } else {
18944                         /*
18945                          * Not pacing yet so set it into our local
18946                          * rack pcb storage.
18947                          */
18948                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
18949                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
18950                 }
18951                 break;
18952         case TCP_DEFER_OPTIONS:
18953                 RACK_OPTS_INC(tcp_defer_opt);
18954                 if (optval) {
18955                         if (rack->gp_ready) {
18956                                 /* Too late */
18957                                 error = EINVAL;
18958                                 break;
18959                         }
18960                         rack->defer_options = 1;
18961                 } else
18962                         rack->defer_options = 0;
18963                 break;
18964         case TCP_RACK_MEASURE_CNT:
18965                 RACK_OPTS_INC(tcp_rack_measure_cnt);
18966                 if (optval && (optval <= 0xff)) {
18967                         rack->r_ctl.req_measurements = optval;
18968                 } else
18969                         error = EINVAL;
18970                 break;
18971         case TCP_REC_ABC_VAL:
18972                 RACK_OPTS_INC(tcp_rec_abc_val);
18973                 if (optval > 0)
18974                         rack->r_use_labc_for_rec = 1;
18975                 else
18976                         rack->r_use_labc_for_rec = 0;
18977                 break;
18978         case TCP_RACK_ABC_VAL:
18979                 RACK_OPTS_INC(tcp_rack_abc_val);
18980                 if ((optval > 0) && (optval < 255))
18981                         rack->rc_labc = optval;
18982                 else
18983                         error = EINVAL;
18984                 break;
18985         case TCP_HDWR_UP_ONLY:
18986                 RACK_OPTS_INC(tcp_pacing_up_only);
18987                 if (optval)
18988                         rack->r_up_only = 1;
18989                 else
18990                         rack->r_up_only = 0;
18991                 break;
18992         case TCP_PACING_RATE_CAP:
18993                 RACK_OPTS_INC(tcp_pacing_rate_cap);
18994                 rack->r_ctl.bw_rate_cap = loptval;
18995                 break;
18996         case TCP_RACK_PROFILE:
18997                 RACK_OPTS_INC(tcp_profile);
18998                 error = rack_set_profile(rack, optval);
18999                 break;
19000         case TCP_USE_CMP_ACKS:
19001                 RACK_OPTS_INC(tcp_use_cmp_acks);
19002                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19003                         /* You can't turn it off once its on! */
19004                         error = EINVAL;
19005                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19006                         rack->r_use_cmp_ack = 1;
19007                         rack->r_mbuf_queue = 1;
19008                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19009                 }
19010                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19011                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19012                 break;
19013         case TCP_SHARED_CWND_TIME_LIMIT:
19014                 RACK_OPTS_INC(tcp_lscwnd);
19015                 if (optval)
19016                         rack->r_limit_scw = 1;
19017                 else
19018                         rack->r_limit_scw = 0;
19019                 break;
19020         case TCP_RACK_PACE_TO_FILL:
19021                 RACK_OPTS_INC(tcp_fillcw);
19022                 if (optval == 0)
19023                         rack->rc_pace_to_cwnd = 0;
19024                 else {
19025                         rack->rc_pace_to_cwnd = 1;
19026                         if (optval > 1)
19027                                 rack->r_fill_less_agg = 1;
19028                 }
19029                 if ((optval >= rack_gp_rtt_maxmul) &&
19030                     rack_gp_rtt_maxmul &&
19031                     (optval < 0xf)) {
19032                         rack->rc_pace_fill_if_rttin_range = 1;
19033                         rack->rtt_limit_mul = optval;
19034                 } else {
19035                         rack->rc_pace_fill_if_rttin_range = 0;
19036                         rack->rtt_limit_mul = 0;
19037                 }
19038                 break;
19039         case TCP_RACK_NO_PUSH_AT_MAX:
19040                 RACK_OPTS_INC(tcp_npush);
19041                 if (optval == 0)
19042                         rack->r_ctl.rc_no_push_at_mrtt = 0;
19043                 else if (optval < 0xff)
19044                         rack->r_ctl.rc_no_push_at_mrtt = optval;
19045                 else
19046                         error = EINVAL;
19047                 break;
19048         case TCP_SHARED_CWND_ENABLE:
19049                 RACK_OPTS_INC(tcp_rack_scwnd);
19050                 if (optval == 0)
19051                         rack->rack_enable_scwnd = 0;
19052                 else
19053                         rack->rack_enable_scwnd = 1;
19054                 break;
19055         case TCP_RACK_MBUF_QUEUE:
19056                 /* Now do we use the LRO mbuf-queue feature */
19057                 RACK_OPTS_INC(tcp_rack_mbufq);
19058                 if (optval || rack->r_use_cmp_ack)
19059                         rack->r_mbuf_queue = 1;
19060                 else
19061                         rack->r_mbuf_queue = 0;
19062                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19063                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19064                 else
19065                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19066                 break;
19067         case TCP_RACK_NONRXT_CFG_RATE:
19068                 RACK_OPTS_INC(tcp_rack_cfg_rate);
19069                 if (optval == 0)
19070                         rack->rack_rec_nonrxt_use_cr = 0;
19071                 else
19072                         rack->rack_rec_nonrxt_use_cr = 1;
19073                 break;
19074         case TCP_NO_PRR:
19075                 RACK_OPTS_INC(tcp_rack_noprr);
19076                 if (optval == 0)
19077                         rack->rack_no_prr = 0;
19078                 else if (optval == 1)
19079                         rack->rack_no_prr = 1;
19080                 else if (optval == 2)
19081                         rack->no_prr_addback = 1;
19082                 else
19083                         error = EINVAL;
19084                 break;
19085         case TCP_TIMELY_DYN_ADJ:
19086                 RACK_OPTS_INC(tcp_timely_dyn);
19087                 if (optval == 0)
19088                         rack->rc_gp_dyn_mul = 0;
19089                 else {
19090                         rack->rc_gp_dyn_mul = 1;
19091                         if (optval >= 100) {
19092                                 /*
19093                                  * If the user sets something 100 or more
19094                                  * its the gp_ca value.
19095                                  */
19096                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19097                         }
19098                 }
19099                 break;
19100         case TCP_RACK_DO_DETECTION:
19101                 RACK_OPTS_INC(tcp_rack_do_detection);
19102                 if (optval == 0)
19103                         rack->do_detection = 0;
19104                 else
19105                         rack->do_detection = 1;
19106                 break;
19107         case TCP_RACK_TLP_USE:
19108                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19109                         error = EINVAL;
19110                         break;
19111                 }
19112                 RACK_OPTS_INC(tcp_tlp_use);
19113                 rack->rack_tlp_threshold_use = optval;
19114                 break;
19115         case TCP_RACK_TLP_REDUCE:
19116                 /* RACK TLP cwnd reduction (bool) */
19117                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19118                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19119                 break;
19120         /*  Pacing related ones */
19121         case TCP_RACK_PACE_ALWAYS:
19122                 /*
19123                  * zero is old rack method, 1 is new
19124                  * method using a pacing rate.
19125                  */
19126                 RACK_OPTS_INC(tcp_rack_pace_always);
19127                 if (optval > 0) {
19128                         if (rack->rc_always_pace) {
19129                                 error = EALREADY;
19130                                 break;
19131                         } else if (tcp_can_enable_pacing()) {
19132                                 rack->rc_always_pace = 1;
19133                                 if (rack->use_fixed_rate || rack->gp_ready)
19134                                         rack_set_cc_pacing(rack);
19135                         }
19136                         else {
19137                                 error = ENOSPC;
19138                                 break;
19139                         }
19140                 } else {
19141                         if (rack->rc_always_pace) {
19142                                 tcp_decrement_paced_conn();
19143                                 rack->rc_always_pace = 0;
19144                                 rack_undo_cc_pacing(rack);
19145                         }
19146                 }
19147                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19148                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19149                 else
19150                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19151                 /* A rate may be set irate or other, if so set seg size */
19152                 rack_update_seg(rack);
19153                 break;
19154         case TCP_BBR_RACK_INIT_RATE:
19155                 RACK_OPTS_INC(tcp_initial_rate);
19156                 val = optval;
19157                 /* Change from kbits per second to bytes per second */
19158                 val *= 1000;
19159                 val /= 8;
19160                 rack->r_ctl.init_rate = val;
19161                 if (rack->rc_init_win != rack_default_init_window) {
19162                         uint32_t win, snt;
19163
19164                         /*
19165                          * Options don't always get applied
19166                          * in the order you think. So in order
19167                          * to assure we update a cwnd we need
19168                          * to check and see if we are still
19169                          * where we should raise the cwnd.
19170                          */
19171                         win = rc_init_window(rack);
19172                         if (SEQ_GT(tp->snd_max, tp->iss))
19173                                 snt = tp->snd_max - tp->iss;
19174                         else
19175                                 snt = 0;
19176                         if ((snt < win) &&
19177                             (tp->snd_cwnd < win))
19178                                 tp->snd_cwnd = win;
19179                 }
19180                 if (rack->rc_always_pace)
19181                         rack_update_seg(rack);
19182                 break;
19183         case TCP_BBR_IWINTSO:
19184                 RACK_OPTS_INC(tcp_initial_win);
19185                 if (optval && (optval <= 0xff)) {
19186                         uint32_t win, snt;
19187
19188                         rack->rc_init_win = optval;
19189                         win = rc_init_window(rack);
19190                         if (SEQ_GT(tp->snd_max, tp->iss))
19191                                 snt = tp->snd_max - tp->iss;
19192                         else
19193                                 snt = 0;
19194                         if ((snt < win) &&
19195                             (tp->t_srtt |
19196 #ifdef NETFLIX_PEAKRATE
19197                              tp->t_maxpeakrate |
19198 #endif
19199                              rack->r_ctl.init_rate)) {
19200                                 /*
19201                                  * We are not past the initial window
19202                                  * and we have some bases for pacing,
19203                                  * so we need to possibly adjust up
19204                                  * the cwnd. Note even if we don't set
19205                                  * the cwnd, its still ok to raise the rc_init_win
19206                                  * which can be used coming out of idle when we
19207                                  * would have a rate.
19208                                  */
19209                                 if (tp->snd_cwnd < win)
19210                                         tp->snd_cwnd = win;
19211                         }
19212                         if (rack->rc_always_pace)
19213                                 rack_update_seg(rack);
19214                 } else
19215                         error = EINVAL;
19216                 break;
19217         case TCP_RACK_FORCE_MSEG:
19218                 RACK_OPTS_INC(tcp_rack_force_max_seg);
19219                 if (optval)
19220                         rack->rc_force_max_seg = 1;
19221                 else
19222                         rack->rc_force_max_seg = 0;
19223                 break;
19224         case TCP_RACK_PACE_MAX_SEG:
19225                 /* Max segments size in a pace in bytes */
19226                 RACK_OPTS_INC(tcp_rack_max_seg);
19227                 rack->rc_user_set_max_segs = optval;
19228                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19229                 break;
19230         case TCP_RACK_PACE_RATE_REC:
19231                 /* Set the fixed pacing rate in Bytes per second ca */
19232                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19233                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19234                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19235                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19236                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19237                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19238                 rack->use_fixed_rate = 1;
19239                 if (rack->rc_always_pace)
19240                         rack_set_cc_pacing(rack);
19241                 rack_log_pacing_delay_calc(rack,
19242                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19243                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19244                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19245                                            __LINE__, NULL);
19246                 break;
19247
19248         case TCP_RACK_PACE_RATE_SS:
19249                 /* Set the fixed pacing rate in Bytes per second ca */
19250                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19251                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19252                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19253                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19254                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19255                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19256                 rack->use_fixed_rate = 1;
19257                 if (rack->rc_always_pace)
19258                         rack_set_cc_pacing(rack);
19259                 rack_log_pacing_delay_calc(rack,
19260                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19261                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19262                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19263                                            __LINE__, NULL);
19264                 break;
19265
19266         case TCP_RACK_PACE_RATE_CA:
19267                 /* Set the fixed pacing rate in Bytes per second ca */
19268                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19269                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19270                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19271                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19272                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19273                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19274                 rack->use_fixed_rate = 1;
19275                 if (rack->rc_always_pace)
19276                         rack_set_cc_pacing(rack);
19277                 rack_log_pacing_delay_calc(rack,
19278                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19279                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19280                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19281                                            __LINE__, NULL);
19282                 break;
19283         case TCP_RACK_GP_INCREASE_REC:
19284                 RACK_OPTS_INC(tcp_gp_inc_rec);
19285                 rack->r_ctl.rack_per_of_gp_rec = optval;
19286                 rack_log_pacing_delay_calc(rack,
19287                                            rack->r_ctl.rack_per_of_gp_ss,
19288                                            rack->r_ctl.rack_per_of_gp_ca,
19289                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19290                                            __LINE__, NULL);
19291                 break;
19292         case TCP_RACK_GP_INCREASE_CA:
19293                 RACK_OPTS_INC(tcp_gp_inc_ca);
19294                 ca = optval;
19295                 if (ca < 100) {
19296                         /*
19297                          * We don't allow any reduction
19298                          * over the GP b/w.
19299                          */
19300                         error = EINVAL;
19301                         break;
19302                 }
19303                 rack->r_ctl.rack_per_of_gp_ca = ca;
19304                 rack_log_pacing_delay_calc(rack,
19305                                            rack->r_ctl.rack_per_of_gp_ss,
19306                                            rack->r_ctl.rack_per_of_gp_ca,
19307                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19308                                            __LINE__, NULL);
19309                 break;
19310         case TCP_RACK_GP_INCREASE_SS:
19311                 RACK_OPTS_INC(tcp_gp_inc_ss);
19312                 ss = optval;
19313                 if (ss < 100) {
19314                         /*
19315                          * We don't allow any reduction
19316                          * over the GP b/w.
19317                          */
19318                         error = EINVAL;
19319                         break;
19320                 }
19321                 rack->r_ctl.rack_per_of_gp_ss = ss;
19322                 rack_log_pacing_delay_calc(rack,
19323                                            rack->r_ctl.rack_per_of_gp_ss,
19324                                            rack->r_ctl.rack_per_of_gp_ca,
19325                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19326                                            __LINE__, NULL);
19327                 break;
19328         case TCP_RACK_RR_CONF:
19329                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19330                 if (optval && optval <= 3)
19331                         rack->r_rr_config = optval;
19332                 else
19333                         rack->r_rr_config = 0;
19334                 break;
19335         case TCP_HDWR_RATE_CAP:
19336                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
19337                 if (optval) {
19338                         if (rack->r_rack_hw_rate_caps == 0)
19339                                 rack->r_rack_hw_rate_caps = 1;
19340                         else
19341                                 error = EALREADY;
19342                 } else {
19343                         rack->r_rack_hw_rate_caps = 0;
19344                 }
19345                 break;
19346         case TCP_BBR_HDWR_PACE:
19347                 RACK_OPTS_INC(tcp_hdwr_pacing);
19348                 if (optval){
19349                         if (rack->rack_hdrw_pacing == 0) {
19350                                 rack->rack_hdw_pace_ena = 1;
19351                                 rack->rack_attempt_hdwr_pace = 0;
19352                         } else
19353                                 error = EALREADY;
19354                 } else {
19355                         rack->rack_hdw_pace_ena = 0;
19356 #ifdef RATELIMIT
19357                         if (rack->r_ctl.crte != NULL) {
19358                                 rack->rack_hdrw_pacing = 0;
19359                                 rack->rack_attempt_hdwr_pace = 0;
19360                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19361                                 rack->r_ctl.crte = NULL;
19362                         }
19363 #endif
19364                 }
19365                 break;
19366         /*  End Pacing related ones */
19367         case TCP_RACK_PRR_SENDALOT:
19368                 /* Allow PRR to send more than one seg */
19369                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
19370                 rack->r_ctl.rc_prr_sendalot = optval;
19371                 break;
19372         case TCP_RACK_MIN_TO:
19373                 /* Minimum time between rack t-o's in ms */
19374                 RACK_OPTS_INC(tcp_rack_min_to);
19375                 rack->r_ctl.rc_min_to = optval;
19376                 break;
19377         case TCP_RACK_EARLY_SEG:
19378                 /* If early recovery max segments */
19379                 RACK_OPTS_INC(tcp_rack_early_seg);
19380                 rack->r_ctl.rc_early_recovery_segs = optval;
19381                 break;
19382         case TCP_RACK_REORD_THRESH:
19383                 /* RACK reorder threshold (shift amount) */
19384                 RACK_OPTS_INC(tcp_rack_reord_thresh);
19385                 if ((optval > 0) && (optval < 31))
19386                         rack->r_ctl.rc_reorder_shift = optval;
19387                 else
19388                         error = EINVAL;
19389                 break;
19390         case TCP_RACK_REORD_FADE:
19391                 /* Does reordering fade after ms time */
19392                 RACK_OPTS_INC(tcp_rack_reord_fade);
19393                 rack->r_ctl.rc_reorder_fade = optval;
19394                 break;
19395         case TCP_RACK_TLP_THRESH:
19396                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19397                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
19398                 if (optval)
19399                         rack->r_ctl.rc_tlp_threshold = optval;
19400                 else
19401                         error = EINVAL;
19402                 break;
19403         case TCP_BBR_USE_RACK_RR:
19404                 RACK_OPTS_INC(tcp_rack_rr);
19405                 if (optval)
19406                         rack->use_rack_rr = 1;
19407                 else
19408                         rack->use_rack_rr = 0;
19409                 break;
19410         case TCP_FAST_RSM_HACK:
19411                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
19412                 if (optval)
19413                         rack->fast_rsm_hack = 1;
19414                 else
19415                         rack->fast_rsm_hack = 0;
19416                 break;
19417         case TCP_RACK_PKT_DELAY:
19418                 /* RACK added ms i.e. rack-rtt + reord + N */
19419                 RACK_OPTS_INC(tcp_rack_pkt_delay);
19420                 rack->r_ctl.rc_pkt_delay = optval;
19421                 break;
19422         case TCP_DELACK:
19423                 RACK_OPTS_INC(tcp_rack_delayed_ack);
19424                 if (optval == 0)
19425                         tp->t_delayed_ack = 0;
19426                 else
19427                         tp->t_delayed_ack = 1;
19428                 if (tp->t_flags & TF_DELACK) {
19429                         tp->t_flags &= ~TF_DELACK;
19430                         tp->t_flags |= TF_ACKNOW;
19431                         NET_EPOCH_ENTER(et);
19432                         rack_output(tp);
19433                         NET_EPOCH_EXIT(et);
19434                 }
19435                 break;
19436
19437         case TCP_BBR_RACK_RTT_USE:
19438                 RACK_OPTS_INC(tcp_rack_rtt_use);
19439                 if ((optval != USE_RTT_HIGH) &&
19440                     (optval != USE_RTT_LOW) &&
19441                     (optval != USE_RTT_AVG))
19442                         error = EINVAL;
19443                 else
19444                         rack->r_ctl.rc_rate_sample_method = optval;
19445                 break;
19446         case TCP_DATA_AFTER_CLOSE:
19447                 RACK_OPTS_INC(tcp_data_after_close);
19448                 if (optval)
19449                         rack->rc_allow_data_af_clo = 1;
19450                 else
19451                         rack->rc_allow_data_af_clo = 0;
19452                 break;
19453         default:
19454                 break;
19455         }
19456 #ifdef NETFLIX_STATS
19457         tcp_log_socket_option(tp, sopt_name, optval, error);
19458 #endif
19459         return (error);
19460 }
19461
19462
19463 static void
19464 rack_apply_deferred_options(struct tcp_rack *rack)
19465 {
19466         struct deferred_opt_list *dol, *sdol;
19467         uint32_t s_optval;
19468
19469         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
19470                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
19471                 /* Disadvantage of deferal is you loose the error return */
19472                 s_optval = (uint32_t)dol->optval;
19473                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
19474                 free(dol, M_TCPDO);
19475         }
19476 }
19477
19478 static void
19479 rack_hw_tls_change(struct tcpcb *tp, int chg)
19480 {
19481         /*
19482          * HW tls state has changed.. fix all
19483          * rsm's in flight.
19484          */
19485         struct tcp_rack *rack;
19486         struct rack_sendmap *rsm;
19487
19488         rack = (struct tcp_rack *)tp->t_fb_ptr;
19489         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
19490                 if (chg)
19491                         rsm->r_hw_tls = 1;
19492                 else
19493                         rsm->r_hw_tls = 0;
19494         }
19495         if (chg)
19496                 rack->r_ctl.fsb.hw_tls = 1;
19497         else
19498                 rack->r_ctl.fsb.hw_tls = 0;
19499 }
19500
19501 static int
19502 rack_pru_options(struct tcpcb *tp, int flags)
19503 {
19504         if (flags & PRUS_OOB)
19505                 return (EOPNOTSUPP);
19506         return (0);
19507 }
19508
19509 static struct tcp_function_block __tcp_rack = {
19510         .tfb_tcp_block_name = __XSTRING(STACKNAME),
19511         .tfb_tcp_output = rack_output,
19512         .tfb_do_queued_segments = ctf_do_queued_segments,
19513         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
19514         .tfb_tcp_do_segment = rack_do_segment,
19515         .tfb_tcp_ctloutput = rack_ctloutput,
19516         .tfb_tcp_fb_init = rack_init,
19517         .tfb_tcp_fb_fini = rack_fini,
19518         .tfb_tcp_timer_stop_all = rack_stopall,
19519         .tfb_tcp_timer_activate = rack_timer_activate,
19520         .tfb_tcp_timer_active = rack_timer_active,
19521         .tfb_tcp_timer_stop = rack_timer_stop,
19522         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
19523         .tfb_tcp_handoff_ok = rack_handoff_ok,
19524         .tfb_tcp_mtu_chg = rack_mtu_change,
19525         .tfb_pru_options = rack_pru_options,
19526         .tfb_hwtls_change = rack_hw_tls_change,
19527 };
19528
19529 /*
19530  * rack_ctloutput() must drop the inpcb lock before performing copyin on
19531  * socket option arguments.  When it re-acquires the lock after the copy, it
19532  * has to revalidate that the connection is still valid for the socket
19533  * option.
19534  */
19535 static int
19536 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
19537     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19538 {
19539         uint64_t loptval;
19540         int32_t error = 0, optval;
19541
19542         switch (sopt->sopt_name) {
19543         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
19544         /*  Pacing related ones */
19545         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
19546         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
19547         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
19548         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
19549         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
19550         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
19551         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
19552         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
19553         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
19554         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
19555         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
19556         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
19557         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
19558         case TCP_HDWR_RATE_CAP:                 /*  URL: hdwrcap boolean */
19559         case TCP_PACING_RATE_CAP:               /*  URL:cap-- used by side-channel */
19560         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
19561        /* End pacing related */
19562         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
19563         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
19564         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
19565         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
19566         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
19567         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
19568         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
19569         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
19570         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
19571         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
19572         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
19573         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
19574         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
19575         case TCP_NO_PRR:                        /*  URL:noprr */
19576         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
19577         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
19578         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
19579         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
19580         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
19581         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
19582         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
19583         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
19584         case TCP_RACK_PROFILE:                  /*  URL:profile */
19585         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
19586         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
19587         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
19588         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
19589         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
19590         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
19591         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
19592         case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
19593                 break;
19594         default:
19595                 /* Filter off all unknown options to the base stack */
19596                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19597                 break;
19598         }
19599         INP_WUNLOCK(inp);
19600         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
19601                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
19602                 /*
19603                  * We truncate it down to 32 bits for the socket-option trace this
19604                  * means rates > 34Gbps won't show right, but thats probably ok.
19605                  */
19606                 optval = (uint32_t)loptval;
19607         } else {
19608                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
19609                 /* Save it in 64 bit form too */
19610                 loptval = optval;
19611         }
19612         if (error)
19613                 return (error);
19614         INP_WLOCK(inp);
19615         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
19616                 INP_WUNLOCK(inp);
19617                 return (ECONNRESET);
19618         }
19619         if (tp->t_fb != &__tcp_rack) {
19620                 INP_WUNLOCK(inp);
19621                 return (ENOPROTOOPT);
19622         }
19623         if (rack->defer_options && (rack->gp_ready == 0) &&
19624             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
19625             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
19626             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
19627             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
19628                 /* Options are beind deferred */
19629                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
19630                         INP_WUNLOCK(inp);
19631                         return (0);
19632                 } else {
19633                         /* No memory to defer, fail */
19634                         INP_WUNLOCK(inp);
19635                         return (ENOMEM);
19636                 }
19637         }
19638         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
19639         INP_WUNLOCK(inp);
19640         return (error);
19641 }
19642
19643 static void
19644 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
19645 {
19646
19647         INP_WLOCK_ASSERT(tp->t_inpcb);
19648         bzero(ti, sizeof(*ti));
19649
19650         ti->tcpi_state = tp->t_state;
19651         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
19652                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
19653         if (tp->t_flags & TF_SACK_PERMIT)
19654                 ti->tcpi_options |= TCPI_OPT_SACK;
19655         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
19656                 ti->tcpi_options |= TCPI_OPT_WSCALE;
19657                 ti->tcpi_snd_wscale = tp->snd_scale;
19658                 ti->tcpi_rcv_wscale = tp->rcv_scale;
19659         }
19660         if (tp->t_flags2 & TF2_ECN_PERMIT)
19661                 ti->tcpi_options |= TCPI_OPT_ECN;
19662         if (tp->t_flags & TF_FASTOPEN)
19663                 ti->tcpi_options |= TCPI_OPT_TFO;
19664         /* still kept in ticks is t_rcvtime */
19665         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
19666         /* Since we hold everything in precise useconds this is easy */
19667         ti->tcpi_rtt = tp->t_srtt;
19668         ti->tcpi_rttvar = tp->t_rttvar;
19669         ti->tcpi_rto = tp->t_rxtcur;
19670         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
19671         ti->tcpi_snd_cwnd = tp->snd_cwnd;
19672         /*
19673          * FreeBSD-specific extension fields for tcp_info.
19674          */
19675         ti->tcpi_rcv_space = tp->rcv_wnd;
19676         ti->tcpi_rcv_nxt = tp->rcv_nxt;
19677         ti->tcpi_snd_wnd = tp->snd_wnd;
19678         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
19679         ti->tcpi_snd_nxt = tp->snd_nxt;
19680         ti->tcpi_snd_mss = tp->t_maxseg;
19681         ti->tcpi_rcv_mss = tp->t_maxseg;
19682         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
19683         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
19684         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
19685 #ifdef NETFLIX_STATS
19686         ti->tcpi_total_tlp = tp->t_sndtlppack;
19687         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
19688         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
19689 #endif
19690 #ifdef TCP_OFFLOAD
19691         if (tp->t_flags & TF_TOE) {
19692                 ti->tcpi_options |= TCPI_OPT_TOE;
19693                 tcp_offload_tcp_info(tp, ti);
19694         }
19695 #endif
19696 }
19697
19698 static int
19699 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
19700     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19701 {
19702         int32_t error, optval;
19703         uint64_t val, loptval;
19704         struct  tcp_info ti;
19705         /*
19706          * Because all our options are either boolean or an int, we can just
19707          * pull everything into optval and then unlock and copy. If we ever
19708          * add a option that is not a int, then this will have quite an
19709          * impact to this routine.
19710          */
19711         error = 0;
19712         switch (sopt->sopt_name) {
19713         case TCP_INFO:
19714                 /* First get the info filled */
19715                 rack_fill_info(tp, &ti);
19716                 /* Fix up the rtt related fields if needed */
19717                 INP_WUNLOCK(inp);
19718                 error = sooptcopyout(sopt, &ti, sizeof ti);
19719                 return (error);
19720         /*
19721          * Beta is the congestion control value for NewReno that influences how
19722          * much of a backoff happens when loss is detected. It is normally set
19723          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
19724          * when you exit recovery.
19725          */
19726         case TCP_RACK_PACING_BETA:
19727                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19728                         error = EINVAL;
19729                 else if (rack->rc_pacing_cc_set == 0)
19730                         optval = rack->r_ctl.rc_saved_beta.beta;
19731                 else {
19732                         /*
19733                          * Reach out into the CC data and report back what
19734                          * I have previously set. Yeah it looks hackish but
19735                          * we don't want to report the saved values.
19736                          */
19737                         if (tp->ccv->cc_data)
19738                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta;
19739                         else
19740                                 error = EINVAL;
19741                 }
19742                 break;
19743                 /*
19744                  * Beta_ecn is the congestion control value for NewReno that influences how
19745                  * much of a backoff happens when a ECN mark is detected. It is normally set
19746                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
19747                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
19748                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
19749                  */
19750
19751         case TCP_RACK_PACING_BETA_ECN:
19752                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19753                         error = EINVAL;
19754                 else if (rack->rc_pacing_cc_set == 0)
19755                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
19756                 else {
19757                         /*
19758                          * Reach out into the CC data and report back what
19759                          * I have previously set. Yeah it looks hackish but
19760                          * we don't want to report the saved values.
19761                          */
19762                         if (tp->ccv->cc_data)
19763                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
19764                         else
19765                                 error = EINVAL;
19766                 }
19767                 break;
19768         case TCP_FAST_RSM_HACK:
19769                 optval = rack->fast_rsm_hack;
19770                 break;
19771         case TCP_DEFER_OPTIONS:
19772                 optval = rack->defer_options;
19773                 break;
19774         case TCP_RACK_MEASURE_CNT:
19775                 optval = rack->r_ctl.req_measurements;
19776                 break;
19777         case TCP_REC_ABC_VAL:
19778                 optval = rack->r_use_labc_for_rec;
19779                 break;
19780         case TCP_RACK_ABC_VAL:
19781                 optval = rack->rc_labc;
19782                 break;
19783         case TCP_HDWR_UP_ONLY:
19784                 optval= rack->r_up_only;
19785                 break;
19786         case TCP_PACING_RATE_CAP:
19787                 loptval = rack->r_ctl.bw_rate_cap;
19788                 break;
19789         case TCP_RACK_PROFILE:
19790                 /* You cannot retrieve a profile, its write only */
19791                 error = EINVAL;
19792                 break;
19793         case TCP_USE_CMP_ACKS:
19794                 optval = rack->r_use_cmp_ack;
19795                 break;
19796         case TCP_RACK_PACE_TO_FILL:
19797                 optval = rack->rc_pace_to_cwnd;
19798                 if (optval && rack->r_fill_less_agg)
19799                         optval++;
19800                 break;
19801         case TCP_RACK_NO_PUSH_AT_MAX:
19802                 optval = rack->r_ctl.rc_no_push_at_mrtt;
19803                 break;
19804         case TCP_SHARED_CWND_ENABLE:
19805                 optval = rack->rack_enable_scwnd;
19806                 break;
19807         case TCP_RACK_NONRXT_CFG_RATE:
19808                 optval = rack->rack_rec_nonrxt_use_cr;
19809                 break;
19810         case TCP_NO_PRR:
19811                 if (rack->rack_no_prr  == 1)
19812                         optval = 1;
19813                 else if (rack->no_prr_addback == 1)
19814                         optval = 2;
19815                 else
19816                         optval = 0;
19817                 break;
19818         case TCP_RACK_DO_DETECTION:
19819                 optval = rack->do_detection;
19820                 break;
19821         case TCP_RACK_MBUF_QUEUE:
19822                 /* Now do we use the LRO mbuf-queue feature */
19823                 optval = rack->r_mbuf_queue;
19824                 break;
19825         case TCP_TIMELY_DYN_ADJ:
19826                 optval = rack->rc_gp_dyn_mul;
19827                 break;
19828         case TCP_BBR_IWINTSO:
19829                 optval = rack->rc_init_win;
19830                 break;
19831         case TCP_RACK_TLP_REDUCE:
19832                 /* RACK TLP cwnd reduction (bool) */
19833                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
19834                 break;
19835         case TCP_BBR_RACK_INIT_RATE:
19836                 val = rack->r_ctl.init_rate;
19837                 /* convert to kbits per sec */
19838                 val *= 8;
19839                 val /= 1000;
19840                 optval = (uint32_t)val;
19841                 break;
19842         case TCP_RACK_FORCE_MSEG:
19843                 optval = rack->rc_force_max_seg;
19844                 break;
19845         case TCP_RACK_PACE_MAX_SEG:
19846                 /* Max segments in a pace */
19847                 optval = rack->rc_user_set_max_segs;
19848                 break;
19849         case TCP_RACK_PACE_ALWAYS:
19850                 /* Use the always pace method */
19851                 optval = rack->rc_always_pace;
19852                 break;
19853         case TCP_RACK_PRR_SENDALOT:
19854                 /* Allow PRR to send more than one seg */
19855                 optval = rack->r_ctl.rc_prr_sendalot;
19856                 break;
19857         case TCP_RACK_MIN_TO:
19858                 /* Minimum time between rack t-o's in ms */
19859                 optval = rack->r_ctl.rc_min_to;
19860                 break;
19861         case TCP_RACK_EARLY_SEG:
19862                 /* If early recovery max segments */
19863                 optval = rack->r_ctl.rc_early_recovery_segs;
19864                 break;
19865         case TCP_RACK_REORD_THRESH:
19866                 /* RACK reorder threshold (shift amount) */
19867                 optval = rack->r_ctl.rc_reorder_shift;
19868                 break;
19869         case TCP_RACK_REORD_FADE:
19870                 /* Does reordering fade after ms time */
19871                 optval = rack->r_ctl.rc_reorder_fade;
19872                 break;
19873         case TCP_BBR_USE_RACK_RR:
19874                 /* Do we use the rack cheat for rxt */
19875                 optval = rack->use_rack_rr;
19876                 break;
19877         case TCP_RACK_RR_CONF:
19878                 optval = rack->r_rr_config;
19879                 break;
19880         case TCP_HDWR_RATE_CAP:
19881                 optval = rack->r_rack_hw_rate_caps;
19882                 break;
19883         case TCP_BBR_HDWR_PACE:
19884                 optval = rack->rack_hdw_pace_ena;
19885                 break;
19886         case TCP_RACK_TLP_THRESH:
19887                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19888                 optval = rack->r_ctl.rc_tlp_threshold;
19889                 break;
19890         case TCP_RACK_PKT_DELAY:
19891                 /* RACK added ms i.e. rack-rtt + reord + N */
19892                 optval = rack->r_ctl.rc_pkt_delay;
19893                 break;
19894         case TCP_RACK_TLP_USE:
19895                 optval = rack->rack_tlp_threshold_use;
19896                 break;
19897         case TCP_RACK_PACE_RATE_CA:
19898                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
19899                 break;
19900         case TCP_RACK_PACE_RATE_SS:
19901                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
19902                 break;
19903         case TCP_RACK_PACE_RATE_REC:
19904                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
19905                 break;
19906         case TCP_RACK_GP_INCREASE_SS:
19907                 optval = rack->r_ctl.rack_per_of_gp_ca;
19908                 break;
19909         case TCP_RACK_GP_INCREASE_CA:
19910                 optval = rack->r_ctl.rack_per_of_gp_ss;
19911                 break;
19912         case TCP_BBR_RACK_RTT_USE:
19913                 optval = rack->r_ctl.rc_rate_sample_method;
19914                 break;
19915         case TCP_DELACK:
19916                 optval = tp->t_delayed_ack;
19917                 break;
19918         case TCP_DATA_AFTER_CLOSE:
19919                 optval = rack->rc_allow_data_af_clo;
19920                 break;
19921         case TCP_SHARED_CWND_TIME_LIMIT:
19922                 optval = rack->r_limit_scw;
19923                 break;
19924         case TCP_RACK_TIMER_SLOP:
19925                 optval = rack->r_ctl.timer_slop;
19926                 break;
19927         default:
19928                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19929                 break;
19930         }
19931         INP_WUNLOCK(inp);
19932         if (error == 0) {
19933                 if (TCP_PACING_RATE_CAP)
19934                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
19935                 else
19936                         error = sooptcopyout(sopt, &optval, sizeof optval);
19937         }
19938         return (error);
19939 }
19940
19941 static int
19942 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
19943 {
19944         int32_t error = EINVAL;
19945         struct tcp_rack *rack;
19946
19947         rack = (struct tcp_rack *)tp->t_fb_ptr;
19948         if (rack == NULL) {
19949                 /* Huh? */
19950                 goto out;
19951         }
19952         if (sopt->sopt_dir == SOPT_SET) {
19953                 return (rack_set_sockopt(so, sopt, inp, tp, rack));
19954         } else if (sopt->sopt_dir == SOPT_GET) {
19955                 return (rack_get_sockopt(so, sopt, inp, tp, rack));
19956         }
19957 out:
19958         INP_WUNLOCK(inp);
19959         return (error);
19960 }
19961
19962 static const char *rack_stack_names[] = {
19963         __XSTRING(STACKNAME),
19964 #ifdef STACKALIAS
19965         __XSTRING(STACKALIAS),
19966 #endif
19967 };
19968
19969 static int
19970 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
19971 {
19972         memset(mem, 0, size);
19973         return (0);
19974 }
19975
19976 static void
19977 rack_dtor(void *mem, int32_t size, void *arg)
19978 {
19979
19980 }
19981
19982 static bool rack_mod_inited = false;
19983
19984 static int
19985 tcp_addrack(module_t mod, int32_t type, void *data)
19986 {
19987         int32_t err = 0;
19988         int num_stacks;
19989
19990         switch (type) {
19991         case MOD_LOAD:
19992                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
19993                     sizeof(struct rack_sendmap),
19994                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
19995
19996                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
19997                     sizeof(struct tcp_rack),
19998                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
19999
20000                 sysctl_ctx_init(&rack_sysctl_ctx);
20001                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20002                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20003                     OID_AUTO,
20004 #ifdef STACKALIAS
20005                     __XSTRING(STACKALIAS),
20006 #else
20007                     __XSTRING(STACKNAME),
20008 #endif
20009                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20010                     "");
20011                 if (rack_sysctl_root == NULL) {
20012                         printf("Failed to add sysctl node\n");
20013                         err = EFAULT;
20014                         goto free_uma;
20015                 }
20016                 rack_init_sysctls();
20017                 num_stacks = nitems(rack_stack_names);
20018                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20019                     rack_stack_names, &num_stacks);
20020                 if (err) {
20021                         printf("Failed to register %s stack name for "
20022                             "%s module\n", rack_stack_names[num_stacks],
20023                             __XSTRING(MODNAME));
20024                         sysctl_ctx_free(&rack_sysctl_ctx);
20025 free_uma:
20026                         uma_zdestroy(rack_zone);
20027                         uma_zdestroy(rack_pcb_zone);
20028                         rack_counter_destroy();
20029                         printf("Failed to register rack module -- err:%d\n", err);
20030                         return (err);
20031                 }
20032                 tcp_lro_reg_mbufq();
20033                 rack_mod_inited = true;
20034                 break;
20035         case MOD_QUIESCE:
20036                 err = deregister_tcp_functions(&__tcp_rack, true, false);
20037                 break;
20038         case MOD_UNLOAD:
20039                 err = deregister_tcp_functions(&__tcp_rack, false, true);
20040                 if (err == EBUSY)
20041                         break;
20042                 if (rack_mod_inited) {
20043                         uma_zdestroy(rack_zone);
20044                         uma_zdestroy(rack_pcb_zone);
20045                         sysctl_ctx_free(&rack_sysctl_ctx);
20046                         rack_counter_destroy();
20047                         rack_mod_inited = false;
20048                 }
20049                 tcp_lro_dereg_mbufq();
20050                 err = 0;
20051                 break;
20052         default:
20053                 return (EOPNOTSUPP);
20054         }
20055         return (err);
20056 }
20057
20058 static moduledata_t tcp_rack = {
20059         .name = __XSTRING(MODNAME),
20060         .evhand = tcp_addrack,
20061         .priv = 0
20062 };
20063
20064 MODULE_VERSION(MODNAME, 1);
20065 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20066 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);