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
 462 static void
 463 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
 464    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
 465    struct rack_sendmap *rsm, uint8_t quality);
 466 static struct rack_sendmap *
 467 rack_find_high_nonack(struct tcp_rack *rack,
 468     struct rack_sendmap *rsm);
 469 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
 470 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
 471 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
 472 static int
 473 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
 474     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
 475 static void
 476 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
 477                             tcp_seq th_ack, int line, uint8_t quality);
 478 static uint32_t
 479 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
 480 static int32_t rack_handoff_ok(struct tcpcb *tp);
 481 static int32_t rack_init(struct tcpcb *tp);
 482 static void rack_init_sysctls(void);
 483 static void
 484 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
 485     struct tcphdr *th, int entered_rec, int dup_ack_struck);
 486 static void
 487 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
 488     uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t ts,
 489     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
 490
 491 static void
 492 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
 493     struct rack_sendmap *rsm);
 494 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
 495 static int32_t rack_output(struct tcpcb *tp);
 496
 497 static uint32_t
 498 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
 499     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
 500     uint32_t cts, int *moved_two);
 501 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
 502 static void rack_remxt_tmr(struct tcpcb *tp);
 503 static int
 504 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
 505     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
 506 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
 507 static int32_t rack_stopall(struct tcpcb *tp);
 508 static void
 509 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
 510     uint32_t delta);
 511 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
 512 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
 513 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
 514 static uint32_t
 515 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
 516     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
 517 static void
 518 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
 519     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
 520 static int
 521 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
 522     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
 523 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
 524 static int
 525 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
 526     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 527     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 528 static int
 529 rack_do_closing(struct mbuf *m, struct tcphdr *th,
 530     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 531     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 532 static int
 533 rack_do_established(struct mbuf *m, struct tcphdr *th,
 534     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 535     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 536 static int
 537 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
 538     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 539     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
 540 static int
 541 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
 542     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 543     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 544 static int
 545 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
 546     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 547     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 548 static int
 549 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
 550     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 551     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 552 static int
 553 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
 554     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 555     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 556 static int
 557 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
 558     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 559     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 560 struct rack_sendmap *
 561 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
 562     uint32_t tsused);
 563 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
 564     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
 565 static void
 566      tcp_rack_partialack(struct tcpcb *tp);
 567 static int
 568 rack_set_profile(struct tcp_rack *rack, int prof);
 569 static void
 570 rack_apply_deferred_options(struct tcp_rack *rack);
 571
 572 int32_t rack_clear_counter=0;
 573
 574 static void
 575 rack_set_cc_pacing(struct tcp_rack *rack)
 576 {
 577         struct sockopt sopt;
 578         struct cc_newreno_opts opt;
 579         struct newreno old, *ptr;
 580         struct tcpcb *tp;
 581         int error;
 582
 583         if (rack->rc_pacing_cc_set)
 584                 return;
 585
 586         tp = rack->rc_tp;
 587         if (tp->cc_algo == NULL) {
 588                 /* Tcb is leaving */
 589                 printf("No cc algorithm?\n");
 590                 return;
 591         }
 592         rack->rc_pacing_cc_set = 1;
 593         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 594                 /* Not new-reno we can't play games with beta! */
 595                 goto out;
 596         }
 597         ptr = ((struct newreno *)tp->ccv->cc_data);
 598         if (CC_ALGO(tp)->ctl_output == NULL)  {
 599                 /* Huh, why does new_reno no longer have a set function? */
 600                 printf("no ctl_output for algo:%s\n", tp->cc_algo->name);
 601                 goto out;
 602         }
 603         if (ptr == NULL) {
 604                 /* Just the default values */
 605                 old.beta = V_newreno_beta_ecn;
 606                 old.beta_ecn = V_newreno_beta_ecn;
 607                 old.newreno_flags = 0;
 608         } else {
 609                 old.beta = ptr->beta;
 610                 old.beta_ecn = ptr->beta_ecn;
 611                 old.newreno_flags = ptr->newreno_flags;
 612         }
 613         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 614         sopt.sopt_dir = SOPT_SET;
 615         opt.name = CC_NEWRENO_BETA;
 616         opt.val = rack->r_ctl.rc_saved_beta.beta;
 617         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 618         if (error)  {
 619                 printf("Error returned by ctl_output %d\n", error);
 620                 goto out;
 621         }
 622         /*
 623          * Hack alert we need to set in our newreno_flags
 624          * so that Abe behavior is also applied.
 625          */
 626         ((struct newreno *)tp->ccv->cc_data)->newreno_flags = CC_NEWRENO_BETA_ECN;
 627         opt.name = CC_NEWRENO_BETA_ECN;
 628         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 629         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 630         if (error) {
 631                 printf("Error returned by ctl_output %d\n", error);
 632                 goto out;
 633         }
 634         /* Save off the original values for restoral */
 635         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 636 out:
 637         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 638                 union tcp_log_stackspecific log;
 639                 struct timeval tv;
 640
 641                 ptr = ((struct newreno *)tp->ccv->cc_data);
 642                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 643                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 644                 if (ptr) {
 645                         log.u_bbr.flex1 = ptr->beta;
 646                         log.u_bbr.flex2 = ptr->beta_ecn;
 647                         log.u_bbr.flex3 = ptr->newreno_flags;
 648                 }
 649                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 650                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 651                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 652                 log.u_bbr.flex7 = rack->gp_ready;
 653                 log.u_bbr.flex7 <<= 1;
 654                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 655                 log.u_bbr.flex7 <<= 1;
 656                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 657                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 658                 log.u_bbr.flex8 = 3;
 659                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 660                                0, &log, false, NULL, NULL, 0, &tv);
 661         }
 662 }
 663
 664 static void
 665 rack_undo_cc_pacing(struct tcp_rack *rack)
 666 {
 667         struct newreno old, *ptr;
 668         struct tcpcb *tp;
 669
 670         if (rack->rc_pacing_cc_set == 0)
 671                 return;
 672         tp = rack->rc_tp;
 673         rack->rc_pacing_cc_set = 0;
 674         if (tp->cc_algo == NULL)
 675                 /* Tcb is leaving */
 676                 return;
 677         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 678                 /* Not new-reno nothing to do! */
 679                 return;
 680         }
 681         ptr = ((struct newreno *)tp->ccv->cc_data);
 682         if (ptr == NULL) {
 683                 /*
 684                  * This happens at rack_fini() if the
 685                  * cc module gets freed on us. In that
 686                  * case we loose our "new" settings but
 687                  * thats ok, since the tcb is going away anyway.
 688                  */
 689                 return;
 690         }
 691         /* Grab out our set values */
 692         memcpy(&old, ptr, sizeof(struct newreno));
 693         /* Copy back in the original values */
 694         memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
 695         /* Now save back the values we had set in (for when pacing is restored) */
 696         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 697         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 698                 union tcp_log_stackspecific log;
 699                 struct timeval tv;
 700
 701                 ptr = ((struct newreno *)tp->ccv->cc_data);
 702                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 703                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 704                 log.u_bbr.flex1 = ptr->beta;
 705                 log.u_bbr.flex2 = ptr->beta_ecn;
 706                 log.u_bbr.flex3 = ptr->newreno_flags;
 707                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 708                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 709                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 710                 log.u_bbr.flex7 = rack->gp_ready;
 711                 log.u_bbr.flex7 <<= 1;
 712                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 713                 log.u_bbr.flex7 <<= 1;
 714                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 715                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 716                 log.u_bbr.flex8 = 4;
 717                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
 718                                0, &log, false, NULL, NULL, 0, &tv);
 719         }
 720 }
 721
 722 #ifdef NETFLIX_PEAKRATE
 723 static inline void
 724 rack_update_peakrate_thr(struct tcpcb *tp)
 725 {
 726         /* Keep in mind that t_maxpeakrate is in B/s. */
 727         uint64_t peak;
 728         peak = uqmax((tp->t_maxseg * 2),
 729                      (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
 730         tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
 731 }
 732 #endif
 733
 734 static int
 735 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 736 {
 737         uint32_t stat;
 738         int32_t error;
 739         int i;
 740
 741         error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
 742         if (error || req->newptr == NULL)
 743                 return error;
 744
 745         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
 746         if (error)
 747                 return (error);
 748         if (stat == 1) {
 749 #ifdef INVARIANTS
 750                 printf("Clearing RACK counters\n");
 751 #endif
 752                 counter_u64_zero(rack_badfr);
 753                 counter_u64_zero(rack_badfr_bytes);
 754                 counter_u64_zero(rack_rtm_prr_retran);
 755                 counter_u64_zero(rack_rtm_prr_newdata);
 756                 counter_u64_zero(rack_timestamp_mismatch);
 757                 counter_u64_zero(rack_reorder_seen);
 758                 counter_u64_zero(rack_tlp_tot);
 759                 counter_u64_zero(rack_tlp_newdata);
 760                 counter_u64_zero(rack_tlp_retran);
 761                 counter_u64_zero(rack_tlp_retran_bytes);
 762                 counter_u64_zero(rack_tlp_retran_fail);
 763                 counter_u64_zero(rack_to_tot);
 764                 counter_u64_zero(rack_to_arm_rack);
 765                 counter_u64_zero(rack_to_arm_tlp);
 766                 counter_u64_zero(rack_paced_segments);
 767                 counter_u64_zero(rack_calc_zero);
 768                 counter_u64_zero(rack_calc_nonzero);
 769                 counter_u64_zero(rack_unpaced_segments);
 770                 counter_u64_zero(rack_saw_enobuf);
 771                 counter_u64_zero(rack_saw_enobuf_hw);
 772                 counter_u64_zero(rack_saw_enetunreach);
 773                 counter_u64_zero(rack_per_timer_hole);
 774                 counter_u64_zero(rack_large_ackcmp);
 775                 counter_u64_zero(rack_small_ackcmp);
 776 #ifdef INVARIANTS
 777                 counter_u64_zero(rack_adjust_map_bw);
 778 #endif
 779                 counter_u64_zero(rack_to_alloc_hard);
 780                 counter_u64_zero(rack_to_alloc_emerg);
 781                 counter_u64_zero(rack_sack_proc_all);
 782                 counter_u64_zero(rack_fto_send);
 783                 counter_u64_zero(rack_fto_rsm_send);
 784                 counter_u64_zero(rack_extended_rfo);
 785                 counter_u64_zero(rack_hw_pace_init_fail);
 786                 counter_u64_zero(rack_hw_pace_lost);
 787                 counter_u64_zero(rack_sbsndptr_wrong);
 788                 counter_u64_zero(rack_sbsndptr_right);
 789                 counter_u64_zero(rack_non_fto_send);
 790                 counter_u64_zero(rack_nfto_resend);
 791                 counter_u64_zero(rack_sack_proc_short);
 792                 counter_u64_zero(rack_sack_proc_restart);
 793                 counter_u64_zero(rack_to_alloc);
 794                 counter_u64_zero(rack_to_alloc_limited);
 795                 counter_u64_zero(rack_alloc_limited_conns);
 796                 counter_u64_zero(rack_split_limited);
 797                 for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
 798                         counter_u64_zero(rack_proc_comp_ack[i]);
 799                 }
 800                 counter_u64_zero(rack_multi_single_eq);
 801                 counter_u64_zero(rack_proc_non_comp_ack);
 802                 counter_u64_zero(rack_find_high);
 803                 counter_u64_zero(rack_sack_attacks_detected);
 804                 counter_u64_zero(rack_sack_attacks_reversed);
 805                 counter_u64_zero(rack_sack_used_next_merge);
 806                 counter_u64_zero(rack_sack_used_prev_merge);
 807                 counter_u64_zero(rack_sack_splits);
 808                 counter_u64_zero(rack_sack_skipped_acked);
 809                 counter_u64_zero(rack_ack_total);
 810                 counter_u64_zero(rack_express_sack);
 811                 counter_u64_zero(rack_sack_total);
 812                 counter_u64_zero(rack_move_none);
 813                 counter_u64_zero(rack_move_some);
 814                 counter_u64_zero(rack_used_tlpmethod);
 815                 counter_u64_zero(rack_used_tlpmethod2);
 816                 counter_u64_zero(rack_enter_tlp_calc);
 817                 counter_u64_zero(rack_progress_drops);
 818                 counter_u64_zero(rack_tlp_does_nada);
 819                 counter_u64_zero(rack_try_scwnd);
 820                 counter_u64_zero(rack_collapsed_win);
 821         }
 822         rack_clear_counter = 0;
 823         return (0);
 824 }
 825
 826 static void
 827 rack_init_sysctls(void)
 828 {
 829         int i;
 830         struct sysctl_oid *rack_counters;
 831         struct sysctl_oid *rack_attack;
 832         struct sysctl_oid *rack_pacing;
 833         struct sysctl_oid *rack_timely;
 834         struct sysctl_oid *rack_timers;
 835         struct sysctl_oid *rack_tlp;
 836         struct sysctl_oid *rack_misc;
 837         struct sysctl_oid *rack_measure;
 838         struct sysctl_oid *rack_probertt;
 839         struct sysctl_oid *rack_hw_pacing;
 840
 841         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 842             SYSCTL_CHILDREN(rack_sysctl_root),
 843             OID_AUTO,
 844             "sack_attack",
 845             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 846             "Rack Sack Attack Counters and Controls");
 847         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 848             SYSCTL_CHILDREN(rack_sysctl_root),
 849             OID_AUTO,
 850             "stats",
 851             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 852             "Rack Counters");
 853         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 854             SYSCTL_CHILDREN(rack_sysctl_root),
 855             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 856             &rack_rate_sample_method , USE_RTT_LOW,
 857             "What method should we use for rate sampling 0=high, 1=low ");
 858         /* Probe rtt related controls */
 859         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 860             SYSCTL_CHILDREN(rack_sysctl_root),
 861             OID_AUTO,
 862             "probertt",
 863             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 864             "ProbeRTT related Controls");
 865         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 866             SYSCTL_CHILDREN(rack_probertt),
 867             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 868             &rack_atexit_prtt_hbp, 130,
 869             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 870         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 871             SYSCTL_CHILDREN(rack_probertt),
 872             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 873             &rack_atexit_prtt, 130,
 874             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 875         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 876             SYSCTL_CHILDREN(rack_probertt),
 877             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 878             &rack_per_of_gp_probertt, 60,
 879             "What percentage of goodput do we pace at in probertt");
 880         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 881             SYSCTL_CHILDREN(rack_probertt),
 882             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 883             &rack_per_of_gp_probertt_reduce, 10,
 884             "What percentage of goodput do we reduce every gp_srtt");
 885         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 886             SYSCTL_CHILDREN(rack_probertt),
 887             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 888             &rack_per_of_gp_lowthresh, 40,
 889             "What percentage of goodput do we allow the multiplier to fall to");
 890         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 891             SYSCTL_CHILDREN(rack_probertt),
 892             OID_AUTO, "time_between", CTLFLAG_RW,
 893             & rack_time_between_probertt, 96000000,
 894             "How many useconds between the lowest rtt falling must past before we enter probertt");
 895         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 896             SYSCTL_CHILDREN(rack_probertt),
 897             OID_AUTO, "safety", CTLFLAG_RW,
 898             &rack_probe_rtt_safety_val, 2000000,
 899             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 900         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 901             SYSCTL_CHILDREN(rack_probertt),
 902             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 903             &rack_probe_rtt_sets_cwnd, 0,
 904             "Do we set the cwnd too (if always_lower is on)");
 905         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 906             SYSCTL_CHILDREN(rack_probertt),
 907             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 908             &rack_max_drain_wait, 2,
 909             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 910         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 911             SYSCTL_CHILDREN(rack_probertt),
 912             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 913             &rack_must_drain, 1,
 914             "We must drain this many gp_srtt's waiting for flight to reach goal");
 915         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 916             SYSCTL_CHILDREN(rack_probertt),
 917             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 918             &rack_probertt_use_min_rtt_entry, 1,
 919             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 920         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 921             SYSCTL_CHILDREN(rack_probertt),
 922             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 923             &rack_probertt_use_min_rtt_exit, 0,
 924             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 925         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 926             SYSCTL_CHILDREN(rack_probertt),
 927             OID_AUTO, "length_div", CTLFLAG_RW,
 928             &rack_probertt_gpsrtt_cnt_div, 0,
 929             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 930         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 931             SYSCTL_CHILDREN(rack_probertt),
 932             OID_AUTO, "length_mul", CTLFLAG_RW,
 933             &rack_probertt_gpsrtt_cnt_mul, 0,
 934             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 935         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 936             SYSCTL_CHILDREN(rack_probertt),
 937             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 938             &rack_min_probertt_hold, 200000,
 939             "What is the minimum time we hold probertt at target");
 940         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 941             SYSCTL_CHILDREN(rack_probertt),
 942             OID_AUTO, "filter_life", CTLFLAG_RW,
 943             &rack_probertt_filter_life, 10000000,
 944             "What is the time for the filters life in useconds");
 945         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 946             SYSCTL_CHILDREN(rack_probertt),
 947             OID_AUTO, "lower_within", CTLFLAG_RW,
 948             &rack_probertt_lower_within, 10,
 949             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 950         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 951             SYSCTL_CHILDREN(rack_probertt),
 952             OID_AUTO, "must_move", CTLFLAG_RW,
 953             &rack_min_rtt_movement, 250,
 954             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 955         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 956             SYSCTL_CHILDREN(rack_probertt),
 957             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 958             &rack_probertt_clear_is, 1,
 959             "Do we clear I/S counts on exiting probe-rtt");
 960         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 961             SYSCTL_CHILDREN(rack_probertt),
 962             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 963             &rack_max_drain_hbp, 1,
 964             "How many extra drain gpsrtt's do we get in highly buffered paths");
 965         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 966             SYSCTL_CHILDREN(rack_probertt),
 967             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 968             &rack_hbp_thresh, 3,
 969             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 970         /* Pacing related sysctls */
 971         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 972             SYSCTL_CHILDREN(rack_sysctl_root),
 973             OID_AUTO,
 974             "pacing",
 975             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 976             "Pacing related Controls");
 977         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 978             SYSCTL_CHILDREN(rack_pacing),
 979             OID_AUTO, "max_pace_over", CTLFLAG_RW,
 980             &rack_max_per_above, 30,
 981             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
 982         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 983             SYSCTL_CHILDREN(rack_pacing),
 984             OID_AUTO, "pace_to_one", CTLFLAG_RW,
 985             &rack_pace_one_seg, 0,
 986             "Do we allow low b/w pacing of 1MSS instead of two");
 987         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 988             SYSCTL_CHILDREN(rack_pacing),
 989             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
 990             &rack_limit_time_with_srtt, 0,
 991             "Do we limit pacing time based on srtt");
 992         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 993             SYSCTL_CHILDREN(rack_pacing),
 994             OID_AUTO, "init_win", CTLFLAG_RW,
 995             &rack_default_init_window, 0,
 996             "Do we have a rack initial window 0 = system default");
 997         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 998             SYSCTL_CHILDREN(rack_pacing),
 999             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1000             &rack_per_of_gp_ss, 250,
1001             "If non zero, what percentage of goodput to pace at in slow start");
1002         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1003             SYSCTL_CHILDREN(rack_pacing),
1004             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1005             &rack_per_of_gp_ca, 150,
1006             "If non zero, what percentage of goodput to pace at in congestion avoidance");
1007         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1008             SYSCTL_CHILDREN(rack_pacing),
1009             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1010             &rack_per_of_gp_rec, 200,
1011             "If non zero, what percentage of goodput to pace at in recovery");
1012         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1013             SYSCTL_CHILDREN(rack_pacing),
1014             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1015             &rack_hptsi_segments, 40,
1016             "What size is the max for TSO segments in pacing and burst mitigation");
1017         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1018             SYSCTL_CHILDREN(rack_pacing),
1019             OID_AUTO, "burst_reduces", CTLFLAG_RW,
1020             &rack_slot_reduction, 4,
1021             "When doing only burst mitigation what is the reduce divisor");
1022         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1023             SYSCTL_CHILDREN(rack_sysctl_root),
1024             OID_AUTO, "use_pacing", CTLFLAG_RW,
1025             &rack_pace_every_seg, 0,
1026             "If set we use pacing, if clear we use only the original burst mitigation");
1027         SYSCTL_ADD_U64(&rack_sysctl_ctx,
1028             SYSCTL_CHILDREN(rack_pacing),
1029             OID_AUTO, "rate_cap", CTLFLAG_RW,
1030             &rack_bw_rate_cap, 0,
1031             "If set we apply this value to the absolute rate cap used by pacing");
1032         SYSCTL_ADD_U8(&rack_sysctl_ctx,
1033             SYSCTL_CHILDREN(rack_sysctl_root),
1034             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1035             &rack_req_measurements, 1,
1036             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1037         /* Hardware pacing */
1038         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1039             SYSCTL_CHILDREN(rack_sysctl_root),
1040             OID_AUTO,
1041             "hdwr_pacing",
1042             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1043             "Pacing related Controls");
1044         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1045             SYSCTL_CHILDREN(rack_hw_pacing),
1046             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1047             &rack_hw_rwnd_factor, 2,
1048             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1049         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1050             SYSCTL_CHILDREN(rack_hw_pacing),
1051             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1052             &rack_enobuf_hw_boost_mult, 2,
1053             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1054         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1055             SYSCTL_CHILDREN(rack_hw_pacing),
1056             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1057             &rack_enobuf_hw_max, 2,
1058             "What is the max boost the pacing time if we see a ENOBUFS?");
1059         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1060             SYSCTL_CHILDREN(rack_hw_pacing),
1061             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1062             &rack_enobuf_hw_min, 2,
1063             "What is the min boost the pacing time if we see a ENOBUFS?");
1064         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1065             SYSCTL_CHILDREN(rack_hw_pacing),
1066             OID_AUTO, "enable", CTLFLAG_RW,
1067             &rack_enable_hw_pacing, 0,
1068             "Should RACK attempt to use hw pacing?");
1069         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1070             SYSCTL_CHILDREN(rack_hw_pacing),
1071             OID_AUTO, "rate_cap", CTLFLAG_RW,
1072             &rack_hw_rate_caps, 1,
1073             "Does the highest hardware pacing rate cap the rate we will send at??");
1074         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1075             SYSCTL_CHILDREN(rack_hw_pacing),
1076             OID_AUTO, "rate_min", CTLFLAG_RW,
1077             &rack_hw_rate_min, 0,
1078             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1079         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1080             SYSCTL_CHILDREN(rack_hw_pacing),
1081             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1082             &rack_hw_rate_to_low, 0,
1083             "If we fall below this rate, dis-engage hw pacing?");
1084         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1085             SYSCTL_CHILDREN(rack_hw_pacing),
1086             OID_AUTO, "up_only", CTLFLAG_RW,
1087             &rack_hw_up_only, 1,
1088             "Do we allow hw pacing to lower the rate selected?");
1089         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1090             SYSCTL_CHILDREN(rack_hw_pacing),
1091             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1092             &rack_hw_pace_extra_slots, 2,
1093             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1094         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1095             SYSCTL_CHILDREN(rack_sysctl_root),
1096             OID_AUTO,
1097             "timely",
1098             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1099             "Rack Timely RTT Controls");
1100         /* Timely based GP dynmics */
1101         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1102             SYSCTL_CHILDREN(rack_timely),
1103             OID_AUTO, "upper", CTLFLAG_RW,
1104             &rack_gp_per_bw_mul_up, 2,
1105             "Rack timely upper range for equal b/w (in percentage)");
1106         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1107             SYSCTL_CHILDREN(rack_timely),
1108             OID_AUTO, "lower", CTLFLAG_RW,
1109             &rack_gp_per_bw_mul_down, 4,
1110             "Rack timely lower range for equal b/w (in percentage)");
1111         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1112             SYSCTL_CHILDREN(rack_timely),
1113             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1114             &rack_gp_rtt_maxmul, 3,
1115             "Rack timely multipler of lowest rtt for rtt_max");
1116         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1117             SYSCTL_CHILDREN(rack_timely),
1118             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1119             &rack_gp_rtt_mindiv, 4,
1120             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1121         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1122             SYSCTL_CHILDREN(rack_timely),
1123             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1124             &rack_gp_rtt_minmul, 1,
1125             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1126         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1127             SYSCTL_CHILDREN(rack_timely),
1128             OID_AUTO, "decrease", CTLFLAG_RW,
1129             &rack_gp_decrease_per, 20,
1130             "Rack timely decrease percentage of our GP multiplication factor");
1131         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1132             SYSCTL_CHILDREN(rack_timely),
1133             OID_AUTO, "increase", CTLFLAG_RW,
1134             &rack_gp_increase_per, 2,
1135             "Rack timely increase perentage of our GP multiplication factor");
1136         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1137             SYSCTL_CHILDREN(rack_timely),
1138             OID_AUTO, "lowerbound", CTLFLAG_RW,
1139             &rack_per_lower_bound, 50,
1140             "Rack timely lowest percentage we allow GP multiplier to fall to");
1141         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1142             SYSCTL_CHILDREN(rack_timely),
1143             OID_AUTO, "upperboundss", CTLFLAG_RW,
1144             &rack_per_upper_bound_ss, 0,
1145             "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1146         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1147             SYSCTL_CHILDREN(rack_timely),
1148             OID_AUTO, "upperboundca", CTLFLAG_RW,
1149             &rack_per_upper_bound_ca, 0,
1150             "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1151         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1152             SYSCTL_CHILDREN(rack_timely),
1153             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1154             &rack_do_dyn_mul, 0,
1155             "Rack timely do we enable dynmaic timely goodput by default");
1156         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1157             SYSCTL_CHILDREN(rack_timely),
1158             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1159             &rack_gp_no_rec_chg, 1,
1160             "Rack timely do we prohibit the recovery multiplier from being lowered");
1161         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1162             SYSCTL_CHILDREN(rack_timely),
1163             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1164             &rack_timely_dec_clear, 6,
1165             "Rack timely what threshold do we count to before another boost during b/w decent");
1166         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1167             SYSCTL_CHILDREN(rack_timely),
1168             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1169             &rack_timely_max_push_rise, 3,
1170             "Rack timely how many times do we push up with b/w increase");
1171         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1172             SYSCTL_CHILDREN(rack_timely),
1173             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1174             &rack_timely_max_push_drop, 3,
1175             "Rack timely how many times do we push back on b/w decent");
1176         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1177             SYSCTL_CHILDREN(rack_timely),
1178             OID_AUTO, "min_segs", CTLFLAG_RW,
1179             &rack_timely_min_segs, 4,
1180             "Rack timely when setting the cwnd what is the min num segments");
1181         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1182             SYSCTL_CHILDREN(rack_timely),
1183             OID_AUTO, "noback_max", CTLFLAG_RW,
1184             &rack_use_max_for_nobackoff, 0,
1185             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1186         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1187             SYSCTL_CHILDREN(rack_timely),
1188             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1189             &rack_timely_int_timely_only, 0,
1190             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1191         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1192             SYSCTL_CHILDREN(rack_timely),
1193             OID_AUTO, "nonstop", CTLFLAG_RW,
1194             &rack_timely_no_stopping, 0,
1195             "Rack timely don't stop increase");
1196         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1197             SYSCTL_CHILDREN(rack_timely),
1198             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1199             &rack_down_raise_thresh, 100,
1200             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1201         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1202             SYSCTL_CHILDREN(rack_timely),
1203             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1204             &rack_req_segs, 1,
1205             "Bottom dragging if not these many segments outstanding and room");
1206
1207         /* TLP and Rack related parameters */
1208         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1209             SYSCTL_CHILDREN(rack_sysctl_root),
1210             OID_AUTO,
1211             "tlp",
1212             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1213             "TLP and Rack related Controls");
1214         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1215             SYSCTL_CHILDREN(rack_tlp),
1216             OID_AUTO, "use_rrr", CTLFLAG_RW,
1217             &use_rack_rr, 1,
1218             "Do we use Rack Rapid Recovery");
1219         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1220             SYSCTL_CHILDREN(rack_tlp),
1221             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1222             &rack_max_abc_post_recovery, 2,
1223             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1224         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1225             SYSCTL_CHILDREN(rack_tlp),
1226             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1227             &rack_non_rxt_use_cr, 0,
1228             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1229         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1230             SYSCTL_CHILDREN(rack_tlp),
1231             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1232             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1233             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1234         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1235             SYSCTL_CHILDREN(rack_tlp),
1236             OID_AUTO, "limit", CTLFLAG_RW,
1237             &rack_tlp_limit, 2,
1238             "How many TLP's can be sent without sending new data");
1239         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1240             SYSCTL_CHILDREN(rack_tlp),
1241             OID_AUTO, "use_greater", CTLFLAG_RW,
1242             &rack_tlp_use_greater, 1,
1243             "Should we use the rack_rtt time if its greater than srtt");
1244         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1245             SYSCTL_CHILDREN(rack_tlp),
1246             OID_AUTO, "tlpminto", CTLFLAG_RW,
1247             &rack_tlp_min, 10000,
1248             "TLP minimum timeout per the specification (in microseconds)");
1249         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1250             SYSCTL_CHILDREN(rack_tlp),
1251             OID_AUTO, "send_oldest", CTLFLAG_RW,
1252             &rack_always_send_oldest, 0,
1253             "Should we always send the oldest TLP and RACK-TLP");
1254         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1255             SYSCTL_CHILDREN(rack_tlp),
1256             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1257             &rack_limited_retran, 0,
1258             "How many times can a rack timeout drive out sends");
1259         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1260             SYSCTL_CHILDREN(rack_tlp),
1261             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1262             &rack_lower_cwnd_at_tlp, 0,
1263             "When a TLP completes a retran should we enter recovery");
1264         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1265             SYSCTL_CHILDREN(rack_tlp),
1266             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1267             &rack_reorder_thresh, 2,
1268             "What factor for rack will be added when seeing reordering (shift right)");
1269         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1270             SYSCTL_CHILDREN(rack_tlp),
1271             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1272             &rack_tlp_thresh, 1,
1273             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1274         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1275             SYSCTL_CHILDREN(rack_tlp),
1276             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1277             &rack_reorder_fade, 60000000,
1278             "Does reorder detection fade, if so how many microseconds (0 means never)");
1279         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1280             SYSCTL_CHILDREN(rack_tlp),
1281             OID_AUTO, "pktdelay", CTLFLAG_RW,
1282             &rack_pkt_delay, 1000,
1283             "Extra RACK time (in microseconds) besides reordering thresh");
1284
1285         /* Timer related controls */
1286         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1287             SYSCTL_CHILDREN(rack_sysctl_root),
1288             OID_AUTO,
1289             "timers",
1290             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1291             "Timer related controls");
1292         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1293             SYSCTL_CHILDREN(rack_timers),
1294             OID_AUTO, "persmin", CTLFLAG_RW,
1295             &rack_persist_min, 250000,
1296             "What is the minimum time in microseconds between persists");
1297         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1298             SYSCTL_CHILDREN(rack_timers),
1299             OID_AUTO, "persmax", CTLFLAG_RW,
1300             &rack_persist_max, 2000000,
1301             "What is the largest delay in microseconds between persists");
1302         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1303             SYSCTL_CHILDREN(rack_timers),
1304             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1305             &rack_delayed_ack_time, 40000,
1306             "Delayed ack time (40ms in microseconds)");
1307         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1308             SYSCTL_CHILDREN(rack_timers),
1309             OID_AUTO, "minrto", CTLFLAG_RW,
1310             &rack_rto_min, 30000,
1311             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1312         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1313             SYSCTL_CHILDREN(rack_timers),
1314             OID_AUTO, "maxrto", CTLFLAG_RW,
1315             &rack_rto_max, 4000000,
1316             "Maxiumum RTO in microseconds -- should be at least as large as min_rto");
1317         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1318             SYSCTL_CHILDREN(rack_timers),
1319             OID_AUTO, "minto", CTLFLAG_RW,
1320             &rack_min_to, 1000,
1321             "Minimum rack timeout in microseconds");
1322         /* Measure controls */
1323         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1324             SYSCTL_CHILDREN(rack_sysctl_root),
1325             OID_AUTO,
1326             "measure",
1327             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1328             "Measure related controls");
1329         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1330             SYSCTL_CHILDREN(rack_measure),
1331             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1332             &rack_wma_divisor, 8,
1333             "When doing b/w calculation what is the  divisor for the WMA");
1334         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1335             SYSCTL_CHILDREN(rack_measure),
1336             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1337             &rack_cwnd_block_ends_measure, 0,
1338             "Does a cwnd just-return end the measurement window (app limited)");
1339         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1340             SYSCTL_CHILDREN(rack_measure),
1341             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1342             &rack_rwnd_block_ends_measure, 0,
1343             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1344         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1345             SYSCTL_CHILDREN(rack_measure),
1346             OID_AUTO, "min_target", CTLFLAG_RW,
1347             &rack_def_data_window, 20,
1348             "What is the minimum target window (in mss) for a GP measurements");
1349         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1350             SYSCTL_CHILDREN(rack_measure),
1351             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1352             &rack_goal_bdp, 2,
1353             "What is the goal BDP to measure");
1354         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1355             SYSCTL_CHILDREN(rack_measure),
1356             OID_AUTO, "min_srtts", CTLFLAG_RW,
1357             &rack_min_srtts, 1,
1358             "What is the goal BDP to measure");
1359         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1360             SYSCTL_CHILDREN(rack_measure),
1361             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1362             &rack_min_measure_usec, 0,
1363             "What is the Minimum time time for a measurement if 0, this is off");
1364         /* Misc rack controls */
1365         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1366             SYSCTL_CHILDREN(rack_sysctl_root),
1367             OID_AUTO,
1368             "misc",
1369             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1370             "Misc related controls");
1371 #ifdef TCP_ACCOUNTING
1372         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1373             SYSCTL_CHILDREN(rack_misc),
1374             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1375             &rack_tcp_accounting, 0,
1376             "Should we turn on TCP accounting for all rack sessions?");
1377 #endif
1378         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1379             SYSCTL_CHILDREN(rack_misc),
1380             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1381             &rack_prr_addbackmax, 2,
1382             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1383         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1384             SYSCTL_CHILDREN(rack_misc),
1385             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1386             &rack_stats_gets_ms_rtt, 1,
1387             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1388         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1389             SYSCTL_CHILDREN(rack_misc),
1390             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1391             &rack_client_low_buf, 0,
1392             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1393         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1394             SYSCTL_CHILDREN(rack_misc),
1395             OID_AUTO, "defprofile", CTLFLAG_RW,
1396             &rack_def_profile, 0,
1397             "Should RACK use a default profile (0=no, num == profile num)?");
1398         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1399             SYSCTL_CHILDREN(rack_misc),
1400             OID_AUTO, "cmpack", CTLFLAG_RW,
1401             &rack_use_cmp_acks, 1,
1402             "Should RACK have LRO send compressed acks");
1403         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1404             SYSCTL_CHILDREN(rack_misc),
1405             OID_AUTO, "fsb", CTLFLAG_RW,
1406             &rack_use_fsb, 1,
1407             "Should RACK use the fast send block?");
1408         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1409             SYSCTL_CHILDREN(rack_misc),
1410             OID_AUTO, "rfo", CTLFLAG_RW,
1411             &rack_use_rfo, 1,
1412             "Should RACK use rack_fast_output()?");
1413         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1414             SYSCTL_CHILDREN(rack_misc),
1415             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1416             &rack_use_rsm_rfo, 1,
1417             "Should RACK use rack_fast_rsm_output()?");
1418         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1419             SYSCTL_CHILDREN(rack_misc),
1420             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1421             &rack_enable_shared_cwnd, 1,
1422             "Should RACK try to use the shared cwnd on connections where allowed");
1423         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1424             SYSCTL_CHILDREN(rack_misc),
1425             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1426             &rack_limits_scwnd, 1,
1427             "Should RACK place low end time limits on the shared cwnd feature");
1428         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1429             SYSCTL_CHILDREN(rack_misc),
1430             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1431             &rack_enable_mqueue_for_nonpaced, 0,
1432             "Should RACK use mbuf queuing for non-paced connections");
1433         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1434             SYSCTL_CHILDREN(rack_misc),
1435             OID_AUTO, "iMac_dack", CTLFLAG_RW,
1436             &rack_use_imac_dack, 0,
1437             "Should RACK try to emulate iMac delayed ack");
1438         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1439             SYSCTL_CHILDREN(rack_misc),
1440             OID_AUTO, "no_prr", CTLFLAG_RW,
1441             &rack_disable_prr, 0,
1442             "Should RACK not use prr and only pace (must have pacing on)");
1443         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1444             SYSCTL_CHILDREN(rack_misc),
1445             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1446             &rack_verbose_logging, 0,
1447             "Should RACK black box logging be verbose");
1448         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1449             SYSCTL_CHILDREN(rack_misc),
1450             OID_AUTO, "data_after_close", CTLFLAG_RW,
1451             &rack_ignore_data_after_close, 1,
1452             "Do we hold off sending a RST until all pending data is ack'd");
1453         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1454             SYSCTL_CHILDREN(rack_misc),
1455             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1456             &rack_sack_not_required, 1,
1457             "Do we allow rack to run on connections not supporting SACK");
1458         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1459             SYSCTL_CHILDREN(rack_misc),
1460             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1461             &rack_send_a_lot_in_prr, 1,
1462             "Send a lot in prr");
1463         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1464             SYSCTL_CHILDREN(rack_misc),
1465             OID_AUTO, "autoscale", CTLFLAG_RW,
1466             &rack_autosndbuf_inc, 20,
1467             "What percentage should rack scale up its snd buffer by?");
1468         /* Sack Attacker detection stuff */
1469         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1470             SYSCTL_CHILDREN(rack_attack),
1471             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1472             &rack_highest_sack_thresh_seen, 0,
1473             "Highest sack to ack ratio seen");
1474         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1475             SYSCTL_CHILDREN(rack_attack),
1476             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1477             &rack_highest_move_thresh_seen, 0,
1478             "Highest move to non-move ratio seen");
1479         rack_ack_total = counter_u64_alloc(M_WAITOK);
1480         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1481             SYSCTL_CHILDREN(rack_attack),
1482             OID_AUTO, "acktotal", CTLFLAG_RD,
1483             &rack_ack_total,
1484             "Total number of Ack's");
1485         rack_express_sack = counter_u64_alloc(M_WAITOK);
1486         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1487             SYSCTL_CHILDREN(rack_attack),
1488             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1489             &rack_express_sack,
1490             "Total expresss number of Sack's");
1491         rack_sack_total = counter_u64_alloc(M_WAITOK);
1492         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1493             SYSCTL_CHILDREN(rack_attack),
1494             OID_AUTO, "sacktotal", CTLFLAG_RD,
1495             &rack_sack_total,
1496             "Total number of SACKs");
1497         rack_move_none = counter_u64_alloc(M_WAITOK);
1498         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1499             SYSCTL_CHILDREN(rack_attack),
1500             OID_AUTO, "move_none", CTLFLAG_RD,
1501             &rack_move_none,
1502             "Total number of SACK index reuse of postions under threshold");
1503         rack_move_some = counter_u64_alloc(M_WAITOK);
1504         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1505             SYSCTL_CHILDREN(rack_attack),
1506             OID_AUTO, "move_some", CTLFLAG_RD,
1507             &rack_move_some,
1508             "Total number of SACK index reuse of postions over threshold");
1509         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1510         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1511             SYSCTL_CHILDREN(rack_attack),
1512             OID_AUTO, "attacks", CTLFLAG_RD,
1513             &rack_sack_attacks_detected,
1514             "Total number of SACK attackers that had sack disabled");
1515         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1516         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1517             SYSCTL_CHILDREN(rack_attack),
1518             OID_AUTO, "reversed", CTLFLAG_RD,
1519             &rack_sack_attacks_reversed,
1520             "Total number of SACK attackers that were later determined false positive");
1521         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1522         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1523             SYSCTL_CHILDREN(rack_attack),
1524             OID_AUTO, "nextmerge", CTLFLAG_RD,
1525             &rack_sack_used_next_merge,
1526             "Total number of times we used the next merge");
1527         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1528         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1529             SYSCTL_CHILDREN(rack_attack),
1530             OID_AUTO, "prevmerge", CTLFLAG_RD,
1531             &rack_sack_used_prev_merge,
1532             "Total number of times we used the prev merge");
1533         /* Counters */
1534         rack_fto_send = counter_u64_alloc(M_WAITOK);
1535         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1536             SYSCTL_CHILDREN(rack_counters),
1537             OID_AUTO, "fto_send", CTLFLAG_RD,
1538             &rack_fto_send, "Total number of rack_fast_output sends");
1539         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1540         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1541             SYSCTL_CHILDREN(rack_counters),
1542             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1543             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1544         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1545         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1546             SYSCTL_CHILDREN(rack_counters),
1547             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1548             &rack_nfto_resend, "Total number of rack_output retransmissions");
1549         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1550         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1551             SYSCTL_CHILDREN(rack_counters),
1552             OID_AUTO, "nfto_send", CTLFLAG_RD,
1553             &rack_non_fto_send, "Total number of rack_output first sends");
1554         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1555         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1556             SYSCTL_CHILDREN(rack_counters),
1557             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1558             &rack_extended_rfo, "Total number of times we extended rfo");
1559
1560         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1561         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1562             SYSCTL_CHILDREN(rack_counters),
1563             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1564             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1565         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1566
1567         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1568             SYSCTL_CHILDREN(rack_counters),
1569             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1570             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1571         rack_badfr = counter_u64_alloc(M_WAITOK);
1572         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1573             SYSCTL_CHILDREN(rack_counters),
1574             OID_AUTO, "badfr", CTLFLAG_RD,
1575             &rack_badfr, "Total number of bad FRs");
1576         rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1577         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1578             SYSCTL_CHILDREN(rack_counters),
1579             OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1580             &rack_badfr_bytes, "Total number of bad FRs");
1581         rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1582         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1583             SYSCTL_CHILDREN(rack_counters),
1584             OID_AUTO, "prrsndret", CTLFLAG_RD,
1585             &rack_rtm_prr_retran,
1586             "Total number of prr based retransmits");
1587         rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1588         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1589             SYSCTL_CHILDREN(rack_counters),
1590             OID_AUTO, "prrsndnew", CTLFLAG_RD,
1591             &rack_rtm_prr_newdata,
1592             "Total number of prr based new transmits");
1593         rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1594         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1595             SYSCTL_CHILDREN(rack_counters),
1596             OID_AUTO, "tsnf", CTLFLAG_RD,
1597             &rack_timestamp_mismatch,
1598             "Total number of timestamps that we could not find the reported ts");
1599         rack_find_high = counter_u64_alloc(M_WAITOK);
1600         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1601             SYSCTL_CHILDREN(rack_counters),
1602             OID_AUTO, "findhigh", CTLFLAG_RD,
1603             &rack_find_high,
1604             "Total number of FIN causing find-high");
1605         rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1606         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1607             SYSCTL_CHILDREN(rack_counters),
1608             OID_AUTO, "reordering", CTLFLAG_RD,
1609             &rack_reorder_seen,
1610             "Total number of times we added delay due to reordering");
1611         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1612         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1613             SYSCTL_CHILDREN(rack_counters),
1614             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1615             &rack_tlp_tot,
1616             "Total number of tail loss probe expirations");
1617         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1618         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1619             SYSCTL_CHILDREN(rack_counters),
1620             OID_AUTO, "tlp_new", CTLFLAG_RD,
1621             &rack_tlp_newdata,
1622             "Total number of tail loss probe sending new data");
1623         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1624         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1625             SYSCTL_CHILDREN(rack_counters),
1626             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1627             &rack_tlp_retran,
1628             "Total number of tail loss probe sending retransmitted data");
1629         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1630         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1631             SYSCTL_CHILDREN(rack_counters),
1632             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1633             &rack_tlp_retran_bytes,
1634             "Total bytes of tail loss probe sending retransmitted data");
1635         rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1636         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1637             SYSCTL_CHILDREN(rack_counters),
1638             OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1639             &rack_tlp_retran_fail,
1640             "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1641         rack_to_tot = counter_u64_alloc(M_WAITOK);
1642         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1643             SYSCTL_CHILDREN(rack_counters),
1644             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1645             &rack_to_tot,
1646             "Total number of times the rack to expired");
1647         rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1648         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1649             SYSCTL_CHILDREN(rack_counters),
1650             OID_AUTO, "arm_rack", CTLFLAG_RD,
1651             &rack_to_arm_rack,
1652             "Total number of times the rack timer armed");
1653         rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1654         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1655             SYSCTL_CHILDREN(rack_counters),
1656             OID_AUTO, "arm_tlp", CTLFLAG_RD,
1657             &rack_to_arm_tlp,
1658             "Total number of times the tlp timer armed");
1659         rack_calc_zero = counter_u64_alloc(M_WAITOK);
1660         rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1661         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1662             SYSCTL_CHILDREN(rack_counters),
1663             OID_AUTO, "calc_zero", CTLFLAG_RD,
1664             &rack_calc_zero,
1665             "Total number of times pacing time worked out to zero");
1666         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1667             SYSCTL_CHILDREN(rack_counters),
1668             OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1669             &rack_calc_nonzero,
1670             "Total number of times pacing time worked out to non-zero");
1671         rack_paced_segments = counter_u64_alloc(M_WAITOK);
1672         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1673             SYSCTL_CHILDREN(rack_counters),
1674             OID_AUTO, "paced", CTLFLAG_RD,
1675             &rack_paced_segments,
1676             "Total number of times a segment send caused hptsi");
1677         rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1678         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1679             SYSCTL_CHILDREN(rack_counters),
1680             OID_AUTO, "unpaced", CTLFLAG_RD,
1681             &rack_unpaced_segments,
1682             "Total number of times a segment did not cause hptsi");
1683         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1684         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1685             SYSCTL_CHILDREN(rack_counters),
1686             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1687             &rack_saw_enobuf,
1688             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1689         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1690         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1691             SYSCTL_CHILDREN(rack_counters),
1692             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1693             &rack_saw_enobuf_hw,
1694             "Total number of times a send returned enobuf for hdwr paced connections");
1695         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1696         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1697             SYSCTL_CHILDREN(rack_counters),
1698             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1699             &rack_saw_enetunreach,
1700             "Total number of times a send received a enetunreachable");
1701         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1702         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1703             SYSCTL_CHILDREN(rack_counters),
1704             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1705             &rack_hot_alloc,
1706             "Total allocations from the top of our list");
1707         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1708         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1709             SYSCTL_CHILDREN(rack_counters),
1710             OID_AUTO, "allocs", CTLFLAG_RD,
1711             &rack_to_alloc,
1712             "Total allocations of tracking structures");
1713         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1714         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1715             SYSCTL_CHILDREN(rack_counters),
1716             OID_AUTO, "allochard", CTLFLAG_RD,
1717             &rack_to_alloc_hard,
1718             "Total allocations done with sleeping the hard way");
1719         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1720         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1721             SYSCTL_CHILDREN(rack_counters),
1722             OID_AUTO, "allocemerg", CTLFLAG_RD,
1723             &rack_to_alloc_emerg,
1724             "Total allocations done from emergency cache");
1725         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1726         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1727             SYSCTL_CHILDREN(rack_counters),
1728             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1729             &rack_to_alloc_limited,
1730             "Total allocations dropped due to limit");
1731         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1732         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1733             SYSCTL_CHILDREN(rack_counters),
1734             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1735             &rack_alloc_limited_conns,
1736             "Connections with allocations dropped due to limit");
1737         rack_split_limited = counter_u64_alloc(M_WAITOK);
1738         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1739             SYSCTL_CHILDREN(rack_counters),
1740             OID_AUTO, "split_limited", CTLFLAG_RD,
1741             &rack_split_limited,
1742             "Split allocations dropped due to limit");
1743
1744         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
1745                 char name[32];
1746                 sprintf(name, "cmp_ack_cnt_%d", i);
1747                 rack_proc_comp_ack[i] = counter_u64_alloc(M_WAITOK);
1748                 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1749                                        SYSCTL_CHILDREN(rack_counters),
1750                                        OID_AUTO, name, CTLFLAG_RD,
1751                                        &rack_proc_comp_ack[i],
1752                                        "Number of compressed acks we processed");
1753         }
1754         rack_large_ackcmp = counter_u64_alloc(M_WAITOK);
1755         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756             SYSCTL_CHILDREN(rack_counters),
1757             OID_AUTO, "cmp_large_mbufs", CTLFLAG_RD,
1758             &rack_large_ackcmp,
1759             "Number of TCP connections with large mbuf's for compressed acks");
1760         rack_small_ackcmp = counter_u64_alloc(M_WAITOK);
1761         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1762             SYSCTL_CHILDREN(rack_counters),
1763             OID_AUTO, "cmp_small_mbufs", CTLFLAG_RD,
1764             &rack_small_ackcmp,
1765             "Number of TCP connections with small mbuf's for compressed acks");
1766 #ifdef INVARIANTS
1767         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1768         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1769             SYSCTL_CHILDREN(rack_counters),
1770             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1771             &rack_adjust_map_bw,
1772             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1773 #endif
1774         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1775         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1776             SYSCTL_CHILDREN(rack_counters),
1777             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1778             &rack_multi_single_eq,
1779             "Number of compressed acks total represented");
1780         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1781         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1782             SYSCTL_CHILDREN(rack_counters),
1783             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1784             &rack_proc_non_comp_ack,
1785             "Number of non compresseds acks that we processed");
1786
1787
1788         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1789         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1790             SYSCTL_CHILDREN(rack_counters),
1791             OID_AUTO, "sack_long", CTLFLAG_RD,
1792             &rack_sack_proc_all,
1793             "Total times we had to walk whole list for sack processing");
1794         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1795         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1796             SYSCTL_CHILDREN(rack_counters),
1797             OID_AUTO, "sack_restart", CTLFLAG_RD,
1798             &rack_sack_proc_restart,
1799             "Total times we had to walk whole list due to a restart");
1800         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1801         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1802             SYSCTL_CHILDREN(rack_counters),
1803             OID_AUTO, "sack_short", CTLFLAG_RD,
1804             &rack_sack_proc_short,
1805             "Total times we took shortcut for sack processing");
1806         rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1807         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1808             SYSCTL_CHILDREN(rack_counters),
1809             OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1810             &rack_enter_tlp_calc,
1811             "Total times we called calc-tlp");
1812         rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1813         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1814             SYSCTL_CHILDREN(rack_counters),
1815             OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1816             &rack_used_tlpmethod,
1817             "Total number of runt sacks");
1818         rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1819         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1820             SYSCTL_CHILDREN(rack_counters),
1821             OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1822             &rack_used_tlpmethod2,
1823             "Total number of times we hit TLP method 2");
1824         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1825         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1826             SYSCTL_CHILDREN(rack_attack),
1827             OID_AUTO, "skipacked", CTLFLAG_RD,
1828             &rack_sack_skipped_acked,
1829             "Total number of times we skipped previously sacked");
1830         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1831         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1832             SYSCTL_CHILDREN(rack_attack),
1833             OID_AUTO, "ofsplit", CTLFLAG_RD,
1834             &rack_sack_splits,
1835             "Total number of times we did the old fashion tree split");
1836         rack_progress_drops = counter_u64_alloc(M_WAITOK);
1837         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1838             SYSCTL_CHILDREN(rack_counters),
1839             OID_AUTO, "prog_drops", CTLFLAG_RD,
1840             &rack_progress_drops,
1841             "Total number of progress drops");
1842         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1843         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1844             SYSCTL_CHILDREN(rack_counters),
1845             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1846             &rack_input_idle_reduces,
1847             "Total number of idle reductions on input");
1848         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1849         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1850             SYSCTL_CHILDREN(rack_counters),
1851             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1852             &rack_collapsed_win,
1853             "Total number of collapsed windows");
1854         rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1855         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1856             SYSCTL_CHILDREN(rack_counters),
1857             OID_AUTO, "tlp_nada", CTLFLAG_RD,
1858             &rack_tlp_does_nada,
1859             "Total number of nada tlp calls");
1860         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1861         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1862             SYSCTL_CHILDREN(rack_counters),
1863             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1864             &rack_try_scwnd,
1865             "Total number of scwnd attempts");
1866
1867         rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1868         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1869             SYSCTL_CHILDREN(rack_counters),
1870             OID_AUTO, "timer_hole", CTLFLAG_RD,
1871             &rack_per_timer_hole,
1872             "Total persists start in timer hole");
1873
1874         rack_sbsndptr_wrong = counter_u64_alloc(M_WAITOK);
1875         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1876             SYSCTL_CHILDREN(rack_counters),
1877             OID_AUTO, "sndptr_wrong", CTLFLAG_RD,
1878             &rack_sbsndptr_wrong, "Total number of times the saved sbsndptr was incorret");
1879         rack_sbsndptr_right = counter_u64_alloc(M_WAITOK);
1880         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1881             SYSCTL_CHILDREN(rack_counters),
1882             OID_AUTO, "sndptr_right", CTLFLAG_RD,
1883             &rack_sbsndptr_right, "Total number of times the saved sbsndptr was corret");
1884
1885         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1886         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1887             OID_AUTO, "outsize", CTLFLAG_RD,
1888             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1889         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1890         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1891             OID_AUTO, "opts", CTLFLAG_RD,
1892             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1893         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1894             SYSCTL_CHILDREN(rack_sysctl_root),
1895             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1896             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1897 }
1898
1899 static __inline int
1900 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1901 {
1902         if (SEQ_GEQ(b->r_start, a->r_start) &&
1903             SEQ_LT(b->r_start, a->r_end)) {
1904                 /*
1905                  * The entry b is within the
1906                  * block a. i.e.:
1907                  * a --   |-------------|
1908                  * b --   |----|
1909                  * <or>
1910                  * b --       |------|
1911                  * <or>
1912                  * b --       |-----------|
1913                  */
1914                 return (0);
1915         } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1916                 /*
1917                  * b falls as either the next
1918                  * sequence block after a so a
1919                  * is said to be smaller than b.
1920                  * i.e:
1921                  * a --   |------|
1922                  * b --          |--------|
1923                  * or
1924                  * b --              |-----|
1925                  */
1926                 return (1);
1927         }
1928         /*
1929          * Whats left is where a is
1930          * larger than b. i.e:
1931          * a --         |-------|
1932          * b --  |---|
1933          * or even possibly
1934          * b --   |--------------|
1935          */
1936         return (-1);
1937 }
1938
1939 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1940 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1941
1942 static uint32_t
1943 rc_init_window(struct tcp_rack *rack)
1944 {
1945         uint32_t win;
1946
1947         if (rack->rc_init_win == 0) {
1948                 /*
1949                  * Nothing set by the user, use the system stack
1950                  * default.
1951                  */
1952                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1953         }
1954         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1955         return (win);
1956 }
1957
1958 static uint64_t
1959 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1960 {
1961         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1962                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1963         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1964                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1965         else
1966                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1967 }
1968
1969 static uint64_t
1970 rack_get_bw(struct tcp_rack *rack)
1971 {
1972         if (rack->use_fixed_rate) {
1973                 /* Return the fixed pacing rate */
1974                 return (rack_get_fixed_pacing_bw(rack));
1975         }
1976         if (rack->r_ctl.gp_bw == 0) {
1977                 /*
1978                  * We have yet no b/w measurement,
1979                  * if we have a user set initial bw
1980                  * return it. If we don't have that and
1981                  * we have an srtt, use the tcp IW (10) to
1982                  * calculate a fictional b/w over the SRTT
1983                  * which is more or less a guess. Note
1984                  * we don't use our IW from rack on purpose
1985                  * so if we have like IW=30, we are not
1986                  * calculating a "huge" b/w.
1987                  */
1988                 uint64_t bw, srtt;
1989                 if (rack->r_ctl.init_rate)
1990                         return (rack->r_ctl.init_rate);
1991
1992                 /* Has the user set a max peak rate? */
1993 #ifdef NETFLIX_PEAKRATE
1994                 if (rack->rc_tp->t_maxpeakrate)
1995                         return (rack->rc_tp->t_maxpeakrate);
1996 #endif
1997                 /* Ok lets come up with the IW guess, if we have a srtt */
1998                 if (rack->rc_tp->t_srtt == 0) {
1999                         /*
2000                          * Go with old pacing method
2001                          * i.e. burst mitigation only.
2002                          */
2003                         return (0);
2004                 }
2005                 /* Ok lets get the initial TCP win (not racks) */
2006                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2007                 srtt = (uint64_t)rack->rc_tp->t_srtt;
2008                 bw *= (uint64_t)USECS_IN_SECOND;
2009                 bw /= srtt;
2010                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2011                         bw = rack->r_ctl.bw_rate_cap;
2012                 return (bw);
2013         } else {
2014                 uint64_t bw;
2015
2016                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2017                         /* Averaging is done, we can return the value */
2018                         bw = rack->r_ctl.gp_bw;
2019                 } else {
2020                         /* Still doing initial average must calculate */
2021                         bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
2022                 }
2023 #ifdef NETFLIX_PEAKRATE
2024                 if ((rack->rc_tp->t_maxpeakrate) &&
2025                     (bw > rack->rc_tp->t_maxpeakrate)) {
2026                         /* The user has set a peak rate to pace at
2027                          * don't allow us to pace faster than that.
2028                          */
2029                         return (rack->rc_tp->t_maxpeakrate);
2030                 }
2031 #endif
2032                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2033                         bw = rack->r_ctl.bw_rate_cap;
2034                 return (bw);
2035         }
2036 }
2037
2038 static uint16_t
2039 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2040 {
2041         if (rack->use_fixed_rate) {
2042                 return (100);
2043         } else if (rack->in_probe_rtt && (rsm == NULL))
2044                 return (rack->r_ctl.rack_per_of_gp_probertt);
2045         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2046                   rack->r_ctl.rack_per_of_gp_rec)) {
2047                 if (rsm) {
2048                         /* a retransmission always use the recovery rate */
2049                         return (rack->r_ctl.rack_per_of_gp_rec);
2050                 } else if (rack->rack_rec_nonrxt_use_cr) {
2051                         /* Directed to use the configured rate */
2052                         goto configured_rate;
2053                 } else if (rack->rack_no_prr &&
2054                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2055                         /* No PRR, lets just use the b/w estimate only */
2056                         return (100);
2057                 } else {
2058                         /*
2059                          * Here we may have a non-retransmit but we
2060                          * have no overrides, so just use the recovery
2061                          * rate (prr is in effect).
2062                          */
2063                         return (rack->r_ctl.rack_per_of_gp_rec);
2064                 }
2065         }
2066 configured_rate:
2067         /* For the configured rate we look at our cwnd vs the ssthresh */
2068         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2069                 return (rack->r_ctl.rack_per_of_gp_ss);
2070         else
2071                 return (rack->r_ctl.rack_per_of_gp_ca);
2072 }
2073
2074 static void
2075 rack_log_hdwr_pacing(struct tcp_rack *rack,
2076                      uint64_t rate, uint64_t hw_rate, int line,
2077                      int error, uint16_t mod)
2078 {
2079         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2080                 union tcp_log_stackspecific log;
2081                 struct timeval tv;
2082                 const struct ifnet *ifp;
2083
2084                 memset(&log, 0, sizeof(log));
2085                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2086                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2087                 if (rack->r_ctl.crte) {
2088                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2089                 } else if (rack->rc_inp->inp_route.ro_nh &&
2090                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2091                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2092                 } else
2093                         ifp = NULL;
2094                 if (ifp) {
2095                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2096                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2097                 }
2098                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2099                 log.u_bbr.bw_inuse = rate;
2100                 log.u_bbr.flex5 = line;
2101                 log.u_bbr.flex6 = error;
2102                 log.u_bbr.flex7 = mod;
2103                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2104                 log.u_bbr.flex8 = rack->use_fixed_rate;
2105                 log.u_bbr.flex8 <<= 1;
2106                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2107                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2108                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2109                 if (rack->r_ctl.crte)
2110                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2111                 else
2112                         log.u_bbr.cur_del_rate = 0;
2113                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2114                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2115                     &rack->rc_inp->inp_socket->so_rcv,
2116                     &rack->rc_inp->inp_socket->so_snd,
2117                     BBR_LOG_HDWR_PACE, 0,
2118                     0, &log, false, &tv);
2119         }
2120 }
2121
2122 static uint64_t
2123 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2124 {
2125         /*
2126          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2127          */
2128         uint64_t bw_est, high_rate;
2129         uint64_t gain;
2130
2131         gain = (uint64_t)rack_get_output_gain(rack, rsm);
2132         bw_est = bw * gain;
2133         bw_est /= (uint64_t)100;
2134         /* Never fall below the minimum (def 64kbps) */
2135         if (bw_est < RACK_MIN_BW)
2136                 bw_est = RACK_MIN_BW;
2137         if (rack->r_rack_hw_rate_caps) {
2138                 /* Rate caps are in place */
2139                 if (rack->r_ctl.crte != NULL) {
2140                         /* We have a hdwr rate already */
2141                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2142                         if (bw_est >= high_rate) {
2143                                 /* We are capping bw at the highest rate table entry */
2144                                 rack_log_hdwr_pacing(rack,
2145                                                      bw_est, high_rate, __LINE__,
2146                                                      0, 3);
2147                                 bw_est = high_rate;
2148                                 if (capped)
2149                                         *capped = 1;
2150                         }
2151                 } else if ((rack->rack_hdrw_pacing == 0) &&
2152                            (rack->rack_hdw_pace_ena) &&
2153                            (rack->rack_attempt_hdwr_pace == 0) &&
2154                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2155                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2156                         /*
2157                          * Special case, we have not yet attempted hardware
2158                          * pacing, and yet we may, when we do, find out if we are
2159                          * above the highest rate. We need to know the maxbw for the interface
2160                          * in question (if it supports ratelimiting). We get back
2161                          * a 0, if the interface is not found in the RL lists.
2162                          */
2163                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2164                         if (high_rate) {
2165                                 /* Yep, we have a rate is it above this rate? */
2166                                 if (bw_est > high_rate) {
2167                                         bw_est = high_rate;
2168                                         if (capped)
2169                                                 *capped = 1;
2170                                 }
2171                         }
2172                 }
2173         }
2174         return (bw_est);
2175 }
2176
2177 static void
2178 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2179 {
2180         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2181                 union tcp_log_stackspecific log;
2182                 struct timeval tv;
2183
2184                 if ((mod != 1) && (rack_verbose_logging == 0)) {
2185                         /*
2186                          * We get 3 values currently for mod
2187                          * 1 - We are retransmitting and this tells the reason.
2188                          * 2 - We are clearing a dup-ack count.
2189                          * 3 - We are incrementing a dup-ack count.
2190                          *
2191                          * The clear/increment are only logged
2192                          * if you have BBverbose on.
2193                          */
2194                         return;
2195                 }
2196                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2197                 log.u_bbr.flex1 = tsused;
2198                 log.u_bbr.flex2 = thresh;
2199                 log.u_bbr.flex3 = rsm->r_flags;
2200                 log.u_bbr.flex4 = rsm->r_dupack;
2201                 log.u_bbr.flex5 = rsm->r_start;
2202                 log.u_bbr.flex6 = rsm->r_end;
2203                 log.u_bbr.flex8 = mod;
2204                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2205                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2206                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2207                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2208                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2209                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2210                 log.u_bbr.pacing_gain = rack->r_must_retran;
2211                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2212                     &rack->rc_inp->inp_socket->so_rcv,
2213                     &rack->rc_inp->inp_socket->so_snd,
2214                     BBR_LOG_SETTINGS_CHG, 0,
2215                     0, &log, false, &tv);
2216         }
2217 }
2218
2219 static void
2220 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2221 {
2222         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2223                 union tcp_log_stackspecific log;
2224                 struct timeval tv;
2225
2226                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2227                 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2228                 log.u_bbr.flex2 = to;
2229                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2230                 log.u_bbr.flex4 = slot;
2231                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2232                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2233                 log.u_bbr.flex7 = rack->rc_in_persist;
2234                 log.u_bbr.flex8 = which;
2235                 if (rack->rack_no_prr)
2236                         log.u_bbr.pkts_out = 0;
2237                 else
2238                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2239                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2240                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2241                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2242                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2243                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2244                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2245                 log.u_bbr.pacing_gain = rack->r_must_retran;
2246                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2247                 log.u_bbr.lost = rack_rto_min;
2248                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2249                     &rack->rc_inp->inp_socket->so_rcv,
2250                     &rack->rc_inp->inp_socket->so_snd,
2251                     BBR_LOG_TIMERSTAR, 0,
2252                     0, &log, false, &tv);
2253         }
2254 }
2255
2256 static void
2257 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2258 {
2259         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2260                 union tcp_log_stackspecific log;
2261                 struct timeval tv;
2262
2263                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2264                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2265                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2266                 log.u_bbr.flex8 = to_num;
2267                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2268                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2269                 if (rsm == NULL)
2270                         log.u_bbr.flex3 = 0;
2271                 else
2272                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2273                 if (rack->rack_no_prr)
2274                         log.u_bbr.flex5 = 0;
2275                 else
2276                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2277                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2278                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2279                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2280                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2281                 log.u_bbr.pacing_gain = rack->r_must_retran;
2282                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2283                     &rack->rc_inp->inp_socket->so_rcv,
2284                     &rack->rc_inp->inp_socket->so_snd,
2285                     BBR_LOG_RTO, 0,
2286                     0, &log, false, &tv);
2287         }
2288 }
2289
2290 static void
2291 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2292                  struct rack_sendmap *prev,
2293                  struct rack_sendmap *rsm,
2294                  struct rack_sendmap *next,
2295                  int flag, uint32_t th_ack, int line)
2296 {
2297         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2298                 union tcp_log_stackspecific log;
2299                 struct timeval tv;
2300
2301                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2302                 log.u_bbr.flex8 = flag;
2303                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2304                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2305                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2306                 log.u_bbr.delRate = (uint64_t)rsm;
2307                 log.u_bbr.rttProp = (uint64_t)next;
2308                 log.u_bbr.flex7 = 0;
2309                 if (prev) {
2310                         log.u_bbr.flex1 = prev->r_start;
2311                         log.u_bbr.flex2 = prev->r_end;
2312                         log.u_bbr.flex7 |= 0x4;
2313                 }
2314                 if (rsm) {
2315                         log.u_bbr.flex3 = rsm->r_start;
2316                         log.u_bbr.flex4 = rsm->r_end;
2317                         log.u_bbr.flex7 |= 0x2;
2318                 }
2319                 if (next) {
2320                         log.u_bbr.flex5 = next->r_start;
2321                         log.u_bbr.flex6 = next->r_end;
2322                         log.u_bbr.flex7 |= 0x1;
2323                 }
2324                 log.u_bbr.applimited = line;
2325                 log.u_bbr.pkts_out = th_ack;
2326                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2327                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2328                 if (rack->rack_no_prr)
2329                         log.u_bbr.lost = 0;
2330                 else
2331                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2332                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2333                     &rack->rc_inp->inp_socket->so_rcv,
2334                     &rack->rc_inp->inp_socket->so_snd,
2335                     TCP_LOG_MAPCHG, 0,
2336                     0, &log, false, &tv);
2337         }
2338 }
2339
2340 static void
2341 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2342                  struct rack_sendmap *rsm, int conf)
2343 {
2344         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2345                 union tcp_log_stackspecific log;
2346                 struct timeval tv;
2347                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2348                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2349                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2350                 log.u_bbr.flex1 = t;
2351                 log.u_bbr.flex2 = len;
2352                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2353                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2354                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2355                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2356                 log.u_bbr.flex7 = conf;
2357                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2358                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2359                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2360                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2361                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2362                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2363                 if (rsm) {
2364                         log.u_bbr.pkt_epoch = rsm->r_start;
2365                         log.u_bbr.lost = rsm->r_end;
2366                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2367                         log.u_bbr.pacing_gain = rsm->r_flags;
2368                 } else {
2369                         /* Its a SYN */
2370                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2371                         log.u_bbr.lost = 0;
2372                         log.u_bbr.cwnd_gain = 0;
2373                         log.u_bbr.pacing_gain = 0;
2374                 }
2375                 /* Write out general bits of interest rrs here */
2376                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2377                 log.u_bbr.use_lt_bw <<= 1;
2378                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2379                 log.u_bbr.use_lt_bw <<= 1;
2380                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2381                 log.u_bbr.use_lt_bw <<= 1;
2382                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2383                 log.u_bbr.use_lt_bw <<= 1;
2384                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2385                 log.u_bbr.use_lt_bw <<= 1;
2386                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2387                 log.u_bbr.use_lt_bw <<= 1;
2388                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2389                 log.u_bbr.use_lt_bw <<= 1;
2390                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2391                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2392                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2393                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2394                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2395                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2396                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2397                 log.u_bbr.bw_inuse <<= 32;
2398                 if (rsm)
2399                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2400                 TCP_LOG_EVENTP(tp, NULL,
2401                     &rack->rc_inp->inp_socket->so_rcv,
2402                     &rack->rc_inp->inp_socket->so_snd,
2403                     BBR_LOG_BBRRTT, 0,
2404                     0, &log, false, &tv);
2405
2406
2407         }
2408 }
2409
2410 static void
2411 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2412 {
2413         /*
2414          * Log the rtt sample we are
2415          * applying to the srtt algorithm in
2416          * useconds.
2417          */
2418         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2419                 union tcp_log_stackspecific log;
2420                 struct timeval tv;
2421
2422                 /* Convert our ms to a microsecond */
2423                 memset(&log, 0, sizeof(log));
2424                 log.u_bbr.flex1 = rtt;
2425                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2426                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2427                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2428                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2429                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2430                 log.u_bbr.flex7 = 1;
2431                 log.u_bbr.flex8 = rack->sack_attack_disable;
2432                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2433                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2434                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2435                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2436                 log.u_bbr.pacing_gain = rack->r_must_retran;
2437                 /*
2438                  * We capture in delRate the upper 32 bits as
2439                  * the confidence level we had declared, and the
2440                  * lower 32 bits as the actual RTT using the arrival
2441                  * timestamp.
2442                  */
2443                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2444                 log.u_bbr.delRate <<= 32;
2445                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2446                 /* Lets capture all the things that make up t_rtxcur */
2447                 log.u_bbr.applimited = rack_rto_min;
2448                 log.u_bbr.epoch = rack_rto_max;
2449                 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2450                 log.u_bbr.lost = rack_rto_min;
2451                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2452                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2453                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2454                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2455                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2456                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2457                     &rack->rc_inp->inp_socket->so_rcv,
2458                     &rack->rc_inp->inp_socket->so_snd,
2459                     TCP_LOG_RTT, 0,
2460                     0, &log, false, &tv);
2461         }
2462 }
2463
2464 static void
2465 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2466 {
2467         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2468                 union tcp_log_stackspecific log;
2469                 struct timeval tv;
2470
2471                 /* Convert our ms to a microsecond */
2472                 memset(&log, 0, sizeof(log));
2473                 log.u_bbr.flex1 = rtt;
2474                 log.u_bbr.flex2 = send_time;
2475                 log.u_bbr.flex3 = ack_time;
2476                 log.u_bbr.flex4 = where;
2477                 log.u_bbr.flex7 = 2;
2478                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2479                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2480                     &rack->rc_inp->inp_socket->so_rcv,
2481                     &rack->rc_inp->inp_socket->so_snd,
2482                     TCP_LOG_RTT, 0,
2483                     0, &log, false, &tv);
2484         }
2485 }
2486
2487
2488
2489 static inline void
2490 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2491 {
2492         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2493                 union tcp_log_stackspecific log;
2494                 struct timeval tv;
2495
2496                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2497                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2498                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2499                 log.u_bbr.flex1 = line;
2500                 log.u_bbr.flex2 = tick;
2501                 log.u_bbr.flex3 = tp->t_maxunacktime;
2502                 log.u_bbr.flex4 = tp->t_acktime;
2503                 log.u_bbr.flex8 = event;
2504                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2505                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2506                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2507                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2508                 log.u_bbr.pacing_gain = rack->r_must_retran;
2509                 TCP_LOG_EVENTP(tp, NULL,
2510                     &rack->rc_inp->inp_socket->so_rcv,
2511                     &rack->rc_inp->inp_socket->so_snd,
2512                     BBR_LOG_PROGRESS, 0,
2513                     0, &log, false, &tv);
2514         }
2515 }
2516
2517 static void
2518 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2519 {
2520         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2521                 union tcp_log_stackspecific log;
2522
2523                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2524                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2525                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2526                 log.u_bbr.flex1 = slot;
2527                 if (rack->rack_no_prr)
2528                         log.u_bbr.flex2 = 0;
2529                 else
2530                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2531                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2532                 log.u_bbr.flex8 = rack->rc_in_persist;
2533                 log.u_bbr.timeStamp = cts;
2534                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2535                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2536                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2537                 log.u_bbr.pacing_gain = rack->r_must_retran;
2538                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2539                     &rack->rc_inp->inp_socket->so_rcv,
2540                     &rack->rc_inp->inp_socket->so_snd,
2541                     BBR_LOG_BBRSND, 0,
2542                     0, &log, false, tv);
2543         }
2544 }
2545
2546 static void
2547 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2548 {
2549         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2550                 union tcp_log_stackspecific log;
2551                 struct timeval tv;
2552
2553                 memset(&log, 0, sizeof(log));
2554                 log.u_bbr.flex1 = did_out;
2555                 log.u_bbr.flex2 = nxt_pkt;
2556                 log.u_bbr.flex3 = way_out;
2557                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2558                 if (rack->rack_no_prr)
2559                         log.u_bbr.flex5 = 0;
2560                 else
2561                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2562                 log.u_bbr.flex6 = nsegs;
2563                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2564                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2565                 log.u_bbr.flex7 <<= 1;
2566                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2567                 log.u_bbr.flex7 <<= 1;
2568                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2569                 log.u_bbr.flex8 = rack->rc_in_persist;
2570                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2571                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2572                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2573                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2574                 log.u_bbr.use_lt_bw <<= 1;
2575                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2576                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2577                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2578                 log.u_bbr.pacing_gain = rack->r_must_retran;
2579                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2580                     &rack->rc_inp->inp_socket->so_rcv,
2581                     &rack->rc_inp->inp_socket->so_snd,
2582                     BBR_LOG_DOSEG_DONE, 0,
2583                     0, &log, false, &tv);
2584         }
2585 }
2586
2587 static void
2588 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2589 {
2590         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2591                 union tcp_log_stackspecific log;
2592                 struct timeval tv;
2593                 uint32_t cts;
2594
2595                 memset(&log, 0, sizeof(log));
2596                 cts = tcp_get_usecs(&tv);
2597                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2598                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2599                 log.u_bbr.flex4 = arg1;
2600                 log.u_bbr.flex5 = arg2;
2601                 log.u_bbr.flex6 = arg3;
2602                 log.u_bbr.flex8 = frm;
2603                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2604                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2605                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2606                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2607                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2608                 log.u_bbr.pacing_gain = rack->r_must_retran;
2609                 TCP_LOG_EVENTP(tp, NULL,
2610                     &tp->t_inpcb->inp_socket->so_rcv,
2611                     &tp->t_inpcb->inp_socket->so_snd,
2612                     TCP_HDWR_PACE_SIZE, 0,
2613                     0, &log, false, &tv);
2614         }
2615 }
2616
2617 static void
2618 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2619                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2620 {
2621         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2622                 union tcp_log_stackspecific log;
2623                 struct timeval tv;
2624
2625                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2626                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2627                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2628                 log.u_bbr.flex1 = slot;
2629                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2630                 log.u_bbr.flex4 = reason;
2631                 if (rack->rack_no_prr)
2632                         log.u_bbr.flex5 = 0;
2633                 else
2634                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2635                 log.u_bbr.flex7 = hpts_calling;
2636                 log.u_bbr.flex8 = rack->rc_in_persist;
2637                 log.u_bbr.lt_epoch = cwnd_to_use;
2638                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2639                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2640                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2641                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2642                 log.u_bbr.pacing_gain = rack->r_must_retran;
2643                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2644                     &rack->rc_inp->inp_socket->so_rcv,
2645                     &rack->rc_inp->inp_socket->so_snd,
2646                     BBR_LOG_JUSTRET, 0,
2647                     tlen, &log, false, &tv);
2648         }
2649 }
2650
2651 static void
2652 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2653                    struct timeval *tv, uint32_t flags_on_entry)
2654 {
2655         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2656                 union tcp_log_stackspecific log;
2657
2658                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2659                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2660                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2661                 log.u_bbr.flex1 = line;
2662                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2663                 log.u_bbr.flex3 = flags_on_entry;
2664                 log.u_bbr.flex4 = us_cts;
2665                 if (rack->rack_no_prr)
2666                         log.u_bbr.flex5 = 0;
2667                 else
2668                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2669                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2670                 log.u_bbr.flex7 = hpts_removed;
2671                 log.u_bbr.flex8 = 1;
2672                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2673                 log.u_bbr.timeStamp = us_cts;
2674                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2675                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2676                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2677                 log.u_bbr.pacing_gain = rack->r_must_retran;
2678                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2679                     &rack->rc_inp->inp_socket->so_rcv,
2680                     &rack->rc_inp->inp_socket->so_snd,
2681                     BBR_LOG_TIMERCANC, 0,
2682                     0, &log, false, tv);
2683         }
2684 }
2685
2686 static void
2687 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2688                           uint32_t flex1, uint32_t flex2,
2689                           uint32_t flex3, uint32_t flex4,
2690                           uint32_t flex5, uint32_t flex6,
2691                           uint16_t flex7, uint8_t mod)
2692 {
2693         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2694                 union tcp_log_stackspecific log;
2695                 struct timeval tv;
2696
2697                 if (mod == 1) {
2698                         /* No you can't use 1, its for the real to cancel */
2699                         return;
2700                 }
2701                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2702                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2703                 log.u_bbr.flex1 = flex1;
2704                 log.u_bbr.flex2 = flex2;
2705                 log.u_bbr.flex3 = flex3;
2706                 log.u_bbr.flex4 = flex4;
2707                 log.u_bbr.flex5 = flex5;
2708                 log.u_bbr.flex6 = flex6;
2709                 log.u_bbr.flex7 = flex7;
2710                 log.u_bbr.flex8 = mod;
2711                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2712                     &rack->rc_inp->inp_socket->so_rcv,
2713                     &rack->rc_inp->inp_socket->so_snd,
2714                     BBR_LOG_TIMERCANC, 0,
2715                     0, &log, false, &tv);
2716         }
2717 }
2718
2719 static void
2720 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2721 {
2722         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2723                 union tcp_log_stackspecific log;
2724                 struct timeval tv;
2725
2726                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2727                 log.u_bbr.flex1 = timers;
2728                 log.u_bbr.flex2 = ret;
2729                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2730                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2731                 log.u_bbr.flex5 = cts;
2732                 if (rack->rack_no_prr)
2733                         log.u_bbr.flex6 = 0;
2734                 else
2735                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2736                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2737                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2738                 log.u_bbr.pacing_gain = rack->r_must_retran;
2739                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2740                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2741                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2742                     &rack->rc_inp->inp_socket->so_rcv,
2743                     &rack->rc_inp->inp_socket->so_snd,
2744                     BBR_LOG_TO_PROCESS, 0,
2745                     0, &log, false, &tv);
2746         }
2747 }
2748
2749 static void
2750 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2751 {
2752         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2753                 union tcp_log_stackspecific log;
2754                 struct timeval tv;
2755
2756                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2757                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2758                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2759                 if (rack->rack_no_prr)
2760                         log.u_bbr.flex3 = 0;
2761                 else
2762                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2763                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2764                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2765                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2766                 log.u_bbr.flex8 = frm;
2767                 log.u_bbr.pkts_out = orig_cwnd;
2768                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2769                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2770                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2771                 log.u_bbr.use_lt_bw <<= 1;
2772                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2773                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2774                     &rack->rc_inp->inp_socket->so_rcv,
2775                     &rack->rc_inp->inp_socket->so_snd,
2776                     BBR_LOG_BBRUPD, 0,
2777                     0, &log, false, &tv);
2778         }
2779 }
2780
2781 #ifdef NETFLIX_EXP_DETECTION
2782 static void
2783 rack_log_sad(struct tcp_rack *rack, int event)
2784 {
2785         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2786                 union tcp_log_stackspecific log;
2787                 struct timeval tv;
2788
2789                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2790                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2791                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2792                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2793                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2794                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2795                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2796                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2797                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2798                 log.u_bbr.lt_epoch |= rack->do_detection;
2799                 log.u_bbr.applimited = tcp_map_minimum;
2800                 log.u_bbr.flex7 = rack->sack_attack_disable;
2801                 log.u_bbr.flex8 = event;
2802                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2803                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2804                 log.u_bbr.delivered = tcp_sad_decay_val;
2805                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2806                     &rack->rc_inp->inp_socket->so_rcv,
2807                     &rack->rc_inp->inp_socket->so_snd,
2808                     TCP_SAD_DETECTION, 0,
2809                     0, &log, false, &tv);
2810         }
2811 }
2812 #endif
2813
2814 static void
2815 rack_counter_destroy(void)
2816 {
2817         int i;
2818
2819         counter_u64_free(rack_fto_send);
2820         counter_u64_free(rack_fto_rsm_send);
2821         counter_u64_free(rack_nfto_resend);
2822         counter_u64_free(rack_hw_pace_init_fail);
2823         counter_u64_free(rack_hw_pace_lost);
2824         counter_u64_free(rack_non_fto_send);
2825         counter_u64_free(rack_extended_rfo);
2826         counter_u64_free(rack_ack_total);
2827         counter_u64_free(rack_express_sack);
2828         counter_u64_free(rack_sack_total);
2829         counter_u64_free(rack_move_none);
2830         counter_u64_free(rack_move_some);
2831         counter_u64_free(rack_sack_attacks_detected);
2832         counter_u64_free(rack_sack_attacks_reversed);
2833         counter_u64_free(rack_sack_used_next_merge);
2834         counter_u64_free(rack_sack_used_prev_merge);
2835         counter_u64_free(rack_badfr);
2836         counter_u64_free(rack_badfr_bytes);
2837         counter_u64_free(rack_rtm_prr_retran);
2838         counter_u64_free(rack_rtm_prr_newdata);
2839         counter_u64_free(rack_timestamp_mismatch);
2840         counter_u64_free(rack_find_high);
2841         counter_u64_free(rack_reorder_seen);
2842         counter_u64_free(rack_tlp_tot);
2843         counter_u64_free(rack_tlp_newdata);
2844         counter_u64_free(rack_tlp_retran);
2845         counter_u64_free(rack_tlp_retran_bytes);
2846         counter_u64_free(rack_tlp_retran_fail);
2847         counter_u64_free(rack_to_tot);
2848         counter_u64_free(rack_to_arm_rack);
2849         counter_u64_free(rack_to_arm_tlp);
2850         counter_u64_free(rack_calc_zero);
2851         counter_u64_free(rack_calc_nonzero);
2852         counter_u64_free(rack_paced_segments);
2853         counter_u64_free(rack_unpaced_segments);
2854         counter_u64_free(rack_saw_enobuf);
2855         counter_u64_free(rack_saw_enobuf_hw);
2856         counter_u64_free(rack_saw_enetunreach);
2857         counter_u64_free(rack_hot_alloc);
2858         counter_u64_free(rack_to_alloc);
2859         counter_u64_free(rack_to_alloc_hard);
2860         counter_u64_free(rack_to_alloc_emerg);
2861         counter_u64_free(rack_to_alloc_limited);
2862         counter_u64_free(rack_alloc_limited_conns);
2863         counter_u64_free(rack_split_limited);
2864         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
2865                 counter_u64_free(rack_proc_comp_ack[i]);
2866         }
2867         counter_u64_free(rack_multi_single_eq);
2868         counter_u64_free(rack_proc_non_comp_ack);
2869         counter_u64_free(rack_sack_proc_all);
2870         counter_u64_free(rack_sack_proc_restart);
2871         counter_u64_free(rack_sack_proc_short);
2872         counter_u64_free(rack_enter_tlp_calc);
2873         counter_u64_free(rack_used_tlpmethod);
2874         counter_u64_free(rack_used_tlpmethod2);
2875         counter_u64_free(rack_sack_skipped_acked);
2876         counter_u64_free(rack_sack_splits);
2877         counter_u64_free(rack_progress_drops);
2878         counter_u64_free(rack_input_idle_reduces);
2879         counter_u64_free(rack_collapsed_win);
2880         counter_u64_free(rack_tlp_does_nada);
2881         counter_u64_free(rack_try_scwnd);
2882         counter_u64_free(rack_per_timer_hole);
2883         counter_u64_free(rack_large_ackcmp);
2884         counter_u64_free(rack_small_ackcmp);
2885 #ifdef INVARIANTS
2886         counter_u64_free(rack_adjust_map_bw);
2887 #endif
2888         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2889         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2890 }
2891
2892 static struct rack_sendmap *
2893 rack_alloc(struct tcp_rack *rack)
2894 {
2895         struct rack_sendmap *rsm;
2896
2897         /*
2898          * First get the top of the list it in
2899          * theory is the "hottest" rsm we have,
2900          * possibly just freed by ack processing.
2901          */
2902         if (rack->rc_free_cnt > rack_free_cache) {
2903                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2904                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2905                 counter_u64_add(rack_hot_alloc, 1);
2906                 rack->rc_free_cnt--;
2907                 return (rsm);
2908         }
2909         /*
2910          * Once we get under our free cache we probably
2911          * no longer have a "hot" one available. Lets
2912          * get one from UMA.
2913          */
2914         rsm = uma_zalloc(rack_zone, M_NOWAIT);
2915         if (rsm) {
2916                 rack->r_ctl.rc_num_maps_alloced++;
2917                 counter_u64_add(rack_to_alloc, 1);
2918                 return (rsm);
2919         }
2920         /*
2921          * Dig in to our aux rsm's (the last two) since
2922          * UMA failed to get us one.
2923          */
2924         if (rack->rc_free_cnt) {
2925                 counter_u64_add(rack_to_alloc_emerg, 1);
2926                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2927                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2928                 rack->rc_free_cnt--;
2929                 return (rsm);
2930         }
2931         return (NULL);
2932 }
2933
2934 static struct rack_sendmap *
2935 rack_alloc_full_limit(struct tcp_rack *rack)
2936 {
2937         if ((V_tcp_map_entries_limit > 0) &&
2938             (rack->do_detection == 0) &&
2939             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2940                 counter_u64_add(rack_to_alloc_limited, 1);
2941                 if (!rack->alloc_limit_reported) {
2942                         rack->alloc_limit_reported = 1;
2943                         counter_u64_add(rack_alloc_limited_conns, 1);
2944                 }
2945                 return (NULL);
2946         }
2947         return (rack_alloc(rack));
2948 }
2949
2950 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2951 static struct rack_sendmap *
2952 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2953 {
2954         struct rack_sendmap *rsm;
2955
2956         if (limit_type) {
2957                 /* currently there is only one limit type */
2958                 if (V_tcp_map_split_limit > 0 &&
2959                     (rack->do_detection == 0) &&
2960                     rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2961                         counter_u64_add(rack_split_limited, 1);
2962                         if (!rack->alloc_limit_reported) {
2963                                 rack->alloc_limit_reported = 1;
2964                                 counter_u64_add(rack_alloc_limited_conns, 1);
2965                         }
2966                         return (NULL);
2967                 }
2968         }
2969
2970         /* allocate and mark in the limit type, if set */
2971         rsm = rack_alloc(rack);
2972         if (rsm != NULL && limit_type) {
2973                 rsm->r_limit_type = limit_type;
2974                 rack->r_ctl.rc_num_split_allocs++;
2975         }
2976         return (rsm);
2977 }
2978
2979 static void
2980 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2981 {
2982         if (rsm->r_flags & RACK_APP_LIMITED) {
2983                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2984                         rack->r_ctl.rc_app_limited_cnt--;
2985                 }
2986         }
2987         if (rsm->r_limit_type) {
2988                 /* currently there is only one limit type */
2989                 rack->r_ctl.rc_num_split_allocs--;
2990         }
2991         if (rsm == rack->r_ctl.rc_first_appl) {
2992                 if (rack->r_ctl.rc_app_limited_cnt == 0)
2993                         rack->r_ctl.rc_first_appl = NULL;
2994                 else {
2995                         /* Follow the next one out */
2996                         struct rack_sendmap fe;
2997
2998                         fe.r_start = rsm->r_nseq_appl;
2999                         rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3000                 }
3001         }
3002         if (rsm == rack->r_ctl.rc_resend)
3003                 rack->r_ctl.rc_resend = NULL;
3004         if (rsm == rack->r_ctl.rc_rsm_at_retran)
3005                 rack->r_ctl.rc_rsm_at_retran = NULL;
3006         if (rsm == rack->r_ctl.rc_end_appl)
3007                 rack->r_ctl.rc_end_appl = NULL;
3008         if (rack->r_ctl.rc_tlpsend == rsm)
3009                 rack->r_ctl.rc_tlpsend = NULL;
3010         if (rack->r_ctl.rc_sacklast == rsm)
3011                 rack->r_ctl.rc_sacklast = NULL;
3012         memset(rsm, 0, sizeof(struct rack_sendmap));
3013         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3014         rack->rc_free_cnt++;
3015 }
3016
3017 static void
3018 rack_free_trim(struct tcp_rack *rack)
3019 {
3020         struct rack_sendmap *rsm;
3021
3022         /*
3023          * Free up all the tail entries until
3024          * we get our list down to the limit.
3025          */
3026         while (rack->rc_free_cnt > rack_free_cache) {
3027                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3028                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3029                 rack->rc_free_cnt--;
3030                 uma_zfree(rack_zone, rsm);
3031         }
3032 }
3033
3034
3035 static uint32_t
3036 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3037 {
3038         uint64_t srtt, bw, len, tim;
3039         uint32_t segsiz, def_len, minl;
3040
3041         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3042         def_len = rack_def_data_window * segsiz;
3043         if (rack->rc_gp_filled == 0) {
3044                 /*
3045                  * We have no measurement (IW is in flight?) so
3046                  * we can only guess using our data_window sysctl
3047                  * value (usually 20MSS).
3048                  */
3049                 return (def_len);
3050         }
3051         /*
3052          * Now we have a number of factors to consider.
3053          *
3054          * 1) We have a desired BDP which is usually
3055          *    at least 2.
3056          * 2) We have a minimum number of rtt's usually 1 SRTT
3057          *    but we allow it too to be more.
3058          * 3) We want to make sure a measurement last N useconds (if
3059          *    we have set rack_min_measure_usec.
3060          *
3061          * We handle the first concern here by trying to create a data
3062          * window of max(rack_def_data_window, DesiredBDP). The
3063          * second concern we handle in not letting the measurement
3064          * window end normally until at least the required SRTT's
3065          * have gone by which is done further below in
3066          * rack_enough_for_measurement(). Finally the third concern
3067          * we also handle here by calculating how long that time
3068          * would take at the current BW and then return the
3069          * max of our first calculation and that length. Note
3070          * that if rack_min_measure_usec is 0, we don't deal
3071          * with concern 3. Also for both Concern 1 and 3 an
3072          * application limited period could end the measurement
3073          * earlier.
3074          *
3075          * So lets calculate the BDP with the "known" b/w using
3076          * the SRTT has our rtt and then multiply it by the
3077          * goal.
3078          */
3079         bw = rack_get_bw(rack);
3080         srtt = (uint64_t)tp->t_srtt;
3081         len = bw * srtt;
3082         len /= (uint64_t)HPTS_USEC_IN_SEC;
3083         len *= max(1, rack_goal_bdp);
3084         /* Now we need to round up to the nearest MSS */
3085         len = roundup(len, segsiz);
3086         if (rack_min_measure_usec) {
3087                 /* Now calculate our min length for this b/w */
3088                 tim = rack_min_measure_usec;
3089                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3090                 if (minl == 0)
3091                         minl = 1;
3092                 minl = roundup(minl, segsiz);
3093                 if (len < minl)
3094                         len = minl;
3095         }
3096         /*
3097          * Now if we have a very small window we want
3098          * to attempt to get the window that is
3099          * as small as possible. This happens on
3100          * low b/w connections and we don't want to
3101          * span huge numbers of rtt's between measurements.
3102          *
3103          * We basically include 2 over our "MIN window" so
3104          * that the measurement can be shortened (possibly) by
3105          * an ack'ed packet.
3106          */
3107         if (len < def_len)
3108                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3109         else
3110                 return (max((uint32_t)len, def_len));
3111
3112 }
3113
3114 static int
3115 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3116 {
3117         uint32_t tim, srtts, segsiz;
3118
3119         /*
3120          * Has enough time passed for the GP measurement to be valid?
3121          */
3122         if ((tp->snd_max == tp->snd_una) ||
3123             (th_ack == tp->snd_max)){
3124                 /* All is acked */
3125                 *quality = RACK_QUALITY_ALLACKED;
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             (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3140                 /*
3141                  * We are up to the app limited send point
3142                  * we have to measure irrespective of the time..
3143                  */
3144                 *quality = RACK_QUALITY_APPLIMITED;
3145                 return (1);
3146         }
3147         /* Now what about time? */
3148         srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3149         tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3150         if (tim >= srtts) {
3151                 *quality = RACK_QUALITY_HIGH;
3152                 return (1);
3153         }
3154         /* Nope not even a full SRTT has passed */
3155         return (0);
3156 }
3157
3158 static void
3159 rack_log_timely(struct tcp_rack *rack,
3160                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3161                 uint64_t up_bnd, int line, uint8_t method)
3162 {
3163         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3164                 union tcp_log_stackspecific log;
3165                 struct timeval tv;
3166
3167                 memset(&log, 0, sizeof(log));
3168                 log.u_bbr.flex1 = logged;
3169                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3170                 log.u_bbr.flex2 <<= 4;
3171                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3172                 log.u_bbr.flex2 <<= 4;
3173                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3174                 log.u_bbr.flex2 <<= 4;
3175                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3176                 log.u_bbr.flex3 = rack->rc_gp_incr;
3177                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3178                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3179                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3180                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3181                 log.u_bbr.flex8 = method;
3182                 log.u_bbr.cur_del_rate = cur_bw;
3183                 log.u_bbr.delRate = low_bnd;
3184                 log.u_bbr.bw_inuse = up_bnd;
3185                 log.u_bbr.rttProp = rack_get_bw(rack);
3186                 log.u_bbr.pkt_epoch = line;
3187                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3188                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3189                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3190                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3191                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3192                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3193                 log.u_bbr.cwnd_gain <<= 1;
3194                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3195                 log.u_bbr.cwnd_gain <<= 1;
3196                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3197                 log.u_bbr.cwnd_gain <<= 1;
3198                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3199                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3200                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3201                     &rack->rc_inp->inp_socket->so_rcv,
3202                     &rack->rc_inp->inp_socket->so_snd,
3203                     TCP_TIMELY_WORK, 0,
3204                     0, &log, false, &tv);
3205         }
3206 }
3207
3208 static int
3209 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3210 {
3211         /*
3212          * Before we increase we need to know if
3213          * the estimate just made was less than
3214          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3215          *
3216          * If we already are pacing at a fast enough
3217          * rate to push us faster there is no sense of
3218          * increasing.
3219          *
3220          * We first caculate our actual pacing rate (ss or ca multipler
3221          * times our cur_bw).
3222          *
3223          * Then we take the last measured rate and multipy by our
3224          * maximum pacing overage to give us a max allowable rate.
3225          *
3226          * If our act_rate is smaller than our max_allowable rate
3227          * then we should increase. Else we should hold steady.
3228          *
3229          */
3230         uint64_t act_rate, max_allow_rate;
3231
3232         if (rack_timely_no_stopping)
3233                 return (1);
3234
3235         if ((cur_bw == 0) || (last_bw_est == 0)) {
3236                 /*
3237                  * Initial startup case or
3238                  * everything is acked case.
3239                  */
3240                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3241                                 __LINE__, 9);
3242                 return (1);
3243         }
3244         if (mult <= 100) {
3245                 /*
3246                  * We can always pace at or slightly above our rate.
3247                  */
3248                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3249                                 __LINE__, 9);
3250                 return (1);
3251         }
3252         act_rate = cur_bw * (uint64_t)mult;
3253         act_rate /= 100;
3254         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3255         max_allow_rate /= 100;
3256         if (act_rate < max_allow_rate) {
3257                 /*
3258                  * Here the rate we are actually pacing at
3259                  * is smaller than 10% above our last measurement.
3260                  * This means we are pacing below what we would
3261                  * like to try to achieve (plus some wiggle room).
3262                  */
3263                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3264                                 __LINE__, 9);
3265                 return (1);
3266         } else {
3267                 /*
3268                  * Here we are already pacing at least rack_max_per_above(10%)
3269                  * what we are getting back. This indicates most likely
3270                  * that we are being limited (cwnd/rwnd/app) and can't
3271                  * get any more b/w. There is no sense of trying to
3272                  * raise up the pacing rate its not speeding us up
3273                  * and we already are pacing faster than we are getting.
3274                  */
3275                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3276                                 __LINE__, 8);
3277                 return (0);
3278         }
3279 }
3280
3281 static void
3282 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3283 {
3284         /*
3285          * When we drag bottom, we want to assure
3286          * that no multiplier is below 1.0, if so
3287          * we want to restore it to at least that.
3288          */
3289         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3290                 /* This is unlikely we usually do not touch recovery */
3291                 rack->r_ctl.rack_per_of_gp_rec = 100;
3292         }
3293         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3294                 rack->r_ctl.rack_per_of_gp_ca = 100;
3295         }
3296         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3297                 rack->r_ctl.rack_per_of_gp_ss = 100;
3298         }
3299 }
3300
3301 static void
3302 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3303 {
3304         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3305                 rack->r_ctl.rack_per_of_gp_ca = 100;
3306         }
3307         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3308                 rack->r_ctl.rack_per_of_gp_ss = 100;
3309         }
3310 }
3311
3312 static void
3313 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3314 {
3315         int32_t  calc, logged, plus;
3316
3317         logged = 0;
3318
3319         if (override) {
3320                 /*
3321                  * override is passed when we are
3322                  * loosing b/w and making one last
3323                  * gasp at trying to not loose out
3324                  * to a new-reno flow.
3325                  */
3326                 goto extra_boost;
3327         }
3328         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3329         if (rack->rc_gp_incr &&
3330             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3331                 /*
3332                  * Reset and get 5 strokes more before the boost. Note
3333                  * that the count is 0 based so we have to add one.
3334                  */
3335 extra_boost:
3336                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3337                 rack->rc_gp_timely_inc_cnt = 0;
3338         } else
3339                 plus = (uint32_t)rack_gp_increase_per;
3340         /* Must be at least 1% increase for true timely increases */
3341         if ((plus < 1) &&
3342             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3343                 plus = 1;
3344         if (rack->rc_gp_saw_rec &&
3345             (rack->rc_gp_no_rec_chg == 0) &&
3346             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3347                                   rack->r_ctl.rack_per_of_gp_rec)) {
3348                 /* We have been in recovery ding it too */
3349                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3350                 if (calc > 0xffff)
3351                         calc = 0xffff;
3352                 logged |= 1;
3353                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3354                 if (rack_per_upper_bound_ss &&
3355                     (rack->rc_dragged_bottom == 0) &&
3356                     (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3357                         rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3358         }
3359         if (rack->rc_gp_saw_ca &&
3360             (rack->rc_gp_saw_ss == 0) &&
3361             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3362                                   rack->r_ctl.rack_per_of_gp_ca)) {
3363                 /* In CA */
3364                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3365                 if (calc > 0xffff)
3366                         calc = 0xffff;
3367                 logged |= 2;
3368                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3369                 if (rack_per_upper_bound_ca &&
3370                     (rack->rc_dragged_bottom == 0) &&
3371                     (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3372                         rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3373         }
3374         if (rack->rc_gp_saw_ss &&
3375             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3376                                   rack->r_ctl.rack_per_of_gp_ss)) {
3377                 /* In SS */
3378                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3379                 if (calc > 0xffff)
3380                         calc = 0xffff;
3381                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3382                 if (rack_per_upper_bound_ss &&
3383                     (rack->rc_dragged_bottom == 0) &&
3384                     (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3385                         rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3386                 logged |= 4;
3387         }
3388         if (logged &&
3389             (rack->rc_gp_incr == 0)){
3390                 /* Go into increment mode */
3391                 rack->rc_gp_incr = 1;
3392                 rack->rc_gp_timely_inc_cnt = 0;
3393         }
3394         if (rack->rc_gp_incr &&
3395             logged &&
3396             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3397                 rack->rc_gp_timely_inc_cnt++;
3398         }
3399         rack_log_timely(rack,  logged, plus, 0, 0,
3400                         __LINE__, 1);
3401 }
3402
3403 static uint32_t
3404 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3405 {
3406         /*
3407          * norm_grad = rtt_diff / minrtt;
3408          * new_per = curper * (1 - B * norm_grad)
3409          *
3410          * B = rack_gp_decrease_per (default 10%)
3411          * rtt_dif = input var current rtt-diff
3412          * curper = input var current percentage
3413          * minrtt = from rack filter
3414          *
3415          */
3416         uint64_t perf;
3417
3418         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3419                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3420                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3421                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3422                      (uint64_t)1000000)) /
3423                 (uint64_t)1000000);
3424         if (perf > curper) {
3425                 /* TSNH */
3426                 perf = curper - 1;
3427         }
3428         return ((uint32_t)perf);
3429 }
3430
3431 static uint32_t
3432 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3433 {
3434         /*
3435          *                                   highrttthresh
3436          * result = curper * (1 - (B * ( 1 -  ------          ))
3437          *                                     gp_srtt
3438          *
3439          * B = rack_gp_decrease_per (default 10%)
3440          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3441          */
3442         uint64_t perf;
3443         uint32_t highrttthresh;
3444
3445         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3446
3447         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3448                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3449                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3450                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3451         return (perf);
3452 }
3453
3454 static void
3455 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3456 {
3457         uint64_t logvar, logvar2, logvar3;
3458         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3459
3460         if (rack->rc_gp_incr) {
3461                 /* Turn off increment counting */
3462                 rack->rc_gp_incr = 0;
3463                 rack->rc_gp_timely_inc_cnt = 0;
3464         }
3465         ss_red = ca_red = rec_red = 0;
3466         logged = 0;
3467         /* Calculate the reduction value */
3468         if (rtt_diff < 0) {
3469                 rtt_diff *= -1;
3470         }
3471         /* Must be at least 1% reduction */
3472         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3473                 /* We have been in recovery ding it too */
3474                 if (timely_says == 2) {
3475                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3476                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3477                         if (alt < new_per)
3478                                 val = alt;
3479                         else
3480                                 val = new_per;
3481                 } else
3482                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3483                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3484                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3485                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3486                 } else {
3487                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3488                         rec_red = 0;
3489                 }
3490                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3491                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3492                 logged |= 1;
3493         }
3494         if (rack->rc_gp_saw_ss) {
3495                 /* Sent in SS */
3496                 if (timely_says == 2) {
3497                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3498                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3499                         if (alt < new_per)
3500                                 val = alt;
3501                         else
3502                                 val = new_per;
3503                 } else
3504                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3505                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3506                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3507                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3508                 } else {
3509                         ss_red = new_per;
3510                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3511                         logvar = new_per;
3512                         logvar <<= 32;
3513                         logvar |= alt;
3514                         logvar2 = (uint32_t)rtt;
3515                         logvar2 <<= 32;
3516                         logvar2 |= (uint32_t)rtt_diff;
3517                         logvar3 = rack_gp_rtt_maxmul;
3518                         logvar3 <<= 32;
3519                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3520                         rack_log_timely(rack, timely_says,
3521                                         logvar2, logvar3,
3522                                         logvar, __LINE__, 10);
3523                 }
3524                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3525                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3526                 logged |= 4;
3527         } else if (rack->rc_gp_saw_ca) {
3528                 /* Sent in CA */
3529                 if (timely_says == 2) {
3530                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3531                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3532                         if (alt < new_per)
3533                                 val = alt;
3534                         else
3535                                 val = new_per;
3536                 } else
3537                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3538                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3539                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3540                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3541                 } else {
3542                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3543                         ca_red = 0;
3544                         logvar = new_per;
3545                         logvar <<= 32;
3546                         logvar |= alt;
3547                         logvar2 = (uint32_t)rtt;
3548                         logvar2 <<= 32;
3549                         logvar2 |= (uint32_t)rtt_diff;
3550                         logvar3 = rack_gp_rtt_maxmul;
3551                         logvar3 <<= 32;
3552                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3553                         rack_log_timely(rack, timely_says,
3554                                         logvar2, logvar3,
3555                                         logvar, __LINE__, 10);
3556                 }
3557                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3558                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3559                 logged |= 2;
3560         }
3561         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3562                 rack->rc_gp_timely_dec_cnt++;
3563                 if (rack_timely_dec_clear &&
3564                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3565                         rack->rc_gp_timely_dec_cnt = 0;
3566         }
3567         logvar = ss_red;
3568         logvar <<= 32;
3569         logvar |= ca_red;
3570         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3571                         __LINE__, 2);
3572 }
3573
3574 static void
3575 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3576                      uint32_t rtt, uint32_t line, uint8_t reas)
3577 {
3578         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3579                 union tcp_log_stackspecific log;
3580                 struct timeval tv;
3581
3582                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3583                 log.u_bbr.flex1 = line;
3584                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3585                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3586                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3587                 log.u_bbr.flex5 = rtt;
3588                 log.u_bbr.flex6 = rack->rc_highly_buffered;
3589                 log.u_bbr.flex6 <<= 1;
3590                 log.u_bbr.flex6 |= rack->forced_ack;
3591                 log.u_bbr.flex6 <<= 1;
3592                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3593                 log.u_bbr.flex6 <<= 1;
3594                 log.u_bbr.flex6 |= rack->in_probe_rtt;
3595                 log.u_bbr.flex6 <<= 1;
3596                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3597                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3598                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3599                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3600                 log.u_bbr.flex8 = reas;
3601                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3602                 log.u_bbr.delRate = rack_get_bw(rack);
3603                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3604                 log.u_bbr.cur_del_rate <<= 32;
3605                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3606                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3607                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3608                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3609                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3610                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3611                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3612                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3613                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3614                 log.u_bbr.rttProp = us_cts;
3615                 log.u_bbr.rttProp <<= 32;
3616                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3617                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3618                     &rack->rc_inp->inp_socket->so_rcv,
3619                     &rack->rc_inp->inp_socket->so_snd,
3620                     BBR_LOG_RTT_SHRINKS, 0,
3621                     0, &log, false, &rack->r_ctl.act_rcv_time);
3622         }
3623 }
3624
3625 static void
3626 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3627 {
3628         uint64_t bwdp;
3629
3630         bwdp = rack_get_bw(rack);
3631         bwdp *= (uint64_t)rtt;
3632         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3633         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3634         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3635                 /*
3636                  * A window protocol must be able to have 4 packets
3637                  * outstanding as the floor in order to function
3638                  * (especially considering delayed ack :D).
3639                  */
3640                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3641         }
3642 }
3643
3644 static void
3645 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3646 {
3647         /**
3648          * ProbeRTT is a bit different in rack_pacing than in
3649          * BBR. It is like BBR in that it uses the lowering of
3650          * the RTT as a signal that we saw something new and
3651          * counts from there for how long between. But it is
3652          * different in that its quite simple. It does not
3653          * play with the cwnd and wait until we get down
3654          * to N segments outstanding and hold that for
3655          * 200ms. Instead it just sets the pacing reduction
3656          * rate to a set percentage (70 by default) and hold
3657          * that for a number of recent GP Srtt's.
3658          */
3659         uint32_t segsiz;
3660
3661         if (rack->rc_gp_dyn_mul == 0)
3662                 return;
3663
3664         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3665                 /* We are idle */
3666                 return;
3667         }
3668         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3669             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3670                 /*
3671                  * Stop the goodput now, the idea here is
3672                  * that future measurements with in_probe_rtt
3673                  * won't register if they are not greater so
3674                  * we want to get what info (if any) is available
3675                  * now.
3676                  */
3677                 rack_do_goodput_measurement(rack->rc_tp, rack,
3678                                             rack->rc_tp->snd_una, __LINE__,
3679                                             RACK_QUALITY_PROBERTT);
3680         }
3681         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3682         rack->r_ctl.rc_time_probertt_entered = us_cts;
3683         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3684                      rack->r_ctl.rc_pace_min_segs);
3685         rack->in_probe_rtt = 1;
3686         rack->measure_saw_probe_rtt = 1;
3687         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3688         rack->r_ctl.rc_time_probertt_starts = 0;
3689         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3690         if (rack_probertt_use_min_rtt_entry)
3691                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3692         else
3693                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3694         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3695                              __LINE__, RACK_RTTS_ENTERPROBE);
3696 }
3697
3698 static void
3699 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3700 {
3701         struct rack_sendmap *rsm;
3702         uint32_t segsiz;
3703
3704         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3705                      rack->r_ctl.rc_pace_min_segs);
3706         rack->in_probe_rtt = 0;
3707         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3708             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3709                 /*
3710                  * Stop the goodput now, the idea here is
3711                  * that future measurements with in_probe_rtt
3712                  * won't register if they are not greater so
3713                  * we want to get what info (if any) is available
3714                  * now.
3715                  */
3716                 rack_do_goodput_measurement(rack->rc_tp, rack,
3717                                             rack->rc_tp->snd_una, __LINE__,
3718                                             RACK_QUALITY_PROBERTT);
3719         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3720                 /*
3721                  * We don't have enough data to make a measurement.
3722                  * So lets just stop and start here after exiting
3723                  * probe-rtt. We probably are not interested in
3724                  * the results anyway.
3725                  */
3726                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3727         }
3728         /*
3729          * Measurements through the current snd_max are going
3730          * to be limited by the slower pacing rate.
3731          *
3732          * We need to mark these as app-limited so we
3733          * don't collapse the b/w.
3734          */
3735         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3736         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3737                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3738                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3739                 else {
3740                         /*
3741                          * Go out to the end app limited and mark
3742                          * this new one as next and move the end_appl up
3743                          * to this guy.
3744                          */
3745                         if (rack->r_ctl.rc_end_appl)
3746                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3747                         rack->r_ctl.rc_end_appl = rsm;
3748                 }
3749                 rsm->r_flags |= RACK_APP_LIMITED;
3750                 rack->r_ctl.rc_app_limited_cnt++;
3751         }
3752         /*
3753          * Now, we need to examine our pacing rate multipliers.
3754          * If its under 100%, we need to kick it back up to
3755          * 100%. We also don't let it be over our "max" above
3756          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3757          * Note setting clamp_atexit_prtt to 0 has the effect
3758          * of setting CA/SS to 100% always at exit (which is
3759          * the default behavior).
3760          */
3761         if (rack_probertt_clear_is) {
3762                 rack->rc_gp_incr = 0;
3763                 rack->rc_gp_bwred = 0;
3764                 rack->rc_gp_timely_inc_cnt = 0;
3765                 rack->rc_gp_timely_dec_cnt = 0;
3766         }
3767         /* Do we do any clamping at exit? */
3768         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3769                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3770                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3771         }
3772         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3773                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3774                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3775         }
3776         /*
3777          * Lets set rtt_diff to 0, so that we will get a "boost"
3778          * after exiting.
3779          */
3780         rack->r_ctl.rc_rtt_diff = 0;
3781
3782         /* Clear all flags so we start fresh */
3783         rack->rc_tp->t_bytes_acked = 0;
3784         rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3785         /*
3786          * If configured to, set the cwnd and ssthresh to
3787          * our targets.
3788          */
3789         if (rack_probe_rtt_sets_cwnd) {
3790                 uint64_t ebdp;
3791                 uint32_t setto;
3792
3793                 /* Set ssthresh so we get into CA once we hit our target */
3794                 if (rack_probertt_use_min_rtt_exit == 1) {
3795                         /* Set to min rtt */
3796                         rack_set_prtt_target(rack, segsiz,
3797                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3798                 } else if (rack_probertt_use_min_rtt_exit == 2) {
3799                         /* Set to current gp rtt */
3800                         rack_set_prtt_target(rack, segsiz,
3801                                              rack->r_ctl.rc_gp_srtt);
3802                 } else if (rack_probertt_use_min_rtt_exit == 3) {
3803                         /* Set to entry gp rtt */
3804                         rack_set_prtt_target(rack, segsiz,
3805                                              rack->r_ctl.rc_entry_gp_rtt);
3806                 } else {
3807                         uint64_t sum;
3808                         uint32_t setval;
3809
3810                         sum = rack->r_ctl.rc_entry_gp_rtt;
3811                         sum *= 10;
3812                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3813                         if (sum >= 20) {
3814                                 /*
3815                                  * A highly buffered path needs
3816                                  * cwnd space for timely to work.
3817                                  * Lets set things up as if
3818                                  * we are heading back here again.
3819                                  */
3820                                 setval = rack->r_ctl.rc_entry_gp_rtt;
3821                         } else if (sum >= 15) {
3822                                 /*
3823                                  * Lets take the smaller of the
3824                                  * two since we are just somewhat
3825                                  * buffered.
3826                                  */
3827                                 setval = rack->r_ctl.rc_gp_srtt;
3828                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3829                                         setval = rack->r_ctl.rc_entry_gp_rtt;
3830                         } else {
3831                                 /*
3832                                  * Here we are not highly buffered
3833                                  * and should pick the min we can to
3834                                  * keep from causing loss.
3835                                  */
3836                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3837                         }
3838                         rack_set_prtt_target(rack, segsiz,
3839                                              setval);
3840                 }
3841                 if (rack_probe_rtt_sets_cwnd > 1) {
3842                         /* There is a percentage here to boost */
3843                         ebdp = rack->r_ctl.rc_target_probertt_flight;
3844                         ebdp *= rack_probe_rtt_sets_cwnd;
3845                         ebdp /= 100;
3846                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3847                 } else
3848                         setto = rack->r_ctl.rc_target_probertt_flight;
3849                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3850                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3851                         /* Enforce a min */
3852                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3853                 }
3854                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3855                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3856         }
3857         rack_log_rtt_shrinks(rack,  us_cts,
3858                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3859                              __LINE__, RACK_RTTS_EXITPROBE);
3860         /* Clear times last so log has all the info */
3861         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3862         rack->r_ctl.rc_time_probertt_entered = us_cts;
3863         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3864         rack->r_ctl.rc_time_of_last_probertt = us_cts;
3865 }
3866
3867 static void
3868 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3869 {
3870         /* Check in on probe-rtt */
3871         if (rack->rc_gp_filled == 0) {
3872                 /* We do not do p-rtt unless we have gp measurements */
3873                 return;
3874         }
3875         if (rack->in_probe_rtt) {
3876                 uint64_t no_overflow;
3877                 uint32_t endtime, must_stay;
3878
3879                 if (rack->r_ctl.rc_went_idle_time &&
3880                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3881                         /*
3882                          * We went idle during prtt, just exit now.
3883                          */
3884                         rack_exit_probertt(rack, us_cts);
3885                 } else if (rack_probe_rtt_safety_val &&
3886                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3887                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3888                         /*
3889                          * Probe RTT safety value triggered!
3890                          */
3891                         rack_log_rtt_shrinks(rack,  us_cts,
3892                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3893                                              __LINE__, RACK_RTTS_SAFETY);
3894                         rack_exit_probertt(rack, us_cts);
3895                 }
3896                 /* Calculate the max we will wait */
3897                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3898                 if (rack->rc_highly_buffered)
3899                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3900                 /* Calculate the min we must wait */
3901                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3902                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3903                     TSTMP_LT(us_cts, endtime)) {
3904                         uint32_t calc;
3905                         /* Do we lower more? */
3906 no_exit:
3907                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3908                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3909                         else
3910                                 calc = 0;
3911                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3912                         if (calc) {
3913                                 /* Maybe */
3914                                 calc *= rack_per_of_gp_probertt_reduce;
3915                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3916                                 /* Limit it too */
3917                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3918                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3919                         }
3920                         /* We must reach target or the time set */
3921                         return;
3922                 }
3923                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3924                         if ((TSTMP_LT(us_cts, must_stay) &&
3925                              rack->rc_highly_buffered) ||
3926                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3927                               rack->r_ctl.rc_target_probertt_flight)) {
3928                                 /* We are not past the must_stay time */
3929                                 goto no_exit;
3930                         }
3931                         rack_log_rtt_shrinks(rack,  us_cts,
3932                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3933                                              __LINE__, RACK_RTTS_REACHTARGET);
3934                         rack->r_ctl.rc_time_probertt_starts = us_cts;
3935                         if (rack->r_ctl.rc_time_probertt_starts == 0)
3936                                 rack->r_ctl.rc_time_probertt_starts = 1;
3937                         /* Restore back to our rate we want to pace at in prtt */
3938                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3939                 }
3940                 /*
3941                  * Setup our end time, some number of gp_srtts plus 200ms.
3942                  */
3943                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3944                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3945                 if (rack_probertt_gpsrtt_cnt_div)
3946                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3947                 else
3948                         endtime = 0;
3949                 endtime += rack_min_probertt_hold;
3950                 endtime += rack->r_ctl.rc_time_probertt_starts;
3951                 if (TSTMP_GEQ(us_cts,  endtime)) {
3952                         /* yes, exit probertt */
3953                         rack_exit_probertt(rack, us_cts);
3954                 }
3955
3956         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3957                 /* Go into probertt, its been too long since we went lower */
3958                 rack_enter_probertt(rack, us_cts);
3959         }
3960 }
3961
3962 static void
3963 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3964                        uint32_t rtt, int32_t rtt_diff)
3965 {
3966         uint64_t cur_bw, up_bnd, low_bnd, subfr;
3967         uint32_t losses;
3968
3969         if ((rack->rc_gp_dyn_mul == 0) ||
3970             (rack->use_fixed_rate) ||
3971             (rack->in_probe_rtt) ||
3972             (rack->rc_always_pace == 0)) {
3973                 /* No dynamic GP multipler in play */
3974                 return;
3975         }
3976         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3977         cur_bw = rack_get_bw(rack);
3978         /* Calculate our up and down range */
3979         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3980         up_bnd /= 100;
3981         up_bnd += rack->r_ctl.last_gp_comp_bw;
3982
3983         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3984         subfr /= 100;
3985         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3986         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3987                 /*
3988                  * This is the case where our RTT is above
3989                  * the max target and we have been configured
3990                  * to just do timely no bonus up stuff in that case.
3991                  *
3992                  * There are two configurations, set to 1, and we
3993                  * just do timely if we are over our max. If its
3994                  * set above 1 then we slam the multipliers down
3995                  * to 100 and then decrement per timely.
3996                  */
3997                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3998                                 __LINE__, 3);
3999                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4000                         rack_validate_multipliers_at_or_below_100(rack);
4001                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4002         } else if ((last_bw_est < low_bnd) && !losses) {
4003                 /*
4004                  * We are decreasing this is a bit complicated this
4005                  * means we are loosing ground. This could be
4006                  * because another flow entered and we are competing
4007                  * for b/w with it. This will push the RTT up which
4008                  * makes timely unusable unless we want to get shoved
4009                  * into a corner and just be backed off (the age
4010                  * old problem with delay based CC).
4011                  *
4012                  * On the other hand if it was a route change we
4013                  * would like to stay somewhat contained and not
4014                  * blow out the buffers.
4015                  */
4016                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4017                                 __LINE__, 3);
4018                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4019                 if (rack->rc_gp_bwred == 0) {
4020                         /* Go into reduction counting */
4021                         rack->rc_gp_bwred = 1;
4022                         rack->rc_gp_timely_dec_cnt = 0;
4023                 }
4024                 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
4025                     (timely_says == 0)) {
4026                         /*
4027                          * Push another time with a faster pacing
4028                          * to try to gain back (we include override to
4029                          * get a full raise factor).
4030                          */
4031                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4032                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4033                             (timely_says == 0) ||
4034                             (rack_down_raise_thresh == 0)) {
4035                                 /*
4036                                  * Do an override up in b/w if we were
4037                                  * below the threshold or if the threshold
4038                                  * is zero we always do the raise.
4039                                  */
4040                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4041                         } else {
4042                                 /* Log it stays the same */
4043                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4044                                                 __LINE__, 11);
4045                         }
4046                         rack->rc_gp_timely_dec_cnt++;
4047                         /* We are not incrementing really no-count */
4048                         rack->rc_gp_incr = 0;
4049                         rack->rc_gp_timely_inc_cnt = 0;
4050                 } else {
4051                         /*
4052                          * Lets just use the RTT
4053                          * information and give up
4054                          * pushing.
4055                          */
4056                         goto use_timely;
4057                 }
4058         } else if ((timely_says != 2) &&
4059                     !losses &&
4060                     (last_bw_est > up_bnd)) {
4061                 /*
4062                  * We are increasing b/w lets keep going, updating
4063                  * our b/w and ignoring any timely input, unless
4064                  * of course we are at our max raise (if there is one).
4065                  */
4066
4067                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4068                                 __LINE__, 3);
4069                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4070                 if (rack->rc_gp_saw_ss &&
4071                     rack_per_upper_bound_ss &&
4072                      (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
4073                             /*
4074                              * In cases where we can't go higher
4075                              * we should just use timely.
4076                              */
4077                             goto use_timely;
4078                 }
4079                 if (rack->rc_gp_saw_ca &&
4080                     rack_per_upper_bound_ca &&
4081                     (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4082                             /*
4083                              * In cases where we can't go higher
4084                              * we should just use timely.
4085                              */
4086                             goto use_timely;
4087                 }
4088                 rack->rc_gp_bwred = 0;
4089                 rack->rc_gp_timely_dec_cnt = 0;
4090                 /* You get a set number of pushes if timely is trying to reduce */
4091                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4092                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4093                 } else {
4094                         /* Log it stays the same */
4095                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4096                             __LINE__, 12);
4097                 }
4098                 return;
4099         } else {
4100                 /*
4101                  * We are staying between the lower and upper range bounds
4102                  * so use timely to decide.
4103                  */
4104                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4105                                 __LINE__, 3);
4106 use_timely:
4107                 if (timely_says) {
4108                         rack->rc_gp_incr = 0;
4109                         rack->rc_gp_timely_inc_cnt = 0;
4110                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4111                             !losses &&
4112                             (last_bw_est < low_bnd)) {
4113                                 /* We are loosing ground */
4114                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4115                                 rack->rc_gp_timely_dec_cnt++;
4116                                 /* We are not incrementing really no-count */
4117                                 rack->rc_gp_incr = 0;
4118                                 rack->rc_gp_timely_inc_cnt = 0;
4119                         } else
4120                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4121                 } else {
4122                         rack->rc_gp_bwred = 0;
4123                         rack->rc_gp_timely_dec_cnt = 0;
4124                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4125                 }
4126         }
4127 }
4128
4129 static int32_t
4130 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4131 {
4132         int32_t timely_says;
4133         uint64_t log_mult, log_rtt_a_diff;
4134
4135         log_rtt_a_diff = rtt;
4136         log_rtt_a_diff <<= 32;
4137         log_rtt_a_diff |= (uint32_t)rtt_diff;
4138         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4139                     rack_gp_rtt_maxmul)) {
4140                 /* Reduce the b/w multipler */
4141                 timely_says = 2;
4142                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4143                 log_mult <<= 32;
4144                 log_mult |= prev_rtt;
4145                 rack_log_timely(rack,  timely_says, log_mult,
4146                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4147                                 log_rtt_a_diff, __LINE__, 4);
4148         } else if (rtt <= (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                 /* Increase the b/w multipler */
4152                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4153                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4154                          max(rack_gp_rtt_mindiv , 1));
4155                 log_mult <<= 32;
4156                 log_mult |= prev_rtt;
4157                 timely_says = 0;
4158                 rack_log_timely(rack,  timely_says, log_mult ,
4159                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4160                                 log_rtt_a_diff, __LINE__, 5);
4161         } else {
4162                 /*
4163                  * Use a gradient to find it the timely gradient
4164                  * is:
4165                  * grad = rc_rtt_diff / min_rtt;
4166                  *
4167                  * anything below or equal to 0 will be
4168                  * a increase indication. Anything above
4169                  * zero is a decrease. Note we take care
4170                  * of the actual gradient calculation
4171                  * in the reduction (its not needed for
4172                  * increase).
4173                  */
4174                 log_mult = prev_rtt;
4175                 if (rtt_diff <= 0) {
4176                         /*
4177                          * Rttdiff is less than zero, increase the
4178                          * b/w multipler (its 0 or negative)
4179                          */
4180                         timely_says = 0;
4181                         rack_log_timely(rack,  timely_says, log_mult,
4182                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4183                 } else {
4184                         /* Reduce the b/w multipler */
4185                         timely_says = 1;
4186                         rack_log_timely(rack,  timely_says, log_mult,
4187                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4188                 }
4189         }
4190         return (timely_says);
4191 }
4192
4193 static void
4194 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4195                             tcp_seq th_ack, int line, uint8_t quality)
4196 {
4197         uint64_t tim, bytes_ps, ltim, stim, utim;
4198         uint32_t segsiz, bytes, reqbytes, us_cts;
4199         int32_t gput, new_rtt_diff, timely_says;
4200         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4201         int did_add = 0;
4202
4203         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4204         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4205         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4206                 tim = us_cts - tp->gput_ts;
4207         else
4208                 tim = 0;
4209         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4210                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4211         else
4212                 stim = 0;
4213         /*
4214          * Use the larger of the send time or ack time. This prevents us
4215          * from being influenced by ack artifacts to come up with too
4216          * high of measurement. Note that since we are spanning over many more
4217          * bytes in most of our measurements hopefully that is less likely to
4218          * occur.
4219          */
4220         if (tim > stim)
4221                 utim = max(tim, 1);
4222         else
4223                 utim = max(stim, 1);
4224         /* Lets get a msec time ltim too for the old stuff */
4225         ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4226         gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4227         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4228         if ((tim == 0) && (stim == 0)) {
4229                 /*
4230                  * Invalid measurement time, maybe
4231                  * all on one ack/one send?
4232                  */
4233                 bytes = 0;
4234                 bytes_ps = 0;
4235                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4236                                            0, 0, 0, 10, __LINE__, NULL, quality);
4237                 goto skip_measurement;
4238         }
4239         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4240                 /* We never made a us_rtt measurement? */
4241                 bytes = 0;
4242                 bytes_ps = 0;
4243                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4244                                            0, 0, 0, 10, __LINE__, NULL, quality);
4245                 goto skip_measurement;
4246         }
4247         /*
4248          * Calculate the maximum possible b/w this connection
4249          * could have. We base our calculation on the lowest
4250          * rtt we have seen during the measurement and the
4251          * largest rwnd the client has given us in that time. This
4252          * forms a BDP that is the maximum that we could ever
4253          * get to the client. Anything larger is not valid.
4254          *
4255          * I originally had code here that rejected measurements
4256          * where the time was less than 1/2 the latest us_rtt.
4257          * But after thinking on that I realized its wrong since
4258          * say you had a 150Mbps or even 1Gbps link, and you
4259          * were a long way away.. example I am in Europe (100ms rtt)
4260          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4261          * bytes my time would be 1.2ms, and yet my rtt would say
4262          * the measurement was invalid the time was < 50ms. The
4263          * same thing is true for 150Mb (8ms of time).
4264          *
4265          * A better way I realized is to look at what the maximum
4266          * the connection could possibly do. This is gated on
4267          * the lowest RTT we have seen and the highest rwnd.
4268          * We should in theory never exceed that, if we are
4269          * then something on the path is storing up packets
4270          * and then feeding them all at once to our endpoint
4271          * messing up our measurement.
4272          */
4273         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4274         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4275         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4276         if (SEQ_LT(th_ack, tp->gput_seq)) {
4277                 /* No measurement can be made */
4278                 bytes = 0;
4279                 bytes_ps = 0;
4280                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4281                                            0, 0, 0, 10, __LINE__, NULL, quality);
4282                 goto skip_measurement;
4283         } else
4284                 bytes = (th_ack - tp->gput_seq);
4285         bytes_ps = (uint64_t)bytes;
4286         /*
4287          * Don't measure a b/w for pacing unless we have gotten at least
4288          * an initial windows worth of data in this measurement interval.
4289          *
4290          * Small numbers of bytes get badly influenced by delayed ack and
4291          * other artifacts. Note we take the initial window or our
4292          * defined minimum GP (defaulting to 10 which hopefully is the
4293          * IW).
4294          */
4295         if (rack->rc_gp_filled == 0) {
4296                 /*
4297                  * The initial estimate is special. We
4298                  * have blasted out an IW worth of packets
4299                  * without a real valid ack ts results. We
4300                  * then setup the app_limited_needs_set flag,
4301                  * this should get the first ack in (probably 2
4302                  * MSS worth) to be recorded as the timestamp.
4303                  * We thus allow a smaller number of bytes i.e.
4304                  * IW - 2MSS.
4305                  */
4306                 reqbytes -= (2 * segsiz);
4307                 /* Also lets fill previous for our first measurement to be neutral */
4308                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4309         }
4310         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4311                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4312                                            rack->r_ctl.rc_app_limited_cnt,
4313                                            0, 0, 10, __LINE__, NULL, quality);
4314                 goto skip_measurement;
4315         }
4316         /*
4317          * We now need to calculate the Timely like status so
4318          * we can update (possibly) the b/w multipliers.
4319          */
4320         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4321         if (rack->rc_gp_filled == 0) {
4322                 /* No previous reading */
4323                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4324         } else {
4325                 if (rack->measure_saw_probe_rtt == 0) {
4326                         /*
4327                          * We don't want a probertt to be counted
4328                          * since it will be negative incorrectly. We
4329                          * expect to be reducing the RTT when we
4330                          * pace at a slower rate.
4331                          */
4332                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4333                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4334                 }
4335         }
4336         timely_says = rack_make_timely_judgement(rack,
4337                 rack->r_ctl.rc_gp_srtt,
4338                 rack->r_ctl.rc_rtt_diff,
4339                 rack->r_ctl.rc_prev_gp_srtt
4340                 );
4341         bytes_ps *= HPTS_USEC_IN_SEC;
4342         bytes_ps /= utim;
4343         if (bytes_ps > rack->r_ctl.last_max_bw) {
4344                 /*
4345                  * Something is on path playing
4346                  * since this b/w is not possible based
4347                  * on our BDP (highest rwnd and lowest rtt
4348                  * we saw in the measurement window).
4349                  *
4350                  * Another option here would be to
4351                  * instead skip the measurement.
4352                  */
4353                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4354                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4355                                            11, __LINE__, NULL, quality);
4356                 bytes_ps = rack->r_ctl.last_max_bw;
4357         }
4358         /* We store gp for b/w in bytes per second */
4359         if (rack->rc_gp_filled == 0) {
4360                 /* Initial measurment */
4361                 if (bytes_ps) {
4362                         rack->r_ctl.gp_bw = bytes_ps;
4363                         rack->rc_gp_filled = 1;
4364                         rack->r_ctl.num_measurements = 1;
4365                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4366                 } else {
4367                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4368                                                    rack->r_ctl.rc_app_limited_cnt,
4369                                                    0, 0, 10, __LINE__, NULL, quality);
4370                 }
4371                 if (rack->rc_inp->inp_in_hpts &&
4372                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4373                         /*
4374                          * Ok we can't trust the pacer in this case
4375                          * where we transition from un-paced to paced.
4376                          * Or for that matter when the burst mitigation
4377                          * was making a wild guess and got it wrong.
4378                          * Stop the pacer and clear up all the aggregate
4379                          * delays etc.
4380                          */
4381                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
4382                         rack->r_ctl.rc_hpts_flags = 0;
4383                         rack->r_ctl.rc_last_output_to = 0;
4384                 }
4385                 did_add = 2;
4386         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4387                 /* Still a small number run an average */
4388                 rack->r_ctl.gp_bw += bytes_ps;
4389                 addpart = rack->r_ctl.num_measurements;
4390                 rack->r_ctl.num_measurements++;
4391                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4392                         /* We have collected enought to move forward */
4393                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4394                 }
4395                 did_add = 3;
4396         } else {
4397                 /*
4398                  * We want to take 1/wma of the goodput and add in to 7/8th
4399                  * of the old value weighted by the srtt. So if your measurement
4400                  * period is say 2 SRTT's long you would get 1/4 as the
4401                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4402                  *
4403                  * But we must be careful not to take too much i.e. if the
4404                  * srtt is say 20ms and the measurement is taken over
4405                  * 400ms our weight would be 400/20 i.e. 20. On the
4406                  * other hand if we get a measurement over 1ms with a
4407                  * 10ms rtt we only want to take a much smaller portion.
4408                  */
4409                 if (rack->r_ctl.num_measurements < 0xff) {
4410                         rack->r_ctl.num_measurements++;
4411                 }
4412                 srtt = (uint64_t)tp->t_srtt;
4413                 if (srtt == 0) {
4414                         /*
4415                          * Strange why did t_srtt go back to zero?
4416                          */
4417                         if (rack->r_ctl.rc_rack_min_rtt)
4418                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4419                         else
4420                                 srtt = HPTS_USEC_IN_MSEC;
4421                 }
4422                 /*
4423                  * XXXrrs: Note for reviewers, in playing with
4424                  * dynamic pacing I discovered this GP calculation
4425                  * as done originally leads to some undesired results.
4426                  * Basically you can get longer measurements contributing
4427                  * too much to the WMA. Thus I changed it if you are doing
4428                  * dynamic adjustments to only do the aportioned adjustment
4429                  * if we have a very small (time wise) measurement. Longer
4430                  * measurements just get there weight (defaulting to 1/8)
4431                  * add to the WMA. We may want to think about changing
4432                  * this to always do that for both sides i.e. dynamic
4433                  * and non-dynamic... but considering lots of folks
4434                  * were playing with this I did not want to change the
4435                  * calculation per.se. without your thoughts.. Lawerence?
4436                  * Peter??
4437                  */
4438                 if (rack->rc_gp_dyn_mul == 0) {
4439                         subpart = rack->r_ctl.gp_bw * utim;
4440                         subpart /= (srtt * 8);
4441                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4442                                 /*
4443                                  * The b/w update takes no more
4444                                  * away then 1/2 our running total
4445                                  * so factor it in.
4446                                  */
4447                                 addpart = bytes_ps * utim;
4448                                 addpart /= (srtt * 8);
4449                         } else {
4450                                 /*
4451                                  * Don't allow a single measurement
4452                                  * to account for more than 1/2 of the
4453                                  * WMA. This could happen on a retransmission
4454                                  * where utim becomes huge compared to
4455                                  * srtt (multiple retransmissions when using
4456                                  * the sending rate which factors in all the
4457                                  * transmissions from the first one).
4458                                  */
4459                                 subpart = rack->r_ctl.gp_bw / 2;
4460                                 addpart = bytes_ps / 2;
4461                         }
4462                         resid_bw = rack->r_ctl.gp_bw - subpart;
4463                         rack->r_ctl.gp_bw = resid_bw + addpart;
4464                         did_add = 1;
4465                 } else {
4466                         if ((utim / srtt) <= 1) {
4467                                 /*
4468                                  * The b/w update was over a small period
4469                                  * of time. The idea here is to prevent a small
4470                                  * measurement time period from counting
4471                                  * too much. So we scale it based on the
4472                                  * time so it attributes less than 1/rack_wma_divisor
4473                                  * of its measurement.
4474                                  */
4475                                 subpart = rack->r_ctl.gp_bw * utim;
4476                                 subpart /= (srtt * rack_wma_divisor);
4477                                 addpart = bytes_ps * utim;
4478                                 addpart /= (srtt * rack_wma_divisor);
4479                         } else {
4480                                 /*
4481                                  * The scaled measurement was long
4482                                  * enough so lets just add in the
4483                                  * portion of the measurment i.e. 1/rack_wma_divisor
4484                                  */
4485                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4486                                 addpart = bytes_ps / rack_wma_divisor;
4487                         }
4488                         if ((rack->measure_saw_probe_rtt == 0) ||
4489                             (bytes_ps > rack->r_ctl.gp_bw)) {
4490                                 /*
4491                                  * For probe-rtt we only add it in
4492                                  * if its larger, all others we just
4493                                  * add in.
4494                                  */
4495                                 did_add = 1;
4496                                 resid_bw = rack->r_ctl.gp_bw - subpart;
4497                                 rack->r_ctl.gp_bw = resid_bw + addpart;
4498                         }
4499                 }
4500         }
4501         if ((rack->gp_ready == 0) &&
4502             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4503                 /* We have enough measurements now */
4504                 rack->gp_ready = 1;
4505                 rack_set_cc_pacing(rack);
4506                 if (rack->defer_options)
4507                         rack_apply_deferred_options(rack);
4508         }
4509         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4510                                    rack_get_bw(rack), 22, did_add, NULL, quality);
4511         /* We do not update any multipliers if we are in or have seen a probe-rtt */
4512         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4513                 rack_update_multiplier(rack, timely_says, bytes_ps,
4514                                        rack->r_ctl.rc_gp_srtt,
4515                                        rack->r_ctl.rc_rtt_diff);
4516         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4517                                    rack_get_bw(rack), 3, line, NULL, quality);
4518         /* reset the gp srtt and setup the new prev */
4519         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4520         /* Record the lost count for the next measurement */
4521         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4522         /*
4523          * We restart our diffs based on the gpsrtt in the
4524          * measurement window.
4525          */
4526         rack->rc_gp_rtt_set = 0;
4527         rack->rc_gp_saw_rec = 0;
4528         rack->rc_gp_saw_ca = 0;
4529         rack->rc_gp_saw_ss = 0;
4530         rack->rc_dragged_bottom = 0;
4531 skip_measurement:
4532
4533 #ifdef STATS
4534         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4535                                  gput);
4536         /*
4537          * XXXLAS: This is a temporary hack, and should be
4538          * chained off VOI_TCP_GPUT when stats(9) grows an
4539          * API to deal with chained VOIs.
4540          */
4541         if (tp->t_stats_gput_prev > 0)
4542                 stats_voi_update_abs_s32(tp->t_stats,
4543                                          VOI_TCP_GPUT_ND,
4544                                          ((gput - tp->t_stats_gput_prev) * 100) /
4545                                          tp->t_stats_gput_prev);
4546 #endif
4547         tp->t_flags &= ~TF_GPUTINPROG;
4548         tp->t_stats_gput_prev = gput;
4549         /*
4550          * Now are we app limited now and there is space from where we
4551          * were to where we want to go?
4552          *
4553          * We don't do the other case i.e. non-applimited here since
4554          * the next send will trigger us picking up the missing data.
4555          */
4556         if (rack->r_ctl.rc_first_appl &&
4557             TCPS_HAVEESTABLISHED(tp->t_state) &&
4558             rack->r_ctl.rc_app_limited_cnt &&
4559             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4560             ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4561              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4562                 /*
4563                  * Yep there is enough outstanding to make a measurement here.
4564                  */
4565                 struct rack_sendmap *rsm, fe;
4566
4567                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4568                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4569                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4570                 rack->app_limited_needs_set = 0;
4571                 tp->gput_seq = th_ack;
4572                 if (rack->in_probe_rtt)
4573                         rack->measure_saw_probe_rtt = 1;
4574                 else if ((rack->measure_saw_probe_rtt) &&
4575                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4576                         rack->measure_saw_probe_rtt = 0;
4577                 if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4578                         /* There is a full window to gain info from */
4579                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4580                 } else {
4581                         /* We can only measure up to the applimited point */
4582                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4583                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4584                                 /*
4585                                  * We don't have enough to make a measurement.
4586                                  */
4587                                 tp->t_flags &= ~TF_GPUTINPROG;
4588                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4589                                                            0, 0, 0, 6, __LINE__, NULL, quality);
4590                                 return;
4591                         }
4592                 }
4593                 if (tp->t_state >= TCPS_FIN_WAIT_1) {
4594                         /*
4595                          * We will get no more data into the SB
4596                          * this means we need to have the data available
4597                          * before we start a measurement.
4598                          */
4599                         if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4600                                 /* Nope not enough data. */
4601                                 return;
4602                         }
4603                 }
4604                 tp->t_flags |= TF_GPUTINPROG;
4605                 /*
4606                  * Now we need to find the timestamp of the send at tp->gput_seq
4607                  * for the send based measurement.
4608                  */
4609                 fe.r_start = tp->gput_seq;
4610                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4611                 if (rsm) {
4612                         /* Ok send-based limit is set */
4613                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4614                                 /*
4615                                  * Move back to include the earlier part
4616                                  * so our ack time lines up right (this may
4617                                  * make an overlapping measurement but thats
4618                                  * ok).
4619                                  */
4620                                 tp->gput_seq = rsm->r_start;
4621                         }
4622                         if (rsm->r_flags & RACK_ACKED)
4623                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4624                         else
4625                                 rack->app_limited_needs_set = 1;
4626                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4627                 } else {
4628                         /*
4629                          * If we don't find the rsm due to some
4630                          * send-limit set the current time, which
4631                          * basically disables the send-limit.
4632                          */
4633                         struct timeval tv;
4634
4635                         microuptime(&tv);
4636                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4637                 }
4638                 rack_log_pacing_delay_calc(rack,
4639                                            tp->gput_seq,
4640                                            tp->gput_ack,
4641                                            (uint64_t)rsm,
4642                                            tp->gput_ts,
4643                                            rack->r_ctl.rc_app_limited_cnt,
4644                                            9,
4645                                            __LINE__, NULL, quality);
4646         }
4647 }
4648
4649 /*
4650  * CC wrapper hook functions
4651  */
4652 static void
4653 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4654     uint16_t type, int32_t recovery)
4655 {
4656         uint32_t prior_cwnd, acked;
4657         struct tcp_log_buffer *lgb = NULL;
4658         uint8_t labc_to_use, quality;
4659
4660         INP_WLOCK_ASSERT(tp->t_inpcb);
4661         tp->ccv->nsegs = nsegs;
4662         acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4663         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4664                 uint32_t max;
4665
4666                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4667                 if (tp->ccv->bytes_this_ack > max) {
4668                         tp->ccv->bytes_this_ack = max;
4669                 }
4670         }
4671 #ifdef STATS
4672         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4673             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4674 #endif
4675         quality = RACK_QUALITY_NONE;
4676         if ((tp->t_flags & TF_GPUTINPROG) &&
4677             rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4678                 /* Measure the Goodput */
4679                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4680 #ifdef NETFLIX_PEAKRATE
4681                 if ((type == CC_ACK) &&
4682                     (tp->t_maxpeakrate)) {
4683                         /*
4684                          * We update t_peakrate_thr. This gives us roughly
4685                          * one update per round trip time. Note
4686                          * it will only be used if pace_always is off i.e
4687                          * we don't do this for paced flows.
4688                          */
4689                         rack_update_peakrate_thr(tp);
4690                 }
4691 #endif
4692         }
4693         /* Which way our we limited, if not cwnd limited no advance in CA */
4694         if (tp->snd_cwnd <= tp->snd_wnd)
4695                 tp->ccv->flags |= CCF_CWND_LIMITED;
4696         else
4697                 tp->ccv->flags &= ~CCF_CWND_LIMITED;
4698         if (tp->snd_cwnd > tp->snd_ssthresh) {
4699                 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4700                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4701                 /* For the setting of a window past use the actual scwnd we are using */
4702                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4703                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4704                         tp->ccv->flags |= CCF_ABC_SENTAWND;
4705                 }
4706         } else {
4707                 tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4708                 tp->t_bytes_acked = 0;
4709         }
4710         prior_cwnd = tp->snd_cwnd;
4711         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4712             (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4713                 labc_to_use = rack->rc_labc;
4714         else
4715                 labc_to_use = rack_max_abc_post_recovery;
4716         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4717                 union tcp_log_stackspecific log;
4718                 struct timeval tv;
4719
4720                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4721                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4722                 log.u_bbr.flex1 = th_ack;
4723                 log.u_bbr.flex2 = tp->ccv->flags;
4724                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4725                 log.u_bbr.flex4 = tp->ccv->nsegs;
4726                 log.u_bbr.flex5 = labc_to_use;
4727                 log.u_bbr.flex6 = prior_cwnd;
4728                 log.u_bbr.flex7 = V_tcp_do_newsack;
4729                 log.u_bbr.flex8 = 1;
4730                 lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4731                                      0, &log, false, NULL, NULL, 0, &tv);
4732         }
4733         if (CC_ALGO(tp)->ack_received != NULL) {
4734                 /* XXXLAS: Find a way to live without this */
4735                 tp->ccv->curack = th_ack;
4736                 tp->ccv->labc = labc_to_use;
4737                 tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4738                 CC_ALGO(tp)->ack_received(tp->ccv, type);
4739         }
4740         if (lgb) {
4741                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4742         }
4743         if (rack->r_must_retran) {
4744                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4745                         /*
4746                          * We now are beyond the rxt point so lets disable
4747                          * the flag.
4748                          */
4749                         rack->r_ctl.rc_out_at_rto = 0;
4750                         rack->r_must_retran = 0;
4751                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4752                         /*
4753                          * Only decrement the rc_out_at_rto if the cwnd advances
4754                          * at least a whole segment. Otherwise next time the peer
4755                          * acks, we won't be able to send this generaly happens
4756                          * when we are in Congestion Avoidance.
4757                          */
4758                         if (acked <= rack->r_ctl.rc_out_at_rto){
4759                                 rack->r_ctl.rc_out_at_rto -= acked;
4760                         } else {
4761                                 rack->r_ctl.rc_out_at_rto = 0;
4762                         }
4763                 }
4764         }
4765 #ifdef STATS
4766         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4767 #endif
4768         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4769                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4770         }
4771 #ifdef NETFLIX_PEAKRATE
4772         /* we enforce max peak rate if it is set and we are not pacing */
4773         if ((rack->rc_always_pace == 0) &&
4774             tp->t_peakrate_thr &&
4775             (tp->snd_cwnd > tp->t_peakrate_thr)) {
4776                 tp->snd_cwnd = tp->t_peakrate_thr;
4777         }
4778 #endif
4779 }
4780
4781 static void
4782 tcp_rack_partialack(struct tcpcb *tp)
4783 {
4784         struct tcp_rack *rack;
4785
4786         rack = (struct tcp_rack *)tp->t_fb_ptr;
4787         INP_WLOCK_ASSERT(tp->t_inpcb);
4788         /*
4789          * If we are doing PRR and have enough
4790          * room to send <or> we are pacing and prr
4791          * is disabled we will want to see if we
4792          * can send data (by setting r_wanted_output to
4793          * true).
4794          */
4795         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4796             rack->rack_no_prr)
4797                 rack->r_wanted_output = 1;
4798 }
4799
4800 static void
4801 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4802 {
4803         struct tcp_rack *rack;
4804         uint32_t orig_cwnd;
4805
4806         orig_cwnd = tp->snd_cwnd;
4807         INP_WLOCK_ASSERT(tp->t_inpcb);
4808         rack = (struct tcp_rack *)tp->t_fb_ptr;
4809         /* only alert CC if we alerted when we entered */
4810         if (CC_ALGO(tp)->post_recovery != NULL) {
4811                 tp->ccv->curack = th_ack;
4812                 CC_ALGO(tp)->post_recovery(tp->ccv);
4813                 if (tp->snd_cwnd < tp->snd_ssthresh) {
4814                         /*
4815                          * Rack has burst control and pacing
4816                          * so lets not set this any lower than
4817                          * snd_ssthresh per RFC-6582 (option 2).
4818                          */
4819                         tp->snd_cwnd = tp->snd_ssthresh;
4820                 }
4821         }
4822         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4823                 union tcp_log_stackspecific log;
4824                 struct timeval tv;
4825
4826                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4827                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4828                 log.u_bbr.flex1 = th_ack;
4829                 log.u_bbr.flex2 = tp->ccv->flags;
4830                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4831                 log.u_bbr.flex4 = tp->ccv->nsegs;
4832                 log.u_bbr.flex5 = V_tcp_abc_l_var;
4833                 log.u_bbr.flex6 = orig_cwnd;
4834                 log.u_bbr.flex7 = V_tcp_do_newsack;
4835                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4836                 log.u_bbr.flex8 = 2;
4837                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4838                                0, &log, false, NULL, NULL, 0, &tv);
4839         }
4840         if ((rack->rack_no_prr == 0) &&
4841             (rack->no_prr_addback == 0) &&
4842             (rack->r_ctl.rc_prr_sndcnt > 0)) {
4843                 /*
4844                  * Suck the next prr cnt back into cwnd, but
4845                  * only do that if we are not application limited.
4846                  */
4847                 if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4848                         /*
4849                          * We are allowed to add back to the cwnd the amount we did
4850                          * not get out if:
4851                          * a) no_prr_addback is off.
4852                          * b) we are not app limited
4853                          * c) we are doing prr
4854                          * <and>
4855                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4856                          */
4857                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4858                                             rack->r_ctl.rc_prr_sndcnt);
4859                 }
4860                 rack->r_ctl.rc_prr_sndcnt = 0;
4861                 rack_log_to_prr(rack, 1, 0);
4862         }
4863         rack_log_to_prr(rack, 14, orig_cwnd);
4864         tp->snd_recover = tp->snd_una;
4865         EXIT_RECOVERY(tp->t_flags);
4866 }
4867
4868 static void
4869 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack)
4870 {
4871         struct tcp_rack *rack;
4872         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4873
4874         INP_WLOCK_ASSERT(tp->t_inpcb);
4875 #ifdef STATS
4876         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4877 #endif
4878         if (IN_RECOVERY(tp->t_flags) == 0) {
4879                 in_rec_at_entry = 0;
4880                 ssthresh_enter = tp->snd_ssthresh;
4881                 cwnd_enter = tp->snd_cwnd;
4882         } else
4883                 in_rec_at_entry = 1;
4884         rack = (struct tcp_rack *)tp->t_fb_ptr;
4885         switch (type) {
4886         case CC_NDUPACK:
4887                 tp->t_flags &= ~TF_WASFRECOVERY;
4888                 tp->t_flags &= ~TF_WASCRECOVERY;
4889                 if (!IN_FASTRECOVERY(tp->t_flags)) {
4890                         rack->r_ctl.rc_prr_delivered = 0;
4891                         rack->r_ctl.rc_prr_out = 0;
4892                         if (rack->rack_no_prr == 0) {
4893                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4894                                 rack_log_to_prr(rack, 2, in_rec_at_entry);
4895                         }
4896                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4897                         tp->snd_recover = tp->snd_max;
4898                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4899                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4900                 }
4901                 break;
4902         case CC_ECN:
4903                 if (!IN_CONGRECOVERY(tp->t_flags) ||
4904                     /*
4905                      * Allow ECN reaction on ACK to CWR, if
4906                      * that data segment was also CE marked.
4907                      */
4908                     SEQ_GEQ(ack, tp->snd_recover)) {
4909                         EXIT_CONGRECOVERY(tp->t_flags);
4910                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4911                         tp->snd_recover = tp->snd_max + 1;
4912                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4913                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4914                 }
4915                 break;
4916         case CC_RTO:
4917                 tp->t_dupacks = 0;
4918                 tp->t_bytes_acked = 0;
4919                 EXIT_RECOVERY(tp->t_flags);
4920                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4921                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4922                 orig_cwnd = tp->snd_cwnd;
4923                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4924                 rack_log_to_prr(rack, 16, orig_cwnd);
4925                 if (tp->t_flags2 & TF2_ECN_PERMIT)
4926                         tp->t_flags2 |= TF2_ECN_SND_CWR;
4927                 break;
4928         case CC_RTO_ERR:
4929                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4930                 /* RTO was unnecessary, so reset everything. */
4931                 tp->snd_cwnd = tp->snd_cwnd_prev;
4932                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4933                 tp->snd_recover = tp->snd_recover_prev;
4934                 if (tp->t_flags & TF_WASFRECOVERY) {
4935                         ENTER_FASTRECOVERY(tp->t_flags);
4936                         tp->t_flags &= ~TF_WASFRECOVERY;
4937                 }
4938                 if (tp->t_flags & TF_WASCRECOVERY) {
4939                         ENTER_CONGRECOVERY(tp->t_flags);
4940                         tp->t_flags &= ~TF_WASCRECOVERY;
4941                 }
4942                 tp->snd_nxt = tp->snd_max;
4943                 tp->t_badrxtwin = 0;
4944                 break;
4945         }
4946         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4947             (type != CC_RTO)){
4948                 tp->ccv->curack = ack;
4949                 CC_ALGO(tp)->cong_signal(tp->ccv, type);
4950         }
4951         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4952                 rack_log_to_prr(rack, 15, cwnd_enter);
4953                 rack->r_ctl.dsack_byte_cnt = 0;
4954                 rack->r_ctl.retran_during_recovery = 0;
4955                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4956                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4957                 rack->r_ent_rec_ns = 1;
4958         }
4959 }
4960
4961 static inline void
4962 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4963 {
4964         uint32_t i_cwnd;
4965
4966         INP_WLOCK_ASSERT(tp->t_inpcb);
4967
4968 #ifdef NETFLIX_STATS
4969         KMOD_TCPSTAT_INC(tcps_idle_restarts);
4970         if (tp->t_state == TCPS_ESTABLISHED)
4971                 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4972 #endif
4973         if (CC_ALGO(tp)->after_idle != NULL)
4974                 CC_ALGO(tp)->after_idle(tp->ccv);
4975
4976         if (tp->snd_cwnd == 1)
4977                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
4978         else
4979                 i_cwnd = rc_init_window(rack);
4980
4981         /*
4982          * Being idle is no differnt than the initial window. If the cc
4983          * clamps it down below the initial window raise it to the initial
4984          * window.
4985          */
4986         if (tp->snd_cwnd < i_cwnd) {
4987                 tp->snd_cwnd = i_cwnd;
4988         }
4989 }
4990
4991 /*
4992  * Indicate whether this ack should be delayed.  We can delay the ack if
4993  * following conditions are met:
4994  *      - There is no delayed ack timer in progress.
4995  *      - Our last ack wasn't a 0-sized window. We never want to delay
4996  *        the ack that opens up a 0-sized window.
4997  *      - LRO wasn't used for this segment. We make sure by checking that the
4998  *        segment size is not larger than the MSS.
4999  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
5000  *        connection.
5001  */
5002 #define DELAY_ACK(tp, tlen)                      \
5003         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
5004         ((tp->t_flags & TF_DELACK) == 0) &&      \
5005         (tlen <= tp->t_maxseg) &&                \
5006         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
5007
5008 static struct rack_sendmap *
5009 rack_find_lowest_rsm(struct tcp_rack *rack)
5010 {
5011         struct rack_sendmap *rsm;
5012
5013         /*
5014          * Walk the time-order transmitted list looking for an rsm that is
5015          * not acked. This will be the one that was sent the longest time
5016          * ago that is still outstanding.
5017          */
5018         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
5019                 if (rsm->r_flags & RACK_ACKED) {
5020                         continue;
5021                 }
5022                 goto finish;
5023         }
5024 finish:
5025         return (rsm);
5026 }
5027
5028 static struct rack_sendmap *
5029 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5030 {
5031         struct rack_sendmap *prsm;
5032
5033         /*
5034          * Walk the sequence order list backward until we hit and arrive at
5035          * the highest seq not acked. In theory when this is called it
5036          * should be the last segment (which it was not).
5037          */
5038         counter_u64_add(rack_find_high, 1);
5039         prsm = rsm;
5040         RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
5041                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5042                         continue;
5043                 }
5044                 return (prsm);
5045         }
5046         return (NULL);
5047 }
5048
5049 static uint32_t
5050 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
5051 {
5052         int32_t lro;
5053         uint32_t thresh;
5054
5055         /*
5056          * lro is the flag we use to determine if we have seen reordering.
5057          * If it gets set we have seen reordering. The reorder logic either
5058          * works in one of two ways:
5059          *
5060          * If reorder-fade is configured, then we track the last time we saw
5061          * re-ordering occur. If we reach the point where enough time as
5062          * passed we no longer consider reordering has occuring.
5063          *
5064          * Or if reorder-face is 0, then once we see reordering we consider
5065          * the connection to alway be subject to reordering and just set lro
5066          * to 1.
5067          *
5068          * In the end if lro is non-zero we add the extra time for
5069          * reordering in.
5070          */
5071         if (srtt == 0)
5072                 srtt = 1;
5073         if (rack->r_ctl.rc_reorder_ts) {
5074                 if (rack->r_ctl.rc_reorder_fade) {
5075                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5076                                 lro = cts - rack->r_ctl.rc_reorder_ts;
5077                                 if (lro == 0) {
5078                                         /*
5079                                          * No time as passed since the last
5080                                          * reorder, mark it as reordering.
5081                                          */
5082                                         lro = 1;
5083                                 }
5084                         } else {
5085                                 /* Negative time? */
5086                                 lro = 0;
5087                         }
5088                         if (lro > rack->r_ctl.rc_reorder_fade) {
5089                                 /* Turn off reordering seen too */
5090                                 rack->r_ctl.rc_reorder_ts = 0;
5091                                 lro = 0;
5092                         }
5093                 } else {
5094                         /* Reodering does not fade */
5095                         lro = 1;
5096                 }
5097         } else {
5098                 lro = 0;
5099         }
5100         thresh = srtt + rack->r_ctl.rc_pkt_delay;
5101         if (lro) {
5102                 /* It must be set, if not you get 1/4 rtt */
5103                 if (rack->r_ctl.rc_reorder_shift)
5104                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5105                 else
5106                         thresh += (srtt >> 2);
5107         } else {
5108                 thresh += 1;
5109         }
5110         /* We don't let the rack timeout be above a RTO */
5111         if (thresh > rack->rc_tp->t_rxtcur) {
5112                 thresh = rack->rc_tp->t_rxtcur;
5113         }
5114         /* And we don't want it above the RTO max either */
5115         if (thresh > rack_rto_max) {
5116                 thresh = rack_rto_max;
5117         }
5118         return (thresh);
5119 }
5120
5121 static uint32_t
5122 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5123                      struct rack_sendmap *rsm, uint32_t srtt)
5124 {
5125         struct rack_sendmap *prsm;
5126         uint32_t thresh, len;
5127         int segsiz;
5128
5129         if (srtt == 0)
5130                 srtt = 1;
5131         if (rack->r_ctl.rc_tlp_threshold)
5132                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5133         else
5134                 thresh = (srtt * 2);
5135
5136         /* Get the previous sent packet, if any */
5137         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5138         counter_u64_add(rack_enter_tlp_calc, 1);
5139         len = rsm->r_end - rsm->r_start;
5140         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5141                 /* Exactly like the ID */
5142                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5143                         uint32_t alt_thresh;
5144                         /*
5145                          * Compensate for delayed-ack with the d-ack time.
5146                          */
5147                         counter_u64_add(rack_used_tlpmethod, 1);
5148                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5149                         if (alt_thresh > thresh)
5150                                 thresh = alt_thresh;
5151                 }
5152         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5153                 /* 2.1 behavior */
5154                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5155                 if (prsm && (len <= segsiz)) {
5156                         /*
5157                          * Two packets outstanding, thresh should be (2*srtt) +
5158                          * possible inter-packet delay (if any).
5159                          */
5160                         uint32_t inter_gap = 0;
5161                         int idx, nidx;
5162
5163                         counter_u64_add(rack_used_tlpmethod, 1);
5164                         idx = rsm->r_rtr_cnt - 1;
5165                         nidx = prsm->r_rtr_cnt - 1;
5166                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5167                                 /* Yes it was sent later (or at the same time) */
5168                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5169                         }
5170                         thresh += inter_gap;
5171                 } else if (len <= segsiz) {
5172                         /*
5173                          * Possibly compensate for delayed-ack.
5174                          */
5175                         uint32_t alt_thresh;
5176
5177                         counter_u64_add(rack_used_tlpmethod2, 1);
5178                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5179                         if (alt_thresh > thresh)
5180                                 thresh = alt_thresh;
5181                 }
5182         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5183                 /* 2.2 behavior */
5184                 if (len <= segsiz) {
5185                         uint32_t alt_thresh;
5186                         /*
5187                          * Compensate for delayed-ack with the d-ack time.
5188                          */
5189                         counter_u64_add(rack_used_tlpmethod, 1);
5190                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5191                         if (alt_thresh > thresh)
5192                                 thresh = alt_thresh;
5193                 }
5194         }
5195         /* Not above an RTO */
5196         if (thresh > tp->t_rxtcur) {
5197                 thresh = tp->t_rxtcur;
5198         }
5199         /* Not above a RTO max */
5200         if (thresh > rack_rto_max) {
5201                 thresh = rack_rto_max;
5202         }
5203         /* Apply user supplied min TLP */
5204         if (thresh < rack_tlp_min) {
5205                 thresh = rack_tlp_min;
5206         }
5207         return (thresh);
5208 }
5209
5210 static uint32_t
5211 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5212 {
5213         /*
5214          * We want the rack_rtt which is the
5215          * last rtt we measured. However if that
5216          * does not exist we fallback to the srtt (which
5217          * we probably will never do) and then as a last
5218          * resort we use RACK_INITIAL_RTO if no srtt is
5219          * yet set.
5220          */
5221         if (rack->rc_rack_rtt)
5222                 return (rack->rc_rack_rtt);
5223         else if (tp->t_srtt == 0)
5224                 return (RACK_INITIAL_RTO);
5225         return (tp->t_srtt);
5226 }
5227
5228 static struct rack_sendmap *
5229 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5230 {
5231         /*
5232          * Check to see that we don't need to fall into recovery. We will
5233          * need to do so if our oldest transmit is past the time we should
5234          * have had an ack.
5235          */
5236         struct tcp_rack *rack;
5237         struct rack_sendmap *rsm;
5238         int32_t idx;
5239         uint32_t srtt, thresh;
5240
5241         rack = (struct tcp_rack *)tp->t_fb_ptr;
5242         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5243                 return (NULL);
5244         }
5245         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5246         if (rsm == NULL)
5247                 return (NULL);
5248
5249         if (rsm->r_flags & RACK_ACKED) {
5250                 rsm = rack_find_lowest_rsm(rack);
5251                 if (rsm == NULL)
5252                         return (NULL);
5253         }
5254         idx = rsm->r_rtr_cnt - 1;
5255         srtt = rack_grab_rtt(tp, rack);
5256         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5257         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5258                 return (NULL);
5259         }
5260         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5261                 return (NULL);
5262         }
5263         /* Ok if we reach here we are over-due and this guy can be sent */
5264         if (IN_RECOVERY(tp->t_flags) == 0) {
5265                 /*
5266                  * For the one that enters us into recovery record undo
5267                  * info.
5268                  */
5269                 rack->r_ctl.rc_rsm_start = rsm->r_start;
5270                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
5271                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
5272         }
5273         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
5274         return (rsm);
5275 }
5276
5277 static uint32_t
5278 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5279 {
5280         int32_t t;
5281         int32_t tt;
5282         uint32_t ret_val;
5283
5284         t = (tp->t_srtt + (tp->t_rttvar << 2));
5285         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5286             rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5287         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
5288                 tp->t_rxtshift++;
5289         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5290         ret_val = (uint32_t)tt;
5291         return (ret_val);
5292 }
5293
5294 static uint32_t
5295 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5296 {
5297         /*
5298          * Start the FR timer, we do this based on getting the first one in
5299          * the rc_tmap. Note that if its NULL we must stop the timer. in all
5300          * events we need to stop the running timer (if its running) before
5301          * starting the new one.
5302          */
5303         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5304         uint32_t srtt_cur;
5305         int32_t idx;
5306         int32_t is_tlp_timer = 0;
5307         struct rack_sendmap *rsm;
5308
5309         if (rack->t_timers_stopped) {
5310                 /* All timers have been stopped none are to run */
5311                 return (0);
5312         }
5313         if (rack->rc_in_persist) {
5314                 /* We can't start any timer in persists */
5315                 return (rack_get_persists_timer_val(tp, rack));
5316         }
5317         rack->rc_on_min_to = 0;
5318         if ((tp->t_state < TCPS_ESTABLISHED) ||
5319             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5320                 goto activate_rxt;
5321         }
5322         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5323         if ((rsm == NULL) || sup_rack) {
5324                 /* Nothing on the send map or no rack */
5325 activate_rxt:
5326                 time_since_sent = 0;
5327                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5328                 if (rsm) {
5329                         /*
5330                          * Should we discount the RTX timer any?
5331                          *
5332                          * We want to discount it the smallest amount.
5333                          * If a timer (Rack/TLP or RXT) has gone off more
5334                          * recently thats the discount we want to use (now - timer time).
5335                          * If the retransmit of the oldest packet was more recent then
5336                          * we want to use that (now - oldest-packet-last_transmit_time).
5337                          *
5338                          */
5339                         idx = rsm->r_rtr_cnt - 1;
5340                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5341                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5342                         else
5343                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5344                         if (TSTMP_GT(cts, tstmp_touse))
5345                             time_since_sent = cts - tstmp_touse;
5346                 }
5347                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5348                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5349                         to = tp->t_rxtcur;
5350                         if (to > time_since_sent)
5351                                 to -= time_since_sent;
5352                         else
5353                                 to = rack->r_ctl.rc_min_to;
5354                         if (to == 0)
5355                                 to = 1;
5356                         /* Special case for KEEPINIT */
5357                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5358                             (TP_KEEPINIT(tp) != 0) &&
5359                             rsm) {
5360                                 /*
5361                                  * We have to put a ceiling on the rxt timer
5362                                  * of the keep-init timeout.
5363                                  */
5364                                 uint32_t max_time, red;
5365
5366                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5367                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5368                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5369                                         if (red < max_time)
5370                                                 max_time -= red;
5371                                         else
5372                                                 max_time = 1;
5373                                 }
5374                                 /* Reduce timeout to the keep value if needed */
5375                                 if (max_time < to)
5376                                         to = max_time;
5377                         }
5378                         return (to);
5379                 }
5380                 return (0);
5381         }
5382         if (rsm->r_flags & RACK_ACKED) {
5383                 rsm = rack_find_lowest_rsm(rack);
5384                 if (rsm == NULL) {
5385                         /* No lowest? */
5386                         goto activate_rxt;
5387                 }
5388         }
5389         if (rack->sack_attack_disable) {
5390                 /*
5391                  * We don't want to do
5392                  * any TLP's if you are an attacker.
5393                  * Though if you are doing what
5394                  * is expected you may still have
5395                  * SACK-PASSED marks.
5396                  */
5397                 goto activate_rxt;
5398         }
5399         /* Convert from ms to usecs */
5400         if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5401                 if ((tp->t_flags & TF_SENTFIN) &&
5402                     ((tp->snd_max - tp->snd_una) == 1) &&
5403                     (rsm->r_flags & RACK_HAS_FIN)) {
5404                         /*
5405                          * We don't start a rack timer if all we have is a
5406                          * FIN outstanding.
5407                          */
5408                         goto activate_rxt;
5409                 }
5410                 if ((rack->use_rack_rr == 0) &&
5411                     (IN_FASTRECOVERY(tp->t_flags)) &&
5412                     (rack->rack_no_prr == 0) &&
5413                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5414                         /*
5415                          * We are not cheating, in recovery  and
5416                          * not enough ack's to yet get our next
5417                          * retransmission out.
5418                          *
5419                          * Note that classified attackers do not
5420                          * get to use the rack-cheat.
5421                          */
5422                         goto activate_tlp;
5423                 }
5424                 srtt = rack_grab_rtt(tp, rack);
5425                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5426                 idx = rsm->r_rtr_cnt - 1;
5427                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5428                 if (SEQ_GEQ(exp, cts)) {
5429                         to = exp - cts;
5430                         if (to < rack->r_ctl.rc_min_to) {
5431                                 to = rack->r_ctl.rc_min_to;
5432                                 if (rack->r_rr_config == 3)
5433                                         rack->rc_on_min_to = 1;
5434                         }
5435                 } else {
5436                         to = rack->r_ctl.rc_min_to;
5437                         if (rack->r_rr_config == 3)
5438                                 rack->rc_on_min_to = 1;
5439                 }
5440         } else {
5441                 /* Ok we need to do a TLP not RACK */
5442 activate_tlp:
5443                 if ((rack->rc_tlp_in_progress != 0) &&
5444                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5445                         /*
5446                          * The previous send was a TLP and we have sent
5447                          * N TLP's without sending new data.
5448                          */
5449                         goto activate_rxt;
5450                 }
5451                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5452                 if (rsm == NULL) {
5453                         /* We found no rsm to TLP with. */
5454                         goto activate_rxt;
5455                 }
5456                 if (rsm->r_flags & RACK_HAS_FIN) {
5457                         /* If its a FIN we dont do TLP */
5458                         rsm = NULL;
5459                         goto activate_rxt;
5460                 }
5461                 idx = rsm->r_rtr_cnt - 1;
5462                 time_since_sent = 0;
5463                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5464                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5465                 else
5466                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5467                 if (TSTMP_GT(cts, tstmp_touse))
5468                     time_since_sent = cts - tstmp_touse;
5469                 is_tlp_timer = 1;
5470                 if (tp->t_srtt) {
5471                         if ((rack->rc_srtt_measure_made == 0) &&
5472                             (tp->t_srtt == 1)) {
5473                                 /*
5474                                  * If another stack as run and set srtt to 1,
5475                                  * then the srtt was 0, so lets use the initial.
5476                                  */
5477                                 srtt = RACK_INITIAL_RTO;
5478                         } else {
5479                                 srtt_cur = tp->t_srtt;
5480                                 srtt = srtt_cur;
5481                         }
5482                 } else
5483                         srtt = RACK_INITIAL_RTO;
5484                 /*
5485                  * If the SRTT is not keeping up and the
5486                  * rack RTT has spiked we want to use
5487                  * the last RTT not the smoothed one.
5488                  */
5489                 if (rack_tlp_use_greater &&
5490                     tp->t_srtt &&
5491                     (srtt < rack_grab_rtt(tp, rack))) {
5492                         srtt = rack_grab_rtt(tp, rack);
5493                 }
5494                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5495                 if (thresh > time_since_sent) {
5496                         to = thresh - time_since_sent;
5497                 } else {
5498                         to = rack->r_ctl.rc_min_to;
5499                         rack_log_alt_to_to_cancel(rack,
5500                                                   thresh,               /* flex1 */
5501                                                   time_since_sent,      /* flex2 */
5502                                                   tstmp_touse,          /* flex3 */
5503                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5504                                                   (uint32_t)rsm->r_tim_lastsent[idx],
5505                                                   srtt,
5506                                                   idx, 99);
5507                 }
5508                 if (to < rack_tlp_min) {
5509                         to = rack_tlp_min;
5510                 }
5511                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5512                         /*
5513                          * If the TLP time works out to larger than the max
5514                          * RTO lets not do TLP.. just RTO.
5515                          */
5516                         goto activate_rxt;
5517                 }
5518         }
5519         if (is_tlp_timer == 0) {
5520                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5521         } else {
5522                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5523         }
5524         if (to == 0)
5525                 to = 1;
5526         return (to);
5527 }
5528
5529 static void
5530 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5531 {
5532         if (rack->rc_in_persist == 0) {
5533                 if (tp->t_flags & TF_GPUTINPROG) {
5534                         /*
5535                          * Stop the goodput now, the calling of the
5536                          * measurement function clears the flag.
5537                          */
5538                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5539                                                     RACK_QUALITY_PERSIST);
5540                 }
5541 #ifdef NETFLIX_SHARED_CWND
5542                 if (rack->r_ctl.rc_scw) {
5543                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5544                         rack->rack_scwnd_is_idle = 1;
5545                 }
5546 #endif
5547                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5548                 if (rack->r_ctl.rc_went_idle_time == 0)
5549                         rack->r_ctl.rc_went_idle_time = 1;
5550                 rack_timer_cancel(tp, rack, cts, __LINE__);
5551                 tp->t_rxtshift = 0;
5552                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5553                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5554                 rack->rc_in_persist = 1;
5555         }
5556 }
5557
5558 static void
5559 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5560 {
5561         if (rack->rc_inp->inp_in_hpts) {
5562                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5563                 rack->r_ctl.rc_hpts_flags = 0;
5564         }
5565 #ifdef NETFLIX_SHARED_CWND
5566         if (rack->r_ctl.rc_scw) {
5567                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5568                 rack->rack_scwnd_is_idle = 0;
5569         }
5570 #endif
5571         if (rack->rc_gp_dyn_mul &&
5572             (rack->use_fixed_rate == 0) &&
5573             (rack->rc_always_pace)) {
5574                 /*
5575                  * Do we count this as if a probe-rtt just
5576                  * finished?
5577                  */
5578                 uint32_t time_idle, idle_min;
5579
5580                 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5581                 idle_min = rack_min_probertt_hold;
5582                 if (rack_probertt_gpsrtt_cnt_div) {
5583                         uint64_t extra;
5584                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5585                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5586                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5587                         idle_min += (uint32_t)extra;
5588                 }
5589                 if (time_idle >= idle_min) {
5590                         /* Yes, we count it as a probe-rtt. */
5591                         uint32_t us_cts;
5592
5593                         us_cts = tcp_get_usecs(NULL);
5594                         if (rack->in_probe_rtt == 0) {
5595                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5596                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5597                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5598                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5599                         } else {
5600                                 rack_exit_probertt(rack, us_cts);
5601                         }
5602                 }
5603         }
5604         rack->rc_in_persist = 0;
5605         rack->r_ctl.rc_went_idle_time = 0;
5606         tp->t_rxtshift = 0;
5607         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5608            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5609         rack->r_ctl.rc_agg_delayed = 0;
5610         rack->r_early = 0;
5611         rack->r_late = 0;
5612         rack->r_ctl.rc_agg_early = 0;
5613 }
5614
5615 static void
5616 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5617                    struct hpts_diag *diag, struct timeval *tv)
5618 {
5619         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5620                 union tcp_log_stackspecific log;
5621
5622                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5623                 log.u_bbr.flex1 = diag->p_nxt_slot;
5624                 log.u_bbr.flex2 = diag->p_cur_slot;
5625                 log.u_bbr.flex3 = diag->slot_req;
5626                 log.u_bbr.flex4 = diag->inp_hptsslot;
5627                 log.u_bbr.flex5 = diag->slot_remaining;
5628                 log.u_bbr.flex6 = diag->need_new_to;
5629                 log.u_bbr.flex7 = diag->p_hpts_active;
5630                 log.u_bbr.flex8 = diag->p_on_min_sleep;
5631                 /* Hijack other fields as needed */
5632                 log.u_bbr.epoch = diag->have_slept;
5633                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5634                 log.u_bbr.pkts_out = diag->co_ret;
5635                 log.u_bbr.applimited = diag->hpts_sleep_time;
5636                 log.u_bbr.delivered = diag->p_prev_slot;
5637                 log.u_bbr.inflight = diag->p_runningslot;
5638                 log.u_bbr.bw_inuse = diag->wheel_slot;
5639                 log.u_bbr.rttProp = diag->wheel_cts;
5640                 log.u_bbr.timeStamp = cts;
5641                 log.u_bbr.delRate = diag->maxslots;
5642                 log.u_bbr.cur_del_rate = diag->p_curtick;
5643                 log.u_bbr.cur_del_rate <<= 32;
5644                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5645                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5646                     &rack->rc_inp->inp_socket->so_rcv,
5647                     &rack->rc_inp->inp_socket->so_snd,
5648                     BBR_LOG_HPTSDIAG, 0,
5649                     0, &log, false, tv);
5650         }
5651
5652 }
5653
5654 static void
5655 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5656 {
5657         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5658                 union tcp_log_stackspecific log;
5659                 struct timeval tv;
5660
5661                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5662                 log.u_bbr.flex1 = sb->sb_flags;
5663                 log.u_bbr.flex2 = len;
5664                 log.u_bbr.flex3 = sb->sb_state;
5665                 log.u_bbr.flex8 = type;
5666                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5667                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5668                     &rack->rc_inp->inp_socket->so_rcv,
5669                     &rack->rc_inp->inp_socket->so_snd,
5670                     TCP_LOG_SB_WAKE, 0,
5671                     len, &log, false, &tv);
5672         }
5673 }
5674
5675 static void
5676 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5677       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5678 {
5679         struct hpts_diag diag;
5680         struct inpcb *inp;
5681         struct timeval tv;
5682         uint32_t delayed_ack = 0;
5683         uint32_t hpts_timeout;
5684         uint32_t entry_slot = slot;
5685         uint8_t stopped;
5686         uint32_t left = 0;
5687         uint32_t us_cts;
5688
5689         inp = tp->t_inpcb;
5690         if ((tp->t_state == TCPS_CLOSED) ||
5691             (tp->t_state == TCPS_LISTEN)) {
5692                 return;
5693         }
5694         if (inp->inp_in_hpts) {
5695                 /* Already on the pacer */
5696                 return;
5697         }
5698         stopped = rack->rc_tmr_stopped;
5699         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5700                 left = rack->r_ctl.rc_timer_exp - cts;
5701         }
5702         rack->r_ctl.rc_timer_exp = 0;
5703         rack->r_ctl.rc_hpts_flags = 0;
5704         us_cts = tcp_get_usecs(&tv);
5705         /* Now early/late accounting */
5706         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5707         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5708                 /*
5709                  * We have a early carry over set,
5710                  * we can always add more time so we
5711                  * can always make this compensation.
5712                  *
5713                  * Note if ack's are allowed to wake us do not
5714                  * penalize the next timer for being awoke
5715                  * by an ack aka the rc_agg_early (non-paced mode).
5716                  */
5717                 slot += rack->r_ctl.rc_agg_early;
5718                 rack->r_early = 0;
5719                 rack->r_ctl.rc_agg_early = 0;
5720         }
5721         if (rack->r_late) {
5722                 /*
5723                  * This is harder, we can
5724                  * compensate some but it
5725                  * really depends on what
5726                  * the current pacing time is.
5727                  */
5728                 if (rack->r_ctl.rc_agg_delayed >= slot) {
5729                         /*
5730                          * We can't compensate for it all.
5731                          * And we have to have some time
5732                          * on the clock. We always have a min
5733                          * 10 slots (10 x 10 i.e. 100 usecs).
5734                          */
5735                         if (slot <= HPTS_TICKS_PER_SLOT) {
5736                                 /* We gain delay */
5737                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5738                                 slot = HPTS_TICKS_PER_SLOT;
5739                         } else {
5740                                 /* We take off some */
5741                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5742                                 slot = HPTS_TICKS_PER_SLOT;
5743                         }
5744                 } else {
5745                         slot -= rack->r_ctl.rc_agg_delayed;
5746                         rack->r_ctl.rc_agg_delayed = 0;
5747                         /* Make sure we have 100 useconds at minimum */
5748                         if (slot < HPTS_TICKS_PER_SLOT) {
5749                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5750                                 slot = HPTS_TICKS_PER_SLOT;
5751                         }
5752                         if (rack->r_ctl.rc_agg_delayed == 0)
5753                                 rack->r_late = 0;
5754                 }
5755         }
5756         if (slot) {
5757                 /* We are pacing too */
5758                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5759         }
5760         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5761 #ifdef NETFLIX_EXP_DETECTION
5762         if (rack->sack_attack_disable &&
5763             (slot < tcp_sad_pacing_interval)) {
5764                 /*
5765                  * We have a potential attacker on
5766                  * the line. We have possibly some
5767                  * (or now) pacing time set. We want to
5768                  * slow down the processing of sacks by some
5769                  * amount (if it is an attacker). Set the default
5770                  * slot for attackers in place (unless the orginal
5771                  * interval is longer). Its stored in
5772                  * micro-seconds, so lets convert to msecs.
5773                  */
5774                 slot = tcp_sad_pacing_interval;
5775         }
5776 #endif
5777         if (tp->t_flags & TF_DELACK) {
5778                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5779                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5780         }
5781         if (delayed_ack && ((hpts_timeout == 0) ||
5782                             (delayed_ack < hpts_timeout)))
5783                 hpts_timeout = delayed_ack;
5784         else
5785                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5786         /*
5787          * If no timers are going to run and we will fall off the hptsi
5788          * wheel, we resort to a keep-alive timer if its configured.
5789          */
5790         if ((hpts_timeout == 0) &&
5791             (slot == 0)) {
5792                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5793                     (tp->t_state <= TCPS_CLOSING)) {
5794                         /*
5795                          * Ok we have no timer (persists, rack, tlp, rxt  or
5796                          * del-ack), we don't have segments being paced. So
5797                          * all that is left is the keepalive timer.
5798                          */
5799                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5800                                 /* Get the established keep-alive time */
5801                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5802                         } else {
5803                                 /*
5804                                  * Get the initial setup keep-alive time,
5805                                  * note that this is probably not going to
5806                                  * happen, since rack will be running a rxt timer
5807                                  * if a SYN of some sort is outstanding. It is
5808                                  * actually handled in rack_timeout_rxt().
5809                                  */
5810                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5811                         }
5812                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5813                         if (rack->in_probe_rtt) {
5814                                 /*
5815                                  * We want to instead not wake up a long time from
5816                                  * now but to wake up about the time we would
5817                                  * exit probe-rtt and initiate a keep-alive ack.
5818                                  * This will get us out of probe-rtt and update
5819                                  * our min-rtt.
5820                                  */
5821                                 hpts_timeout = rack_min_probertt_hold;
5822                         }
5823                 }
5824         }
5825         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5826             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5827                 /*
5828                  * RACK, TLP, persists and RXT timers all are restartable
5829                  * based on actions input .. i.e we received a packet (ack
5830                  * or sack) and that changes things (rw, or snd_una etc).
5831                  * Thus we can restart them with a new value. For
5832                  * keep-alive, delayed_ack we keep track of what was left
5833                  * and restart the timer with a smaller value.
5834                  */
5835                 if (left < hpts_timeout)
5836                         hpts_timeout = left;
5837         }
5838         if (hpts_timeout) {
5839                 /*
5840                  * Hack alert for now we can't time-out over 2,147,483
5841                  * seconds (a bit more than 596 hours), which is probably ok
5842                  * :).
5843                  */
5844                 if (hpts_timeout > 0x7ffffffe)
5845                         hpts_timeout = 0x7ffffffe;
5846                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5847         }
5848         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5849         if ((rack->gp_ready == 0) &&
5850             (rack->use_fixed_rate == 0) &&
5851             (hpts_timeout < slot) &&
5852             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5853                 /*
5854                  * We have no good estimate yet for the
5855                  * old clunky burst mitigation or the
5856                  * real pacing. And the tlp or rxt is smaller
5857                  * than the pacing calculation. Lets not
5858                  * pace that long since we know the calculation
5859                  * so far is not accurate.
5860                  */
5861                 slot = hpts_timeout;
5862         }
5863         rack->r_ctl.last_pacing_time = slot;
5864         /**
5865          * Turn off all the flags for queuing by default. The
5866          * flags have important meanings to what happens when
5867          * LRO interacts with the transport. Most likely (by default now)
5868          * mbuf_queueing and ack compression are on. So the transport
5869          * has a couple of flags that control what happens (if those
5870          * are not on then these flags won't have any effect since it
5871          * won't go through the queuing LRO path).
5872          *
5873          * INP_MBUF_QUEUE_READY - This flags says that I am busy
5874          *                        pacing output, so don't disturb. But
5875          *                        it also means LRO can wake me if there
5876          *                        is a SACK arrival.
5877          *
5878          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5879          *                       with the above flag (QUEUE_READY) and
5880          *                       when present it says don't even wake me
5881          *                       if a SACK arrives.
5882          *
5883          * The idea behind these flags is that if we are pacing we
5884          * set the MBUF_QUEUE_READY and only get woken up if
5885          * a SACK arrives (which could change things) or if
5886          * our pacing timer expires. If, however, we have a rack
5887          * timer running, then we don't even want a sack to wake
5888          * us since the rack timer has to expire before we can send.
5889          *
5890          * Other cases should usually have none of the flags set
5891          * so LRO can call into us.
5892          */
5893         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5894         if (slot) {
5895                 rack->r_ctl.rc_last_output_to = us_cts + slot;
5896                 /*
5897                  * A pacing timer (slot) is being set, in
5898                  * such a case we cannot send (we are blocked by
5899                  * the timer). So lets tell LRO that it should not
5900                  * wake us unless there is a SACK. Note this only
5901                  * will be effective if mbuf queueing is on or
5902                  * compressed acks are being processed.
5903                  */
5904                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5905                 /*
5906                  * But wait if we have a Rack timer running
5907                  * even a SACK should not disturb us (with
5908                  * the exception of r_rr_config 3).
5909                  */
5910                 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5911                     (rack->r_rr_config != 3))
5912                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5913                 if (rack->rc_ack_can_sendout_data) {
5914                         /*
5915                          * Ahh but wait, this is that special case
5916                          * where the pacing timer can be disturbed
5917                          * backout the changes (used for non-paced
5918                          * burst limiting).
5919                          */
5920                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5921                 }
5922                 if ((rack->use_rack_rr) &&
5923                     (rack->r_rr_config < 2) &&
5924                     ((hpts_timeout) && (hpts_timeout < slot))) {
5925                         /*
5926                          * Arrange for the hpts to kick back in after the
5927                          * t-o if the t-o does not cause a send.
5928                          */
5929                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5930                                                    __LINE__, &diag);
5931                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5932                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5933                 } else {
5934                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
5935                                                    __LINE__, &diag);
5936                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5937                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5938                 }
5939         } else if (hpts_timeout) {
5940                 /*
5941                  * With respect to inp_flags2 here, lets let any new acks wake
5942                  * us up here. Since we are not pacing (no pacing timer), output
5943                  * can happen so we should let it. If its a Rack timer, then any inbound
5944                  * packet probably won't change the sending (we will be blocked)
5945                  * but it may change the prr stats so letting it in (the set defaults
5946                  * at the start of this block) are good enough.
5947                  */
5948                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5949                                            __LINE__, &diag);
5950                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5951                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5952         } else {
5953                 /* No timer starting */
5954 #ifdef INVARIANTS
5955                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5956                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5957                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5958                 }
5959 #endif
5960         }
5961         rack->rc_tmr_stopped = 0;
5962         if (slot)
5963                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5964 }
5965
5966 /*
5967  * RACK Timer, here we simply do logging and house keeping.
5968  * the normal rack_output() function will call the
5969  * appropriate thing to check if we need to do a RACK retransmit.
5970  * We return 1, saying don't proceed with rack_output only
5971  * when all timers have been stopped (destroyed PCB?).
5972  */
5973 static int
5974 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5975 {
5976         /*
5977          * This timer simply provides an internal trigger to send out data.
5978          * The check_recovery_mode call will see if there are needed
5979          * retransmissions, if so we will enter fast-recovery. The output
5980          * call may or may not do the same thing depending on sysctl
5981          * settings.
5982          */
5983         struct rack_sendmap *rsm;
5984
5985         if (tp->t_timers->tt_flags & TT_STOPPED) {
5986                 return (1);
5987         }
5988         counter_u64_add(rack_to_tot, 1);
5989         if (rack->r_state && (rack->r_state != tp->t_state))
5990                 rack_set_state(tp, rack);
5991         rack->rc_on_min_to = 0;
5992         rsm = rack_check_recovery_mode(tp, cts);
5993         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5994         if (rsm) {
5995                 rack->r_ctl.rc_resend = rsm;
5996                 rack->r_timer_override = 1;
5997                 if (rack->use_rack_rr) {
5998                         /*
5999                          * Don't accumulate extra pacing delay
6000                          * we are allowing the rack timer to
6001                          * over-ride pacing i.e. rrr takes precedence
6002                          * if the pacing interval is longer than the rrr
6003                          * time (in other words we get the min pacing
6004                          * time versus rrr pacing time).
6005                          */
6006                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6007                 }
6008         }
6009         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6010         if (rsm == NULL) {
6011                 /* restart a timer and return 1 */
6012                 rack_start_hpts_timer(rack, tp, cts,
6013                                       0, 0, 0);
6014                 return (1);
6015         }
6016         return (0);
6017 }
6018
6019 static void
6020 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6021 {
6022         if (rsm->m->m_len > rsm->orig_m_len) {
6023                 /*
6024                  * Mbuf grew, caused by sbcompress, our offset does
6025                  * not change.
6026                  */
6027                 rsm->orig_m_len = rsm->m->m_len;
6028         } else if (rsm->m->m_len < rsm->orig_m_len) {
6029                 /*
6030                  * Mbuf shrank, trimmed off the top by an ack, our
6031                  * offset changes.
6032                  */
6033                 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6034                 rsm->orig_m_len = rsm->m->m_len;
6035         }
6036 }
6037
6038 static void
6039 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6040 {
6041         struct mbuf *m;
6042         uint32_t soff;
6043
6044         if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
6045                 /* Fix up the orig_m_len and possibly the mbuf offset */
6046                 rack_adjust_orig_mlen(src_rsm);
6047         }
6048         m = src_rsm->m;
6049         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6050         while (soff >= m->m_len) {
6051                 /* Move out past this mbuf */
6052                 soff -= m->m_len;
6053                 m = m->m_next;
6054                 KASSERT((m != NULL),
6055                         ("rsm:%p nrsm:%p hit at soff:%u null m",
6056                          src_rsm, rsm, soff));
6057         }
6058         rsm->m = m;
6059         rsm->soff = soff;
6060         rsm->orig_m_len = m->m_len;
6061 }
6062
6063 static __inline void
6064 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
6065                struct rack_sendmap *rsm, uint32_t start)
6066 {
6067         int idx;
6068
6069         nrsm->r_start = start;
6070         nrsm->r_end = rsm->r_end;
6071         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6072         nrsm->r_flags = rsm->r_flags;
6073         nrsm->r_dupack = rsm->r_dupack;
6074         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6075         nrsm->r_rtr_bytes = 0;
6076         rsm->r_end = nrsm->r_start;
6077         nrsm->r_just_ret = rsm->r_just_ret;
6078         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6079                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6080         }
6081         /* Now if we have SYN flag we keep it on the left edge */
6082         if (nrsm->r_flags & RACK_HAS_SYN)
6083                 nrsm->r_flags &= ~RACK_HAS_SYN;
6084         /* Now if we have a FIN flag we keep it on the right edge */
6085         if (rsm->r_flags & RACK_HAS_FIN)
6086                 rsm->r_flags &= ~RACK_HAS_FIN;
6087         /* Push bit must go to the right edge as well */
6088         if (rsm->r_flags & RACK_HAD_PUSH)
6089                 rsm->r_flags &= ~RACK_HAD_PUSH;
6090         /* Clone over the state of the hw_tls flag */
6091         nrsm->r_hw_tls = rsm->r_hw_tls;
6092         /*
6093          * Now we need to find nrsm's new location in the mbuf chain
6094          * we basically calculate a new offset, which is soff +
6095          * how much is left in original rsm. Then we walk out the mbuf
6096          * chain to find the righ postion, it may be the same mbuf
6097          * or maybe not.
6098          */
6099         KASSERT(((rsm->m != NULL) ||
6100                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6101                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6102         if (rsm->m)
6103                 rack_setup_offset_for_rsm(rsm, nrsm);
6104 }
6105
6106 static struct rack_sendmap *
6107 rack_merge_rsm(struct tcp_rack *rack,
6108                struct rack_sendmap *l_rsm,
6109                struct rack_sendmap *r_rsm)
6110 {
6111         /*
6112          * We are merging two ack'd RSM's,
6113          * the l_rsm is on the left (lower seq
6114          * values) and the r_rsm is on the right
6115          * (higher seq value). The simplest way
6116          * to merge these is to move the right
6117          * one into the left. I don't think there
6118          * is any reason we need to try to find
6119          * the oldest (or last oldest retransmitted).
6120          */
6121         struct rack_sendmap *rm;
6122
6123         rack_log_map_chg(rack->rc_tp, rack, NULL,
6124                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6125         l_rsm->r_end = r_rsm->r_end;
6126         if (l_rsm->r_dupack < r_rsm->r_dupack)
6127                 l_rsm->r_dupack = r_rsm->r_dupack;
6128         if (r_rsm->r_rtr_bytes)
6129                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6130         if (r_rsm->r_in_tmap) {
6131                 /* This really should not happen */
6132                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6133                 r_rsm->r_in_tmap = 0;
6134         }
6135
6136         /* Now the flags */
6137         if (r_rsm->r_flags & RACK_HAS_FIN)
6138                 l_rsm->r_flags |= RACK_HAS_FIN;
6139         if (r_rsm->r_flags & RACK_TLP)
6140                 l_rsm->r_flags |= RACK_TLP;
6141         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6142                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6143         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6144             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6145                 /*
6146                  * If both are app-limited then let the
6147                  * free lower the count. If right is app
6148                  * limited and left is not, transfer.
6149                  */
6150                 l_rsm->r_flags |= RACK_APP_LIMITED;
6151                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
6152                 if (r_rsm == rack->r_ctl.rc_first_appl)
6153                         rack->r_ctl.rc_first_appl = l_rsm;
6154         }
6155         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6156 #ifdef INVARIANTS
6157         if (rm != r_rsm) {
6158                 panic("removing head in rack:%p rsm:%p rm:%p",
6159                       rack, r_rsm, rm);
6160         }
6161 #endif
6162         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6163                 /* Transfer the split limit to the map we free */
6164                 r_rsm->r_limit_type = l_rsm->r_limit_type;
6165                 l_rsm->r_limit_type = 0;
6166         }
6167         rack_free(rack, r_rsm);
6168         return (l_rsm);
6169 }
6170
6171 /*
6172  * TLP Timer, here we simply setup what segment we want to
6173  * have the TLP expire on, the normal rack_output() will then
6174  * send it out.
6175  *
6176  * We return 1, saying don't proceed with rack_output only
6177  * when all timers have been stopped (destroyed PCB?).
6178  */
6179 static int
6180 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6181 {
6182         /*
6183          * Tail Loss Probe.
6184          */
6185         struct rack_sendmap *rsm = NULL;
6186         struct rack_sendmap *insret;
6187         struct socket *so;
6188         uint32_t amm;
6189         uint32_t out, avail;
6190         int collapsed_win = 0;
6191
6192         if (tp->t_timers->tt_flags & TT_STOPPED) {
6193                 return (1);
6194         }
6195         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6196                 /* Its not time yet */
6197                 return (0);
6198         }
6199         if (ctf_progress_timeout_check(tp, true)) {
6200                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6201                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6202                 return (1);
6203         }
6204         /*
6205          * A TLP timer has expired. We have been idle for 2 rtts. So we now
6206          * need to figure out how to force a full MSS segment out.
6207          */
6208         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6209         rack->r_ctl.retran_during_recovery = 0;
6210         rack->r_ctl.dsack_byte_cnt = 0;
6211         counter_u64_add(rack_tlp_tot, 1);
6212         if (rack->r_state && (rack->r_state != tp->t_state))
6213                 rack_set_state(tp, rack);
6214         so = tp->t_inpcb->inp_socket;
6215         avail = sbavail(&so->so_snd);
6216         out = tp->snd_max - tp->snd_una;
6217         if (out > tp->snd_wnd) {
6218                 /* special case, we need a retransmission */
6219                 collapsed_win = 1;
6220                 goto need_retran;
6221         }
6222         if ((tp->t_flags & TF_GPUTINPROG) &&
6223             (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6224                 /*
6225                  * If this is the second in a row
6226                  * TLP and we are doing a measurement
6227                  * its time to abandon the measurement.
6228                  * Something is likely broken on
6229                  * the clients network and measuring a
6230                  * broken network does us no good.
6231                  */
6232                 tp->t_flags &= ~TF_GPUTINPROG;
6233                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6234                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6235                                            tp->gput_seq,
6236                                            0, 0, 18, __LINE__, NULL, 0);
6237         }
6238         /*
6239          * Check our send oldest always settings, and if
6240          * there is an oldest to send jump to the need_retran.
6241          */
6242         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6243                 goto need_retran;
6244
6245         if (avail > out) {
6246                 /* New data is available */
6247                 amm = avail - out;
6248                 if (amm > ctf_fixed_maxseg(tp)) {
6249                         amm = ctf_fixed_maxseg(tp);
6250                         if ((amm + out) > tp->snd_wnd) {
6251                                 /* We are rwnd limited */
6252                                 goto need_retran;
6253                         }
6254                 } else if (amm < ctf_fixed_maxseg(tp)) {
6255                         /* not enough to fill a MTU */
6256                         goto need_retran;
6257                 }
6258                 if (IN_FASTRECOVERY(tp->t_flags)) {
6259                         /* Unlikely */
6260                         if (rack->rack_no_prr == 0) {
6261                                 if (out + amm <= tp->snd_wnd) {
6262                                         rack->r_ctl.rc_prr_sndcnt = amm;
6263                                         rack->r_ctl.rc_tlp_new_data = amm;
6264                                         rack_log_to_prr(rack, 4, 0);
6265                                 }
6266                         } else
6267                                 goto need_retran;
6268                 } else {
6269                         /* Set the send-new override */
6270                         if (out + amm <= tp->snd_wnd)
6271                                 rack->r_ctl.rc_tlp_new_data = amm;
6272                         else
6273                                 goto need_retran;
6274                 }
6275                 rack->r_ctl.rc_tlpsend = NULL;
6276                 counter_u64_add(rack_tlp_newdata, 1);
6277                 goto send;
6278         }
6279 need_retran:
6280         /*
6281          * Ok we need to arrange the last un-acked segment to be re-sent, or
6282          * optionally the first un-acked segment.
6283          */
6284         if (collapsed_win == 0) {
6285                 if (rack_always_send_oldest)
6286                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6287                 else {
6288                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6289                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6290                                 rsm = rack_find_high_nonack(rack, rsm);
6291                         }
6292                 }
6293                 if (rsm == NULL) {
6294                         counter_u64_add(rack_tlp_does_nada, 1);
6295 #ifdef TCP_BLACKBOX
6296                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6297 #endif
6298                         goto out;
6299                 }
6300         } else {
6301                 /*
6302                  * We must find the last segment
6303                  * that was acceptable by the client.
6304                  */
6305                 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6306                         if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6307                                 /* Found one */
6308                                 break;
6309                         }
6310                 }
6311                 if (rsm == NULL) {
6312                         /* None? if so send the first */
6313                         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6314                         if (rsm == NULL) {
6315                                 counter_u64_add(rack_tlp_does_nada, 1);
6316 #ifdef TCP_BLACKBOX
6317                                 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6318 #endif
6319                                 goto out;
6320                         }
6321                 }
6322         }
6323         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6324                 /*
6325                  * We need to split this the last segment in two.
6326                  */
6327                 struct rack_sendmap *nrsm;
6328
6329                 nrsm = rack_alloc_full_limit(rack);
6330                 if (nrsm == NULL) {
6331                         /*
6332                          * No memory to split, we will just exit and punt
6333                          * off to the RXT timer.
6334                          */
6335                         counter_u64_add(rack_tlp_does_nada, 1);
6336                         goto out;
6337                 }
6338                 rack_clone_rsm(rack, nrsm, rsm,
6339                                (rsm->r_end - ctf_fixed_maxseg(tp)));
6340                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6341                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6342 #ifdef INVARIANTS
6343                 if (insret != NULL) {
6344                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6345                               nrsm, insret, rack, rsm);
6346                 }
6347 #endif
6348                 if (rsm->r_in_tmap) {
6349                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6350                         nrsm->r_in_tmap = 1;
6351                 }
6352                 rsm->r_flags &= (~RACK_HAS_FIN);
6353                 rsm = nrsm;
6354         }
6355         rack->r_ctl.rc_tlpsend = rsm;
6356 send:
6357         /* Make sure output path knows we are doing a TLP */
6358         *doing_tlp = 1;
6359         rack->r_timer_override = 1;
6360         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6361         return (0);
6362 out:
6363         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6364         return (0);
6365 }
6366
6367 /*
6368  * Delayed ack Timer, here we simply need to setup the
6369  * ACK_NOW flag and remove the DELACK flag. From there
6370  * the output routine will send the ack out.
6371  *
6372  * We only return 1, saying don't proceed, if all timers
6373  * are stopped (destroyed PCB?).
6374  */
6375 static int
6376 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6377 {
6378         if (tp->t_timers->tt_flags & TT_STOPPED) {
6379                 return (1);
6380         }
6381         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6382         tp->t_flags &= ~TF_DELACK;
6383         tp->t_flags |= TF_ACKNOW;
6384         KMOD_TCPSTAT_INC(tcps_delack);
6385         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6386         return (0);
6387 }
6388
6389 /*
6390  * Persists timer, here we simply send the
6391  * same thing as a keepalive will.
6392  * the one byte send.
6393  *
6394  * We only return 1, saying don't proceed, if all timers
6395  * are stopped (destroyed PCB?).
6396  */
6397 static int
6398 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6399 {
6400         struct tcptemp *t_template;
6401         struct inpcb *inp;
6402         int32_t retval = 1;
6403
6404         inp = tp->t_inpcb;
6405
6406         if (tp->t_timers->tt_flags & TT_STOPPED) {
6407                 return (1);
6408         }
6409         if (rack->rc_in_persist == 0)
6410                 return (0);
6411         if (ctf_progress_timeout_check(tp, false)) {
6412                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6413                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6414                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6415                 return (1);
6416         }
6417         KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6418         /*
6419          * Persistence timer into zero window. Force a byte to be output, if
6420          * possible.
6421          */
6422         KMOD_TCPSTAT_INC(tcps_persisttimeo);
6423         /*
6424          * Hack: if the peer is dead/unreachable, we do not time out if the
6425          * window is closed.  After a full backoff, drop the connection if
6426          * the idle time (no responses to probes) reaches the maximum
6427          * backoff that we would use if retransmitting.
6428          */
6429         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6430             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6431              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6432                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6433                 retval = 1;
6434                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6435                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6436                 goto out;
6437         }
6438         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6439             tp->snd_una == tp->snd_max)
6440                 rack_exit_persist(tp, rack, cts);
6441         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6442         /*
6443          * If the user has closed the socket then drop a persisting
6444          * connection after a much reduced timeout.
6445          */
6446         if (tp->t_state > TCPS_CLOSE_WAIT &&
6447             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6448                 retval = 1;
6449                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6450                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6451                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6452                 goto out;
6453         }
6454         t_template = tcpip_maketemplate(rack->rc_inp);
6455         if (t_template) {
6456                 /* only set it if we were answered */
6457                 if (rack->forced_ack == 0) {
6458                         rack->forced_ack = 1;
6459                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6460                 }
6461                 tcp_respond(tp, t_template->tt_ipgen,
6462                             &t_template->tt_t, (struct mbuf *)NULL,
6463                             tp->rcv_nxt, tp->snd_una - 1, 0);
6464                 /* This sends an ack */
6465                 if (tp->t_flags & TF_DELACK)
6466                         tp->t_flags &= ~TF_DELACK;
6467                 free(t_template, M_TEMP);
6468         }
6469         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6470                 tp->t_rxtshift++;
6471 out:
6472         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6473         rack_start_hpts_timer(rack, tp, cts,
6474                               0, 0, 0);
6475         return (retval);
6476 }
6477
6478 /*
6479  * If a keepalive goes off, we had no other timers
6480  * happening. We always return 1 here since this
6481  * routine either drops the connection or sends
6482  * out a segment with respond.
6483  */
6484 static int
6485 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6486 {
6487         struct tcptemp *t_template;
6488         struct inpcb *inp;
6489
6490         if (tp->t_timers->tt_flags & TT_STOPPED) {
6491                 return (1);
6492         }
6493         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6494         inp = tp->t_inpcb;
6495         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6496         /*
6497          * Keep-alive timer went off; send something or drop connection if
6498          * idle for too long.
6499          */
6500         KMOD_TCPSTAT_INC(tcps_keeptimeo);
6501         if (tp->t_state < TCPS_ESTABLISHED)
6502                 goto dropit;
6503         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6504             tp->t_state <= TCPS_CLOSING) {
6505                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6506                         goto dropit;
6507                 /*
6508                  * Send a packet designed to force a response if the peer is
6509                  * up and reachable: either an ACK if the connection is
6510                  * still alive, or an RST if the peer has closed the
6511                  * connection due to timeout or reboot. Using sequence
6512                  * number tp->snd_una-1 causes the transmitted zero-length
6513                  * segment to lie outside the receive window; by the
6514                  * protocol spec, this requires the correspondent TCP to
6515                  * respond.
6516                  */
6517                 KMOD_TCPSTAT_INC(tcps_keepprobe);
6518                 t_template = tcpip_maketemplate(inp);
6519                 if (t_template) {
6520                         if (rack->forced_ack == 0) {
6521                                 rack->forced_ack = 1;
6522                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6523                         }
6524                         tcp_respond(tp, t_template->tt_ipgen,
6525                             &t_template->tt_t, (struct mbuf *)NULL,
6526                             tp->rcv_nxt, tp->snd_una - 1, 0);
6527                         free(t_template, M_TEMP);
6528                 }
6529         }
6530         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6531         return (1);
6532 dropit:
6533         KMOD_TCPSTAT_INC(tcps_keepdrops);
6534         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6535         tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6536         return (1);
6537 }
6538
6539 /*
6540  * Retransmit helper function, clear up all the ack
6541  * flags and take care of important book keeping.
6542  */
6543 static void
6544 rack_remxt_tmr(struct tcpcb *tp)
6545 {
6546         /*
6547          * The retransmit timer went off, all sack'd blocks must be
6548          * un-acked.
6549          */
6550         struct rack_sendmap *rsm, *trsm = NULL;
6551         struct tcp_rack *rack;
6552
6553         rack = (struct tcp_rack *)tp->t_fb_ptr;
6554         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6555         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6556         if (rack->r_state && (rack->r_state != tp->t_state))
6557                 rack_set_state(tp, rack);
6558         /*
6559          * Ideally we would like to be able to
6560          * mark SACK-PASS on anything not acked here.
6561          *
6562          * However, if we do that we would burst out
6563          * all that data 1ms apart. This would be unwise,
6564          * so for now we will just let the normal rxt timer
6565          * and tlp timer take care of it.
6566          *
6567          * Also we really need to stick them back in sequence
6568          * order. This way we send in the proper order and any
6569          * sacks that come floating in will "re-ack" the data.
6570          * To do this we zap the tmap with an INIT and then
6571          * walk through and place every rsm in the RB tree
6572          * back in its seq ordered place.
6573          */
6574         TAILQ_INIT(&rack->r_ctl.rc_tmap);
6575         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6576                 rsm->r_dupack = 0;
6577                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6578                 /* We must re-add it back to the tlist */
6579                 if (trsm == NULL) {
6580                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6581                 } else {
6582                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6583                 }
6584                 rsm->r_in_tmap = 1;
6585                 trsm = rsm;
6586                 if (rsm->r_flags & RACK_ACKED)
6587                         rsm->r_flags |= RACK_WAS_ACKED;
6588                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6589         }
6590         /* Clear the count (we just un-acked them) */
6591         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6592         rack->r_ctl.rc_sacked = 0;
6593         rack->r_ctl.rc_sacklast = NULL;
6594         rack->r_ctl.rc_agg_delayed = 0;
6595         rack->r_early = 0;
6596         rack->r_ctl.rc_agg_early = 0;
6597         rack->r_late = 0;
6598         /* Clear the tlp rtx mark */
6599         rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6600         if (rack->r_ctl.rc_resend != NULL)
6601                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6602         rack->r_ctl.rc_prr_sndcnt = 0;
6603         rack_log_to_prr(rack, 6, 0);
6604         rack->r_timer_override = 1;
6605         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6606 #ifdef NETFLIX_EXP_DETECTION
6607             || (rack->sack_attack_disable != 0)
6608 #endif
6609                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6610                 /*
6611                  * For non-sack customers new data
6612                  * needs to go out as retransmits until
6613                  * we retransmit up to snd_max.
6614                  */
6615                 rack->r_must_retran = 1;
6616                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6617                                                 rack->r_ctl.rc_sacked);
6618         }
6619         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6620 }
6621
6622 static void
6623 rack_convert_rtts(struct tcpcb *tp)
6624 {
6625         if (tp->t_srtt > 1) {
6626                 uint32_t val, frac;
6627
6628                 val = tp->t_srtt >> TCP_RTT_SHIFT;
6629                 frac = tp->t_srtt & 0x1f;
6630                 tp->t_srtt = TICKS_2_USEC(val);
6631                 /*
6632                  * frac is the fractional part of the srtt (if any)
6633                  * but its in ticks and every bit represents
6634                  * 1/32nd of a hz.
6635                  */
6636                 if (frac) {
6637                         if (hz == 1000) {
6638                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6639                         } else {
6640                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6641                         }
6642                         tp->t_srtt += frac;
6643                 }
6644         }
6645         if (tp->t_rttvar) {
6646                 uint32_t val, frac;
6647
6648                 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6649                 frac = tp->t_rttvar & 0x1f;
6650                 tp->t_rttvar = TICKS_2_USEC(val);
6651                 /*
6652                  * frac is the fractional part of the srtt (if any)
6653                  * but its in ticks and every bit represents
6654                  * 1/32nd of a hz.
6655                  */
6656                 if (frac) {
6657                         if (hz == 1000) {
6658                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6659                         } else {
6660                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6661                         }
6662                         tp->t_rttvar += frac;
6663                 }
6664         }
6665         tp->t_rxtcur = RACK_REXMTVAL(tp);
6666         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6667                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6668         }
6669         if (tp->t_rxtcur > rack_rto_max) {
6670                 tp->t_rxtcur = rack_rto_max;
6671         }
6672 }
6673
6674 static void
6675 rack_cc_conn_init(struct tcpcb *tp)
6676 {
6677         struct tcp_rack *rack;
6678         uint32_t srtt;
6679
6680         rack = (struct tcp_rack *)tp->t_fb_ptr;
6681         srtt = tp->t_srtt;
6682         cc_conn_init(tp);
6683         /*
6684          * Now convert to rack's internal format,
6685          * if required.
6686          */
6687         if ((srtt == 0) && (tp->t_srtt != 0))
6688                 rack_convert_rtts(tp);
6689         /*
6690          * We want a chance to stay in slowstart as
6691          * we create a connection. TCP spec says that
6692          * initially ssthresh is infinite. For our
6693          * purposes that is the snd_wnd.
6694          */
6695         if (tp->snd_ssthresh < tp->snd_wnd) {
6696                 tp->snd_ssthresh = tp->snd_wnd;
6697         }
6698         /*
6699          * We also want to assure a IW worth of
6700          * data can get inflight.
6701          */
6702         if (rc_init_window(rack) < tp->snd_cwnd)
6703                 tp->snd_cwnd = rc_init_window(rack);
6704 }
6705
6706 /*
6707  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6708  * we will setup to retransmit the lowest seq number outstanding.
6709  */
6710 static int
6711 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6712 {
6713         int32_t rexmt;
6714         struct inpcb *inp;
6715         int32_t retval = 0;
6716         bool isipv6;
6717
6718         inp = tp->t_inpcb;
6719         if (tp->t_timers->tt_flags & TT_STOPPED) {
6720                 return (1);
6721         }
6722         if ((tp->t_flags & TF_GPUTINPROG) &&
6723             (tp->t_rxtshift)) {
6724                 /*
6725                  * We have had a second timeout
6726                  * measurements on successive rxt's are not profitable.
6727                  * It is unlikely to be of any use (the network is
6728                  * broken or the client went away).
6729                  */
6730                 tp->t_flags &= ~TF_GPUTINPROG;
6731                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6732                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
6733                                            tp->gput_seq,
6734                                            0, 0, 18, __LINE__, NULL, 0);
6735         }
6736         if (ctf_progress_timeout_check(tp, false)) {
6737                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6738                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6739                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6740                 return (1);
6741         }
6742         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6743         rack->r_ctl.retran_during_recovery = 0;
6744         rack->r_ctl.dsack_byte_cnt = 0;
6745         if (IN_FASTRECOVERY(tp->t_flags))
6746                 tp->t_flags |= TF_WASFRECOVERY;
6747         else
6748                 tp->t_flags &= ~TF_WASFRECOVERY;
6749         if (IN_CONGRECOVERY(tp->t_flags))
6750                 tp->t_flags |= TF_WASCRECOVERY;
6751         else
6752                 tp->t_flags &= ~TF_WASCRECOVERY;
6753         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6754             (tp->snd_una == tp->snd_max)) {
6755                 /* Nothing outstanding .. nothing to do */
6756                 return (0);
6757         }
6758         /*
6759          * Rack can only run one timer  at a time, so we cannot
6760          * run a KEEPINIT (gating SYN sending) and a retransmit
6761          * timer for the SYN. So if we are in a front state and
6762          * have a KEEPINIT timer we need to check the first transmit
6763          * against now to see if we have exceeded the KEEPINIT time
6764          * (if one is set).
6765          */
6766         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6767             (TP_KEEPINIT(tp) != 0)) {
6768                 struct rack_sendmap *rsm;
6769
6770                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6771                 if (rsm) {
6772                         /* Ok we have something outstanding to test keepinit with */
6773                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6774                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6775                                 /* We have exceeded the KEEPINIT time */
6776                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6777                                 goto drop_it;
6778                         }
6779                 }
6780         }
6781         /*
6782          * Retransmission timer went off.  Message has not been acked within
6783          * retransmit interval.  Back off to a longer retransmit interval
6784          * and retransmit one segment.
6785          */
6786         rack_remxt_tmr(tp);
6787         if ((rack->r_ctl.rc_resend == NULL) ||
6788             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6789                 /*
6790                  * If the rwnd collapsed on
6791                  * the one we are retransmitting
6792                  * it does not count against the
6793                  * rxt count.
6794                  */
6795                 tp->t_rxtshift++;
6796         }
6797         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6798                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6799 drop_it:
6800                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6801                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6802                 retval = 1;
6803                 tcp_set_inp_to_drop(rack->rc_inp,
6804                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
6805                 goto out;
6806         }
6807         if (tp->t_state == TCPS_SYN_SENT) {
6808                 /*
6809                  * If the SYN was retransmitted, indicate CWND to be limited
6810                  * to 1 segment in cc_conn_init().
6811                  */
6812                 tp->snd_cwnd = 1;
6813         } else if (tp->t_rxtshift == 1) {
6814                 /*
6815                  * first retransmit; record ssthresh and cwnd so they can be
6816                  * recovered if this turns out to be a "bad" retransmit. A
6817                  * retransmit is considered "bad" if an ACK for this segment
6818                  * is received within RTT/2 interval; the assumption here is
6819                  * that the ACK was already in flight.  See "On Estimating
6820                  * End-to-End Network Path Properties" by Allman and Paxson
6821                  * for more details.
6822                  */
6823                 tp->snd_cwnd_prev = tp->snd_cwnd;
6824                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6825                 tp->snd_recover_prev = tp->snd_recover;
6826                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6827                 tp->t_flags |= TF_PREVVALID;
6828         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6829                 tp->t_flags &= ~TF_PREVVALID;
6830         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6831         if ((tp->t_state == TCPS_SYN_SENT) ||
6832             (tp->t_state == TCPS_SYN_RECEIVED))
6833                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6834         else
6835                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6836
6837         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6838            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6839         /*
6840          * We enter the path for PLMTUD if connection is established or, if
6841          * connection is FIN_WAIT_1 status, reason for the last is that if
6842          * amount of data we send is very small, we could send it in couple
6843          * of packets and process straight to FIN. In that case we won't
6844          * catch ESTABLISHED state.
6845          */
6846 #ifdef INET6
6847         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6848 #else
6849         isipv6 = false;
6850 #endif
6851         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6852             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6853             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6854             ((tp->t_state == TCPS_ESTABLISHED) ||
6855             (tp->t_state == TCPS_FIN_WAIT_1))) {
6856                 /*
6857                  * Idea here is that at each stage of mtu probe (usually,
6858                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6859                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6860                  * should take care of that.
6861                  */
6862                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6863                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6864                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6865                     tp->t_rxtshift % 2 == 0)) {
6866                         /*
6867                          * Enter Path MTU Black-hole Detection mechanism: -
6868                          * Disable Path MTU Discovery (IP "DF" bit). -
6869                          * Reduce MTU to lower value than what we negotiated
6870                          * with peer.
6871                          */
6872                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6873                                 /* Record that we may have found a black hole. */
6874                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6875                                 /* Keep track of previous MSS. */
6876                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6877                         }
6878
6879                         /*
6880                          * Reduce the MSS to blackhole value or to the
6881                          * default in an attempt to retransmit.
6882                          */
6883 #ifdef INET6
6884                         if (isipv6 &&
6885                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6886                                 /* Use the sysctl tuneable blackhole MSS. */
6887                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6888                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6889                         } else if (isipv6) {
6890                                 /* Use the default MSS. */
6891                                 tp->t_maxseg = V_tcp_v6mssdflt;
6892                                 /*
6893                                  * Disable Path MTU Discovery when we switch
6894                                  * to minmss.
6895                                  */
6896                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6897                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6898                         }
6899 #endif
6900 #if defined(INET6) && defined(INET)
6901                         else
6902 #endif
6903 #ifdef INET
6904                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6905                                 /* Use the sysctl tuneable blackhole MSS. */
6906                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6907                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6908                         } else {
6909                                 /* Use the default MSS. */
6910                                 tp->t_maxseg = V_tcp_mssdflt;
6911                                 /*
6912                                  * Disable Path MTU Discovery when we switch
6913                                  * to minmss.
6914                                  */
6915                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6916                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6917                         }
6918 #endif
6919                 } else {
6920                         /*
6921                          * If further retransmissions are still unsuccessful
6922                          * with a lowered MTU, maybe this isn't a blackhole
6923                          * and we restore the previous MSS and blackhole
6924                          * detection flags. The limit '6' is determined by
6925                          * giving each probe stage (1448, 1188, 524) 2
6926                          * chances to recover.
6927                          */
6928                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6929                             (tp->t_rxtshift >= 6)) {
6930                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6931                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6932                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6933                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6934                         }
6935                 }
6936         }
6937         /*
6938          * Disable RFC1323 and SACK if we haven't got any response to
6939          * our third SYN to work-around some broken terminal servers
6940          * (most of which have hopefully been retired) that have bad VJ
6941          * header compression code which trashes TCP segments containing
6942          * unknown-to-them TCP options.
6943          */
6944         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6945             (tp->t_rxtshift == 3))
6946                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6947         /*
6948          * If we backed off this far, our srtt estimate is probably bogus.
6949          * Clobber it so we'll take the next rtt measurement as our srtt;
6950          * move the current srtt into rttvar to keep the current retransmit
6951          * times until then.
6952          */
6953         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6954 #ifdef INET6
6955                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6956                         in6_losing(tp->t_inpcb);
6957                 else
6958 #endif
6959                         in_losing(tp->t_inpcb);
6960                 tp->t_rttvar += tp->t_srtt;
6961                 tp->t_srtt = 0;
6962         }
6963         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6964         tp->snd_recover = tp->snd_max;
6965         tp->t_flags |= TF_ACKNOW;
6966         tp->t_rtttime = 0;
6967         rack_cong_signal(tp, CC_RTO, tp->snd_una);
6968 out:
6969         return (retval);
6970 }
6971
6972 static int
6973 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6974 {
6975         int32_t ret = 0;
6976         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6977
6978         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6979             (tp->t_flags & TF_GPUTINPROG)) {
6980                 /*
6981                  * We have a goodput in progress
6982                  * and we have entered a late state.
6983                  * Do we have enough data in the sb
6984                  * to handle the GPUT request?
6985                  */
6986                 uint32_t bytes;
6987
6988                 bytes = tp->gput_ack - tp->gput_seq;
6989                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
6990                         bytes += tp->gput_seq - tp->snd_una;
6991                 if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
6992                         /*
6993                          * There are not enough bytes in the socket
6994                          * buffer that have been sent to cover this
6995                          * measurement. Cancel it.
6996                          */
6997                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6998                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
6999                                                    tp->gput_seq,
7000                                                    0, 0, 18, __LINE__, NULL, 0);
7001                         tp->t_flags &= ~TF_GPUTINPROG;
7002                 }
7003         }
7004         if (timers == 0) {
7005                 return (0);
7006         }
7007         if (tp->t_state == TCPS_LISTEN) {
7008                 /* no timers on listen sockets */
7009                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7010                         return (0);
7011                 return (1);
7012         }
7013         if ((timers & PACE_TMR_RACK) &&
7014             rack->rc_on_min_to) {
7015                 /*
7016                  * For the rack timer when we
7017                  * are on a min-timeout (which means rrr_conf = 3)
7018                  * we don't want to check the timer. It may
7019                  * be going off for a pace and thats ok we
7020                  * want to send the retransmit (if its ready).
7021                  *
7022                  * If its on a normal rack timer (non-min) then
7023                  * we will check if its expired.
7024                  */
7025                 goto skip_time_check;
7026         }
7027         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7028                 uint32_t left;
7029
7030                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
7031                         ret = -1;
7032                         rack_log_to_processing(rack, cts, ret, 0);
7033                         return (0);
7034                 }
7035                 if (hpts_calling == 0) {
7036                         /*
7037                          * A user send or queued mbuf (sack) has called us? We
7038                          * return 0 and let the pacing guards
7039                          * deal with it if they should or
7040                          * should not cause a send.
7041                          */
7042                         ret = -2;
7043                         rack_log_to_processing(rack, cts, ret, 0);
7044                         return (0);
7045                 }
7046                 /*
7047                  * Ok our timer went off early and we are not paced false
7048                  * alarm, go back to sleep.
7049                  */
7050                 ret = -3;
7051                 left = rack->r_ctl.rc_timer_exp - cts;
7052                 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
7053                 rack_log_to_processing(rack, cts, ret, left);
7054                 return (1);
7055         }
7056 skip_time_check:
7057         rack->rc_tmr_stopped = 0;
7058         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
7059         if (timers & PACE_TMR_DELACK) {
7060                 ret = rack_timeout_delack(tp, rack, cts);
7061         } else if (timers & PACE_TMR_RACK) {
7062                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
7063                 rack->r_fast_output = 0;
7064                 ret = rack_timeout_rack(tp, rack, cts);
7065         } else if (timers & PACE_TMR_TLP) {
7066                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
7067                 ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
7068         } else if (timers & PACE_TMR_RXT) {
7069                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
7070                 rack->r_fast_output = 0;
7071                 ret = rack_timeout_rxt(tp, rack, cts);
7072         } else if (timers & PACE_TMR_PERSIT) {
7073                 ret = rack_timeout_persist(tp, rack, cts);
7074         } else if (timers & PACE_TMR_KEEP) {
7075                 ret = rack_timeout_keepalive(tp, rack, cts);
7076         }
7077         rack_log_to_processing(rack, cts, ret, timers);
7078         return (ret);
7079 }
7080
7081 static void
7082 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
7083 {
7084         struct timeval tv;
7085         uint32_t us_cts, flags_on_entry;
7086         uint8_t hpts_removed = 0;
7087
7088         flags_on_entry = rack->r_ctl.rc_hpts_flags;
7089         us_cts = tcp_get_usecs(&tv);
7090         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7091             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7092              ((tp->snd_max - tp->snd_una) == 0))) {
7093                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7094                 hpts_removed = 1;
7095                 /* If we were not delayed cancel out the flag. */
7096                 if ((tp->snd_max - tp->snd_una) == 0)
7097                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7098                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7099         }
7100         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7101                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7102                 if (rack->rc_inp->inp_in_hpts &&
7103                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7104                         /*
7105                          * Canceling timer's when we have no output being
7106                          * paced. We also must remove ourselves from the
7107                          * hpts.
7108                          */
7109                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7110                         hpts_removed = 1;
7111                 }
7112                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7113         }
7114         if (hpts_removed == 0)
7115                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7116 }
7117
7118 static void
7119 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7120 {
7121         return;
7122 }
7123
7124 static int
7125 rack_stopall(struct tcpcb *tp)
7126 {
7127         struct tcp_rack *rack;
7128         rack = (struct tcp_rack *)tp->t_fb_ptr;
7129         rack->t_timers_stopped = 1;
7130         return (0);
7131 }
7132
7133 static void
7134 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7135 {
7136         return;
7137 }
7138
7139 static int
7140 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7141 {
7142         return (0);
7143 }
7144
7145 static void
7146 rack_stop_all_timers(struct tcpcb *tp)
7147 {
7148         struct tcp_rack *rack;
7149
7150         /*
7151          * Assure no timers are running.
7152          */
7153         if (tcp_timer_active(tp, TT_PERSIST)) {
7154                 /* We enter in persists, set the flag appropriately */
7155                 rack = (struct tcp_rack *)tp->t_fb_ptr;
7156                 rack->rc_in_persist = 1;
7157         }
7158         tcp_timer_suspend(tp, TT_PERSIST);
7159         tcp_timer_suspend(tp, TT_REXMT);
7160         tcp_timer_suspend(tp, TT_KEEP);
7161         tcp_timer_suspend(tp, TT_DELACK);
7162 }
7163
7164 static void
7165 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7166     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7167 {
7168         int32_t idx;
7169         uint16_t stripped_flags;
7170
7171         rsm->r_rtr_cnt++;
7172         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7173         rsm->r_dupack = 0;
7174         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7175                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7176                 rsm->r_flags |= RACK_OVERMAX;
7177         }
7178         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7179                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7180                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7181         }
7182         idx = rsm->r_rtr_cnt - 1;
7183         rsm->r_tim_lastsent[idx] = ts;
7184         stripped_flags = rsm->r_flags & ~(RACK_SENT_SP|RACK_SENT_FP);
7185         if (rsm->r_flags & RACK_ACKED) {
7186                 /* Problably MTU discovery messing with us */
7187                 rsm->r_flags &= ~RACK_ACKED;
7188                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7189         }
7190         if (rsm->r_in_tmap) {
7191                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7192                 rsm->r_in_tmap = 0;
7193         }
7194         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7195         rsm->r_in_tmap = 1;
7196         if (rsm->r_flags & RACK_SACK_PASSED) {
7197                 /* We have retransmitted due to the SACK pass */
7198                 rsm->r_flags &= ~RACK_SACK_PASSED;
7199                 rsm->r_flags |= RACK_WAS_SACKPASS;
7200         }
7201 }
7202
7203 static uint32_t
7204 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7205     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7206 {
7207         /*
7208          * We (re-)transmitted starting at rsm->r_start for some length
7209          * (possibly less than r_end.
7210          */
7211         struct rack_sendmap *nrsm, *insret;
7212         uint32_t c_end;
7213         int32_t len;
7214
7215         len = *lenp;
7216         c_end = rsm->r_start + len;
7217         if (SEQ_GEQ(c_end, rsm->r_end)) {
7218                 /*
7219                  * We retransmitted the whole piece or more than the whole
7220                  * slopping into the next rsm.
7221                  */
7222                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7223                 if (c_end == rsm->r_end) {
7224                         *lenp = 0;
7225                         return (0);
7226                 } else {
7227                         int32_t act_len;
7228
7229                         /* Hangs over the end return whats left */
7230                         act_len = rsm->r_end - rsm->r_start;
7231                         *lenp = (len - act_len);
7232                         return (rsm->r_end);
7233                 }
7234                 /* We don't get out of this block. */
7235         }
7236         /*
7237          * Here we retransmitted less than the whole thing which means we
7238          * have to split this into what was transmitted and what was not.
7239          */
7240         nrsm = rack_alloc_full_limit(rack);
7241         if (nrsm == NULL) {
7242                 /*
7243                  * We can't get memory, so lets not proceed.
7244                  */
7245                 *lenp = 0;
7246                 return (0);
7247         }
7248         /*
7249          * So here we are going to take the original rsm and make it what we
7250          * retransmitted. nrsm will be the tail portion we did not
7251          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7252          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7253          * 1, 6 and the new piece will be 6, 11.
7254          */
7255         rack_clone_rsm(rack, nrsm, rsm, c_end);
7256         nrsm->r_dupack = 0;
7257         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7258         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7259 #ifdef INVARIANTS
7260         if (insret != NULL) {
7261                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7262                       nrsm, insret, rack, rsm);
7263         }
7264 #endif
7265         if (rsm->r_in_tmap) {
7266                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7267                 nrsm->r_in_tmap = 1;
7268         }
7269         rsm->r_flags &= (~RACK_HAS_FIN);
7270         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7271         /* Log a split of rsm into rsm and nrsm */
7272         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7273         *lenp = 0;
7274         return (0);
7275 }
7276
7277 static void
7278 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7279                 uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts,
7280                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7281 {
7282         struct tcp_rack *rack;
7283         struct rack_sendmap *rsm, *nrsm, *insret, fe;
7284         register uint32_t snd_max, snd_una;
7285
7286         /*
7287          * Add to the RACK log of packets in flight or retransmitted. If
7288          * there is a TS option we will use the TS echoed, if not we will
7289          * grab a TS.
7290          *
7291          * Retransmissions will increment the count and move the ts to its
7292          * proper place. Note that if options do not include TS's then we
7293          * won't be able to effectively use the ACK for an RTT on a retran.
7294          *
7295          * Notes about r_start and r_end. Lets consider a send starting at
7296          * sequence 1 for 10 bytes. In such an example the r_start would be
7297          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7298          * This means that r_end is actually the first sequence for the next
7299          * slot (11).
7300          *
7301          */
7302         /*
7303          * If err is set what do we do XXXrrs? should we not add the thing?
7304          * -- i.e. return if err != 0 or should we pretend we sent it? --
7305          * i.e. proceed with add ** do this for now.
7306          */
7307         INP_WLOCK_ASSERT(tp->t_inpcb);
7308         if (err)
7309                 /*
7310                  * We don't log errors -- we could but snd_max does not
7311                  * advance in this case either.
7312                  */
7313                 return;
7314
7315         if (th_flags & TH_RST) {
7316                 /*
7317                  * We don't log resets and we return immediately from
7318                  * sending
7319                  */
7320                 return;
7321         }
7322         rack = (struct tcp_rack *)tp->t_fb_ptr;
7323         snd_una = tp->snd_una;
7324         snd_max = tp->snd_max;
7325         if (th_flags & (TH_SYN | TH_FIN)) {
7326                 /*
7327                  * The call to rack_log_output is made before bumping
7328                  * snd_max. This means we can record one extra byte on a SYN
7329                  * or FIN if seq_out is adding more on and a FIN is present
7330                  * (and we are not resending).
7331                  */
7332                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7333                         len++;
7334                 if (th_flags & TH_FIN)
7335                         len++;
7336                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7337                         /*
7338                          * The add/update as not been done for the FIN/SYN
7339                          * yet.
7340                          */
7341                         snd_max = tp->snd_nxt;
7342                 }
7343         }
7344         if (SEQ_LEQ((seq_out + len), snd_una)) {
7345                 /* Are sending an old segment to induce an ack (keep-alive)? */
7346                 return;
7347         }
7348         if (SEQ_LT(seq_out, snd_una)) {
7349                 /* huh? should we panic? */
7350                 uint32_t end;
7351
7352                 end = seq_out + len;
7353                 seq_out = snd_una;
7354                 if (SEQ_GEQ(end, seq_out))
7355                         len = end - seq_out;
7356                 else
7357                         len = 0;
7358         }
7359         if (len == 0) {
7360                 /* We don't log zero window probes */
7361                 return;
7362         }
7363         rack->r_ctl.rc_time_last_sent = cts;
7364         if (IN_FASTRECOVERY(tp->t_flags)) {
7365                 rack->r_ctl.rc_prr_out += len;
7366         }
7367         /* First question is it a retransmission or new? */
7368         if (seq_out == snd_max) {
7369                 /* Its new */
7370 again:
7371                 rsm = rack_alloc(rack);
7372                 if (rsm == NULL) {
7373                         /*
7374                          * Hmm out of memory and the tcb got destroyed while
7375                          * we tried to wait.
7376                          */
7377                         return;
7378                 }
7379                 if (th_flags & TH_FIN) {
7380                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7381                 } else {
7382                         rsm->r_flags = add_flag;
7383                 }
7384                 if (hw_tls)
7385                         rsm->r_hw_tls = 1;
7386                 rsm->r_tim_lastsent[0] = cts;
7387                 rsm->r_rtr_cnt = 1;
7388                 rsm->r_rtr_bytes = 0;
7389                 if (th_flags & TH_SYN) {
7390                         /* The data space is one beyond snd_una */
7391                         rsm->r_flags |= RACK_HAS_SYN;
7392                 }
7393                 rsm->r_start = seq_out;
7394                 rsm->r_end = rsm->r_start + len;
7395                 rsm->r_dupack = 0;
7396                 /*
7397                  * save off the mbuf location that
7398                  * sndmbuf_noadv returned (which is
7399                  * where we started copying from)..
7400                  */
7401                 rsm->m = s_mb;
7402                 rsm->soff = s_moff;
7403                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7404                 if (rsm->m) {
7405                         if (rsm->m->m_len <= rsm->soff) {
7406                                 /*
7407                                  * XXXrrs Question, will this happen?
7408                                  *
7409                                  * If sbsndptr is set at the correct place
7410                                  * then s_moff should always be somewhere
7411                                  * within rsm->m. But if the sbsndptr was
7412                                  * off then that won't be true. If it occurs
7413                                  * we need to walkout to the correct location.
7414                                  */
7415                                 struct mbuf *lm;
7416
7417                                 lm = rsm->m;
7418                                 while (lm->m_len <= rsm->soff) {
7419                                         rsm->soff -= lm->m_len;
7420                                         lm = lm->m_next;
7421                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7422                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7423                                 }
7424                                 rsm->m = lm;
7425                                 counter_u64_add(rack_sbsndptr_wrong, 1);
7426                         } else
7427                                 counter_u64_add(rack_sbsndptr_right, 1);
7428                         rsm->orig_m_len = rsm->m->m_len;
7429                 } else
7430                         rsm->orig_m_len = 0;
7431                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7432                 /* Log a new rsm */
7433                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7434                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7435 #ifdef INVARIANTS
7436                 if (insret != NULL) {
7437                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7438                               nrsm, insret, rack, rsm);
7439                 }
7440 #endif
7441                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7442                 rsm->r_in_tmap = 1;
7443                 /*
7444                  * Special case detection, is there just a single
7445                  * packet outstanding when we are not in recovery?
7446                  *
7447                  * If this is true mark it so.
7448                  */
7449                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7450                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7451                         struct rack_sendmap *prsm;
7452
7453                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7454                         if (prsm)
7455                                 prsm->r_one_out_nr = 1;
7456                 }
7457                 return;
7458         }
7459         /*
7460          * If we reach here its a retransmission and we need to find it.
7461          */
7462         memset(&fe, 0, sizeof(fe));
7463 more:
7464         if (hintrsm && (hintrsm->r_start == seq_out)) {
7465                 rsm = hintrsm;
7466                 hintrsm = NULL;
7467         } else {
7468                 /* No hints sorry */
7469                 rsm = NULL;
7470         }
7471         if ((rsm) && (rsm->r_start == seq_out)) {
7472                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7473                 if (len == 0) {
7474                         return;
7475                 } else {
7476                         goto more;
7477                 }
7478         }
7479         /* Ok it was not the last pointer go through it the hard way. */
7480 refind:
7481         fe.r_start = seq_out;
7482         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7483         if (rsm) {
7484                 if (rsm->r_start == seq_out) {
7485                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7486                         if (len == 0) {
7487                                 return;
7488                         } else {
7489                                 goto refind;
7490                         }
7491                 }
7492                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7493                         /* Transmitted within this piece */
7494                         /*
7495                          * Ok we must split off the front and then let the
7496                          * update do the rest
7497                          */
7498                         nrsm = rack_alloc_full_limit(rack);
7499                         if (nrsm == NULL) {
7500                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7501                                 return;
7502                         }
7503                         /*
7504                          * copy rsm to nrsm and then trim the front of rsm
7505                          * to not include this part.
7506                          */
7507                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7508                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7509                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7510 #ifdef INVARIANTS
7511                         if (insret != NULL) {
7512                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7513                                       nrsm, insret, rack, rsm);
7514                         }
7515 #endif
7516                         if (rsm->r_in_tmap) {
7517                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7518                                 nrsm->r_in_tmap = 1;
7519                         }
7520                         rsm->r_flags &= (~RACK_HAS_FIN);
7521                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7522                         if (len == 0) {
7523                                 return;
7524                         } else if (len > 0)
7525                                 goto refind;
7526                 }
7527         }
7528         /*
7529          * Hmm not found in map did they retransmit both old and on into the
7530          * new?
7531          */
7532         if (seq_out == tp->snd_max) {
7533                 goto again;
7534         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7535 #ifdef INVARIANTS
7536                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7537                        seq_out, len, tp->snd_una, tp->snd_max);
7538                 printf("Starting Dump of all rack entries\n");
7539                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7540                         printf("rsm:%p start:%u end:%u\n",
7541                                rsm, rsm->r_start, rsm->r_end);
7542                 }
7543                 printf("Dump complete\n");
7544                 panic("seq_out not found rack:%p tp:%p",
7545                       rack, tp);
7546 #endif
7547         } else {
7548 #ifdef INVARIANTS
7549                 /*
7550                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7551                  * flag)
7552                  */
7553                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7554                       seq_out, len, tp->snd_max, tp);
7555 #endif
7556         }
7557 }
7558
7559 /*
7560  * Record one of the RTT updates from an ack into
7561  * our sample structure.
7562  */
7563
7564 static void
7565 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7566                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7567 {
7568         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7569             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7570                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7571         }
7572         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7573             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7574                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7575         }
7576         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7577             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7578                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7579             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7580                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7581         }
7582         if ((confidence == 1) &&
7583             ((rsm == NULL) ||
7584              (rsm->r_just_ret) ||
7585              (rsm->r_one_out_nr &&
7586               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7587                 /*
7588                  * If the rsm had a just return
7589                  * hit it then we can't trust the
7590                  * rtt measurement for buffer deterimination
7591                  * Note that a confidence of 2, indicates
7592                  * SACK'd which overrides the r_just_ret or
7593                  * the r_one_out_nr. If it was a CUM-ACK and
7594                  * we had only two outstanding, but get an
7595                  * ack for only 1. Then that also lowers our
7596                  * confidence.
7597                  */
7598                 confidence = 0;
7599         }
7600         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7601             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7602                 if (rack->r_ctl.rack_rs.confidence == 0) {
7603                         /*
7604                          * We take anything with no current confidence
7605                          * saved.
7606                          */
7607                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7608                         rack->r_ctl.rack_rs.confidence = confidence;
7609                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7610                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7611                         /*
7612                          * Once we have a confident number,
7613                          * we can update it with a smaller
7614                          * value since this confident number
7615                          * may include the DSACK time until
7616                          * the next segment (the second one) arrived.
7617                          */
7618                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7619                         rack->r_ctl.rack_rs.confidence = confidence;
7620                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7621                 }
7622         }
7623         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7624         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7625         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7626         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7627 }
7628
7629 /*
7630  * Collect new round-trip time estimate
7631  * and update averages and current timeout.
7632  */
7633 static void
7634 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7635 {
7636         int32_t delta;
7637         uint32_t o_srtt, o_var;
7638         int32_t hrtt_up = 0;
7639         int32_t rtt;
7640
7641         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7642                 /* No valid sample */
7643                 return;
7644         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7645                 /* We are to use the lowest RTT seen in a single ack */
7646                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7647         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7648                 /* We are to use the highest RTT seen in a single ack */
7649                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7650         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7651                 /* We are to use the average RTT seen in a single ack */
7652                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7653                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7654         } else {
7655 #ifdef INVARIANTS
7656                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7657 #endif
7658                 return;
7659         }
7660         if (rtt == 0)
7661                 rtt = 1;
7662         if (rack->rc_gp_rtt_set == 0) {
7663                 /*
7664                  * With no RTT we have to accept
7665                  * even one we are not confident of.
7666                  */
7667                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7668                 rack->rc_gp_rtt_set = 1;
7669         } else if (rack->r_ctl.rack_rs.confidence) {
7670                 /* update the running gp srtt */
7671                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7672                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7673         }
7674         if (rack->r_ctl.rack_rs.confidence) {
7675                 /*
7676                  * record the low and high for highly buffered path computation,
7677                  * we only do this if we are confident (not a retransmission).
7678                  */
7679                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7680                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7681                         hrtt_up = 1;
7682                 }
7683                 if (rack->rc_highly_buffered == 0) {
7684                         /*
7685                          * Currently once we declare a path has
7686                          * highly buffered there is no going
7687                          * back, which may be a problem...
7688                          */
7689                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7690                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7691                                                      rack->r_ctl.rc_highest_us_rtt,
7692                                                      rack->r_ctl.rc_lowest_us_rtt,
7693                                                      RACK_RTTS_SEEHBP);
7694                                 rack->rc_highly_buffered = 1;
7695                         }
7696                 }
7697         }
7698         if ((rack->r_ctl.rack_rs.confidence) ||
7699             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7700                 /*
7701                  * If we are highly confident of it <or> it was
7702                  * never retransmitted we accept it as the last us_rtt.
7703                  */
7704                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7705                 /* The lowest rtt can be set if its was not retransmited */
7706                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7707                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7708                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7709                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7710                 }
7711         }
7712         o_srtt = tp->t_srtt;
7713         o_var = tp->t_rttvar;
7714         rack = (struct tcp_rack *)tp->t_fb_ptr;
7715         if (tp->t_srtt != 0) {
7716                 /*
7717                  * We keep a simple srtt in microseconds, like our rtt
7718                  * measurement. We don't need to do any tricks with shifting
7719                  * etc. Instead we just add in 1/8th of the new measurement
7720                  * and subtract out 1/8 of the old srtt. We do the same with
7721                  * the variance after finding the absolute value of the
7722                  * difference between this sample and the current srtt.
7723                  */
7724                 delta = tp->t_srtt - rtt;
7725                 /* Take off 1/8th of the current sRTT */
7726                 tp->t_srtt -= (tp->t_srtt >> 3);
7727                 /* Add in 1/8th of the new RTT just measured */
7728                 tp->t_srtt += (rtt >> 3);
7729                 if (tp->t_srtt <= 0)
7730                         tp->t_srtt = 1;
7731                 /* Now lets make the absolute value of the variance */
7732                 if (delta < 0)
7733                         delta = -delta;
7734                 /* Subtract out 1/8th */
7735                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7736                 /* Add in 1/8th of the new variance we just saw */
7737                 tp->t_rttvar += (delta >> 3);
7738                 if (tp->t_rttvar <= 0)
7739                         tp->t_rttvar = 1;
7740                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7741                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7742         } else {
7743                 /*
7744                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7745                  * variance to half the rtt (so our first retransmit happens
7746                  * at 3*rtt).
7747                  */
7748                 tp->t_srtt = rtt;
7749                 tp->t_rttvar = rtt >> 1;
7750                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7751         }
7752         rack->rc_srtt_measure_made = 1;
7753         KMOD_TCPSTAT_INC(tcps_rttupdated);
7754         tp->t_rttupdated++;
7755 #ifdef STATS
7756         if (rack_stats_gets_ms_rtt == 0) {
7757                 /* Send in the microsecond rtt used for rxt timeout purposes */
7758                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7759         } else if (rack_stats_gets_ms_rtt == 1) {
7760                 /* Send in the millisecond rtt used for rxt timeout purposes */
7761                 int32_t ms_rtt;
7762
7763                 /* Round up */
7764                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7765                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7766         } else if (rack_stats_gets_ms_rtt == 2) {
7767                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7768                 int32_t ms_rtt;
7769
7770                 /* Round up */
7771                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7772                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7773         }  else {
7774                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7775                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7776         }
7777
7778 #endif
7779         /*
7780          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7781          * way we do the smoothing, srtt and rttvar will each average +1/2
7782          * tick of bias.  When we compute the retransmit timer, we want 1/2
7783          * tick of rounding and 1 extra tick because of +-1/2 tick
7784          * uncertainty in the firing of the timer.  The bias will give us
7785          * exactly the 1.5 tick we need.  But, because the bias is
7786          * statistical, we have to test that we don't drop below the minimum
7787          * feasible timer (which is 2 ticks).
7788          */
7789         tp->t_rxtshift = 0;
7790         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7791                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7792         rack_log_rtt_sample(rack, rtt);
7793         tp->t_softerror = 0;
7794 }
7795
7796
7797 static void
7798 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7799 {
7800         /*
7801          * Apply to filter the inbound us-rtt at us_cts.
7802          */
7803         uint32_t old_rtt;
7804
7805         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7806         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7807                                us_rtt, us_cts);
7808         if (rack->r_ctl.last_pacing_time &&
7809             rack->rc_gp_dyn_mul &&
7810             (rack->r_ctl.last_pacing_time > us_rtt))
7811                 rack->pacing_longer_than_rtt = 1;
7812         else
7813                 rack->pacing_longer_than_rtt = 0;
7814         if (old_rtt > us_rtt) {
7815                 /* We just hit a new lower rtt time */
7816                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7817                                      __LINE__, RACK_RTTS_NEWRTT);
7818                 /*
7819                  * Only count it if its lower than what we saw within our
7820                  * calculated range.
7821                  */
7822                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7823                         if (rack_probertt_lower_within &&
7824                             rack->rc_gp_dyn_mul &&
7825                             (rack->use_fixed_rate == 0) &&
7826                             (rack->rc_always_pace)) {
7827                                 /*
7828                                  * We are seeing a new lower rtt very close
7829                                  * to the time that we would have entered probe-rtt.
7830                                  * This is probably due to the fact that a peer flow
7831                                  * has entered probe-rtt. Lets go in now too.
7832                                  */
7833                                 uint32_t val;
7834
7835                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7836                                 val /= 100;
7837                                 if ((rack->in_probe_rtt == 0)  &&
7838                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7839                                         rack_enter_probertt(rack, us_cts);
7840                                 }
7841                         }
7842                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7843                 }
7844         }
7845 }
7846
7847 static int
7848 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7849     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7850 {
7851         int32_t i, all;
7852         uint32_t t, len_acked;
7853
7854         if ((rsm->r_flags & RACK_ACKED) ||
7855             (rsm->r_flags & RACK_WAS_ACKED))
7856                 /* Already done */
7857                 return (0);
7858         if (rsm->r_no_rtt_allowed) {
7859                 /* Not allowed */
7860                 return (0);
7861         }
7862         if (ack_type == CUM_ACKED) {
7863                 if (SEQ_GT(th_ack, rsm->r_end)) {
7864                         len_acked = rsm->r_end - rsm->r_start;
7865                         all = 1;
7866                 } else {
7867                         len_acked = th_ack - rsm->r_start;
7868                         all = 0;
7869                 }
7870         } else {
7871                 len_acked = rsm->r_end - rsm->r_start;
7872                 all = 0;
7873         }
7874         if (rsm->r_rtr_cnt == 1) {
7875                 uint32_t us_rtt;
7876
7877                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7878                 if ((int)t <= 0)
7879                         t = 1;
7880                 if (!tp->t_rttlow || tp->t_rttlow > t)
7881                         tp->t_rttlow = t;
7882                 if (!rack->r_ctl.rc_rack_min_rtt ||
7883                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7884                         rack->r_ctl.rc_rack_min_rtt = t;
7885                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7886                                 rack->r_ctl.rc_rack_min_rtt = 1;
7887                         }
7888                 }
7889                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7890                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7891                 else
7892                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7893                 if (us_rtt == 0)
7894                         us_rtt = 1;
7895                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7896                 if (ack_type == SACKED) {
7897                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7898                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7899                 } else {
7900                         /*
7901                          * We need to setup what our confidence
7902                          * is in this ack.
7903                          *
7904                          * If the rsm was app limited and it is
7905                          * less than a mss in length (the end
7906                          * of the send) then we have a gap. If we
7907                          * were app limited but say we were sending
7908                          * multiple MSS's then we are more confident
7909                          * int it.
7910                          *
7911                          * When we are not app-limited then we see if
7912                          * the rsm is being included in the current
7913                          * measurement, we tell this by the app_limited_needs_set
7914                          * flag.
7915                          *
7916                          * Note that being cwnd blocked is not applimited
7917                          * as well as the pacing delay between packets which
7918                          * are sending only 1 or 2 MSS's also will show up
7919                          * in the RTT. We probably need to examine this algorithm
7920                          * a bit more and enhance it to account for the delay
7921                          * between rsm's. We could do that by saving off the
7922                          * pacing delay of each rsm (in an rsm) and then
7923                          * factoring that in somehow though for now I am
7924                          * not sure how :)
7925                          */
7926                         int calc_conf = 0;
7927
7928                         if (rsm->r_flags & RACK_APP_LIMITED) {
7929                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7930                                         calc_conf = 0;
7931                                 else
7932                                         calc_conf = 1;
7933                         } else if (rack->app_limited_needs_set == 0) {
7934                                 calc_conf = 1;
7935                         } else {
7936                                 calc_conf = 0;
7937                         }
7938                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7939                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7940                                             calc_conf, rsm, rsm->r_rtr_cnt);
7941                 }
7942                 if ((rsm->r_flags & RACK_TLP) &&
7943                     (!IN_FASTRECOVERY(tp->t_flags))) {
7944                         /* Segment was a TLP and our retrans matched */
7945                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7946                                 rack->r_ctl.rc_rsm_start = tp->snd_max;
7947                                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
7948                                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
7949                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
7950                         }
7951                 }
7952                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7953                         /* New more recent rack_tmit_time */
7954                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7955                         rack->rc_rack_rtt = t;
7956                 }
7957                 return (1);
7958         }
7959         /*
7960          * We clear the soft/rxtshift since we got an ack.
7961          * There is no assurance we will call the commit() function
7962          * so we need to clear these to avoid incorrect handling.
7963          */
7964         tp->t_rxtshift = 0;
7965         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7966                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7967         tp->t_softerror = 0;
7968         if (to && (to->to_flags & TOF_TS) &&
7969             (ack_type == CUM_ACKED) &&
7970             (to->to_tsecr) &&
7971             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7972                 /*
7973                  * Now which timestamp does it match? In this block the ACK
7974                  * must be coming from a previous transmission.
7975                  */
7976                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7977                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7978                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7979                                 if ((int)t <= 0)
7980                                         t = 1;
7981                                 if ((i + 1) < rsm->r_rtr_cnt) {
7982                                         /*
7983                                          * The peer ack'd from our previous
7984                                          * transmission. We have a spurious
7985                                          * retransmission and thus we dont
7986                                          * want to update our rack_rtt.
7987                                          */
7988                                         return (0);
7989                                 }
7990                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7991                                         tp->t_rttlow = t;
7992                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7993                                         rack->r_ctl.rc_rack_min_rtt = t;
7994                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7995                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7996                                         }
7997                                 }
7998                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7999                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
8000                                         /* New more recent rack_tmit_time */
8001                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8002                                         rack->rc_rack_rtt = t;
8003                                 }
8004                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
8005                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
8006                                                     rsm->r_rtr_cnt);
8007                                 return (1);
8008                         }
8009                 }
8010                 goto ts_not_found;
8011         } else {
8012                 /*
8013                  * Ok its a SACK block that we retransmitted. or a windows
8014                  * machine without timestamps. We can tell nothing from the
8015                  * time-stamp since its not there or the time the peer last
8016                  * recieved a segment that moved forward its cum-ack point.
8017                  */
8018 ts_not_found:
8019                 i = rsm->r_rtr_cnt - 1;
8020                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8021                 if ((int)t <= 0)
8022                         t = 1;
8023                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8024                         /*
8025                          * We retransmitted and the ack came back in less
8026                          * than the smallest rtt we have observed. We most
8027                          * likely did an improper retransmit as outlined in
8028                          * 6.2 Step 2 point 2 in the rack-draft so we
8029                          * don't want to update our rack_rtt. We in
8030                          * theory (in future) might want to think about reverting our
8031                          * cwnd state but we won't for now.
8032                          */
8033                         return (0);
8034                 } else if (rack->r_ctl.rc_rack_min_rtt) {
8035                         /*
8036                          * We retransmitted it and the retransmit did the
8037                          * job.
8038                          */
8039                         if (!rack->r_ctl.rc_rack_min_rtt ||
8040                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8041                                 rack->r_ctl.rc_rack_min_rtt = t;
8042                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
8043                                         rack->r_ctl.rc_rack_min_rtt = 1;
8044                                 }
8045                         }
8046                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
8047                                 /* New more recent rack_tmit_time */
8048                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
8049                                 rack->rc_rack_rtt = t;
8050                         }
8051                         return (1);
8052                 }
8053         }
8054         return (0);
8055 }
8056
8057 /*
8058  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
8059  */
8060 static void
8061 rack_log_sack_passed(struct tcpcb *tp,
8062     struct tcp_rack *rack, struct rack_sendmap *rsm)
8063 {
8064         struct rack_sendmap *nrsm;
8065
8066         nrsm = rsm;
8067         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
8068             rack_head, r_tnext) {
8069                 if (nrsm == rsm) {
8070                         /* Skip orginal segment he is acked */
8071                         continue;
8072                 }
8073                 if (nrsm->r_flags & RACK_ACKED) {
8074                         /*
8075                          * Skip ack'd segments, though we
8076                          * should not see these, since tmap
8077                          * should not have ack'd segments.
8078                          */
8079                         continue;
8080                 }
8081                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8082                         /*
8083                          * We found one that is already marked
8084                          * passed, we have been here before and
8085                          * so all others below this are marked.
8086                          */
8087                         break;
8088                 }
8089                 nrsm->r_flags |= RACK_SACK_PASSED;
8090                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8091         }
8092 }
8093
8094 static void
8095 rack_need_set_test(struct tcpcb *tp,
8096                    struct tcp_rack *rack,
8097                    struct rack_sendmap *rsm,
8098                    tcp_seq th_ack,
8099                    int line,
8100                    int use_which)
8101 {
8102
8103         if ((tp->t_flags & TF_GPUTINPROG) &&
8104             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8105                 /*
8106                  * We were app limited, and this ack
8107                  * butts up or goes beyond the point where we want
8108                  * to start our next measurement. We need
8109                  * to record the new gput_ts as here and
8110                  * possibly update the start sequence.
8111                  */
8112                 uint32_t seq, ts;
8113
8114                 if (rsm->r_rtr_cnt > 1) {
8115                         /*
8116                          * This is a retransmit, can we
8117                          * really make any assessment at this
8118                          * point?  We are not really sure of
8119                          * the timestamp, is it this or the
8120                          * previous transmission?
8121                          *
8122                          * Lets wait for something better that
8123                          * is not retransmitted.
8124                          */
8125                         return;
8126                 }
8127                 seq = tp->gput_seq;
8128                 ts = tp->gput_ts;
8129                 rack->app_limited_needs_set = 0;
8130                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8131                 /* Do we start at a new end? */
8132                 if ((use_which == RACK_USE_BEG) &&
8133                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8134                         /*
8135                          * When we get an ACK that just eats
8136                          * up some of the rsm, we set RACK_USE_BEG
8137                          * since whats at r_start (i.e. th_ack)
8138                          * is left unacked and thats where the
8139                          * measurement not starts.
8140                          */
8141                         tp->gput_seq = rsm->r_start;
8142                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8143                 }
8144                 if ((use_which == RACK_USE_END) &&
8145                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8146                             /*
8147                              * We use the end when the cumack
8148                              * is moving forward and completely
8149                              * deleting the rsm passed so basically
8150                              * r_end holds th_ack.
8151                              *
8152                              * For SACK's we also want to use the end
8153                              * since this piece just got sacked and
8154                              * we want to target anything after that
8155                              * in our measurement.
8156                              */
8157                             tp->gput_seq = rsm->r_end;
8158                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8159                 }
8160                 if (use_which == RACK_USE_END_OR_THACK) {
8161                         /*
8162                          * special case for ack moving forward,
8163                          * not a sack, we need to move all the
8164                          * way up to where this ack cum-ack moves
8165                          * to.
8166                          */
8167                         if (SEQ_GT(th_ack, rsm->r_end))
8168                                 tp->gput_seq = th_ack;
8169                         else
8170                                 tp->gput_seq = rsm->r_end;
8171                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8172                 }
8173                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8174                         /*
8175                          * We moved beyond this guy's range, re-calculate
8176                          * the new end point.
8177                          */
8178                         if (rack->rc_gp_filled == 0) {
8179                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8180                         } else {
8181                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8182                         }
8183                 }
8184                 /*
8185                  * We are moving the goal post, we may be able to clear the
8186                  * measure_saw_probe_rtt flag.
8187                  */
8188                 if ((rack->in_probe_rtt == 0) &&
8189                     (rack->measure_saw_probe_rtt) &&
8190                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8191                         rack->measure_saw_probe_rtt = 0;
8192                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8193                                            seq, tp->gput_seq, 0, 5, line, NULL, 0);
8194                 if (rack->rc_gp_filled &&
8195                     ((tp->gput_ack - tp->gput_seq) <
8196                      max(rc_init_window(rack), (MIN_GP_WIN *
8197                                                 ctf_fixed_maxseg(tp))))) {
8198                         uint32_t ideal_amount;
8199
8200                         ideal_amount = rack_get_measure_window(tp, rack);
8201                         if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8202                                 /*
8203                                  * There is no sense of continuing this measurement
8204                                  * because its too small to gain us anything we
8205                                  * trust. Skip it and that way we can start a new
8206                                  * measurement quicker.
8207                                  */
8208                                 tp->t_flags &= ~TF_GPUTINPROG;
8209                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8210                                                            0, 0, 0, 6, __LINE__, NULL, 0);
8211                         } else {
8212                                 /*
8213                                  * Reset the window further out.
8214                                  */
8215                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8216                         }
8217                 }
8218         }
8219 }
8220
8221 static uint32_t
8222 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8223                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8224 {
8225         uint32_t start, end, changed = 0;
8226         struct rack_sendmap stack_map;
8227         struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
8228         int32_t used_ref = 1;
8229         int moved = 0;
8230
8231         start = sack->start;
8232         end = sack->end;
8233         rsm = *prsm;
8234         memset(&fe, 0, sizeof(fe));
8235 do_rest_ofb:
8236         if ((rsm == NULL) ||
8237             (SEQ_LT(end, rsm->r_start)) ||
8238             (SEQ_GEQ(start, rsm->r_end)) ||
8239             (SEQ_LT(start, rsm->r_start))) {
8240                 /*
8241                  * We are not in the right spot,
8242                  * find the correct spot in the tree.
8243                  */
8244                 used_ref = 0;
8245                 fe.r_start = start;
8246                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8247                 moved++;
8248         }
8249         if (rsm == NULL) {
8250                 /* TSNH */
8251                 goto out;
8252         }
8253         /* Ok we have an ACK for some piece of this rsm */
8254         if (rsm->r_start != start) {
8255                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8256                         /**
8257                          * Need to split this in two pieces the before and after,
8258                          * the before remains in the map, the after must be
8259                          * added. In other words we have:
8260                          * rsm        |--------------|
8261                          * sackblk        |------->
8262                          * rsm will become
8263                          *     rsm    |---|
8264                          * and nrsm will be  the sacked piece
8265                          *     nrsm       |----------|
8266                          *
8267                          * But before we start down that path lets
8268                          * see if the sack spans over on top of
8269                          * the next guy and it is already sacked.
8270                          */
8271                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8272                         if (next && (next->r_flags & RACK_ACKED) &&
8273                             SEQ_GEQ(end, next->r_start)) {
8274                                 /**
8275                                  * So the next one is already acked, and
8276                                  * we can thus by hookery use our stack_map
8277                                  * to reflect the piece being sacked and
8278                                  * then adjust the two tree entries moving
8279                                  * the start and ends around. So we start like:
8280                                  *  rsm     |------------|             (not-acked)
8281                                  *  next                 |-----------| (acked)
8282                                  *  sackblk        |-------->
8283                                  *  We want to end like so:
8284                                  *  rsm     |------|                   (not-acked)
8285                                  *  next           |-----------------| (acked)
8286                                  *  nrsm           |-----|
8287                                  * Where nrsm is a temporary stack piece we
8288                                  * use to update all the gizmos.
8289                                  */
8290                                 /* Copy up our fudge block */
8291                                 nrsm = &stack_map;
8292                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8293                                 /* Now adjust our tree blocks */
8294                                 rsm->r_end = start;
8295                                 next->r_start = start;
8296                                 /* Now we must adjust back where next->m is */
8297                                 rack_setup_offset_for_rsm(rsm, next);
8298
8299                                 /* We don't need to adjust rsm, it did not change */
8300                                 /* Clear out the dup ack count of the remainder */
8301                                 rsm->r_dupack = 0;
8302                                 rsm->r_just_ret = 0;
8303                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8304                                 /* Now lets make sure our fudge block is right */
8305                                 nrsm->r_start = start;
8306                                 /* Now lets update all the stats and such */
8307                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8308                                 if (rack->app_limited_needs_set)
8309                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8310                                 changed += (nrsm->r_end - nrsm->r_start);
8311                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8312                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8313                                         counter_u64_add(rack_reorder_seen, 1);
8314                                         rack->r_ctl.rc_reorder_ts = cts;
8315                                 }
8316                                 /*
8317                                  * Now we want to go up from rsm (the
8318                                  * one left un-acked) to the next one
8319                                  * in the tmap. We do this so when
8320                                  * we walk backwards we include marking
8321                                  * sack-passed on rsm (The one passed in
8322                                  * is skipped since it is generally called
8323                                  * on something sacked before removing it
8324                                  * from the tmap).
8325                                  */
8326                                 if (rsm->r_in_tmap) {
8327                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8328                                         /*
8329                                          * Now that we have the next
8330                                          * one walk backwards from there.
8331                                          */
8332                                         if (nrsm && nrsm->r_in_tmap)
8333                                                 rack_log_sack_passed(tp, rack, nrsm);
8334                                 }
8335                                 /* Now are we done? */
8336                                 if (SEQ_LT(end, next->r_end) ||
8337                                     (end == next->r_end)) {
8338                                         /* Done with block */
8339                                         goto out;
8340                                 }
8341                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8342                                 counter_u64_add(rack_sack_used_next_merge, 1);
8343                                 /* Postion for the next block */
8344                                 start = next->r_end;
8345                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8346                                 if (rsm == NULL)
8347                                         goto out;
8348                         } else {
8349                                 /**
8350                                  * We can't use any hookery here, so we
8351                                  * need to split the map. We enter like
8352                                  * so:
8353                                  *  rsm      |--------|
8354                                  *  sackblk       |----->
8355                                  * We will add the new block nrsm and
8356                                  * that will be the new portion, and then
8357                                  * fall through after reseting rsm. So we
8358                                  * split and look like this:
8359                                  *  rsm      |----|
8360                                  *  sackblk       |----->
8361                                  *  nrsm          |---|
8362                                  * We then fall through reseting
8363                                  * rsm to nrsm, so the next block
8364                                  * picks it up.
8365                                  */
8366                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8367                                 if (nrsm == NULL) {
8368                                         /*
8369                                          * failed XXXrrs what can we do but loose the sack
8370                                          * info?
8371                                          */
8372                                         goto out;
8373                                 }
8374                                 counter_u64_add(rack_sack_splits, 1);
8375                                 rack_clone_rsm(rack, nrsm, rsm, start);
8376                                 rsm->r_just_ret = 0;
8377                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8378 #ifdef INVARIANTS
8379                                 if (insret != NULL) {
8380                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8381                                               nrsm, insret, rack, rsm);
8382                                 }
8383 #endif
8384                                 if (rsm->r_in_tmap) {
8385                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8386                                         nrsm->r_in_tmap = 1;
8387                                 }
8388                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8389                                 rsm->r_flags &= (~RACK_HAS_FIN);
8390                                 /* Position us to point to the new nrsm that starts the sack blk */
8391                                 rsm = nrsm;
8392                         }
8393                 } else {
8394                         /* Already sacked this piece */
8395                         counter_u64_add(rack_sack_skipped_acked, 1);
8396                         moved++;
8397                         if (end == rsm->r_end) {
8398                                 /* Done with block */
8399                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8400                                 goto out;
8401                         } else if (SEQ_LT(end, rsm->r_end)) {
8402                                 /* A partial sack to a already sacked block */
8403                                 moved++;
8404                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8405                                 goto out;
8406                         } else {
8407                                 /*
8408                                  * The end goes beyond this guy
8409                                  * repostion the start to the
8410                                  * next block.
8411                                  */
8412                                 start = rsm->r_end;
8413                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8414                                 if (rsm == NULL)
8415                                         goto out;
8416                         }
8417                 }
8418         }
8419         if (SEQ_GEQ(end, rsm->r_end)) {
8420                 /**
8421                  * The end of this block is either beyond this guy or right
8422                  * at this guy. I.e.:
8423                  *  rsm ---                 |-----|
8424                  *  end                     |-----|
8425                  *  <or>
8426                  *  end                     |---------|
8427                  */
8428                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8429                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8430                         changed += (rsm->r_end - rsm->r_start);
8431                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8432                         if (rsm->r_in_tmap) /* should be true */
8433                                 rack_log_sack_passed(tp, rack, rsm);
8434                         /* Is Reordering occuring? */
8435                         if (rsm->r_flags & RACK_SACK_PASSED) {
8436                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8437                                 counter_u64_add(rack_reorder_seen, 1);
8438                                 rack->r_ctl.rc_reorder_ts = cts;
8439                         }
8440                         if (rack->app_limited_needs_set)
8441                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8442                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8443                         rsm->r_flags |= RACK_ACKED;
8444                         rsm->r_flags &= ~RACK_TLP;
8445                         if (rsm->r_in_tmap) {
8446                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8447                                 rsm->r_in_tmap = 0;
8448                         }
8449                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8450                 } else {
8451                         counter_u64_add(rack_sack_skipped_acked, 1);
8452                         moved++;
8453                 }
8454                 if (end == rsm->r_end) {
8455                         /* This block only - done, setup for next */
8456                         goto out;
8457                 }
8458                 /*
8459                  * There is more not coverend by this rsm move on
8460                  * to the next block in the RB tree.
8461                  */
8462                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8463                 start = rsm->r_end;
8464                 rsm = nrsm;
8465                 if (rsm == NULL)
8466                         goto out;
8467                 goto do_rest_ofb;
8468         }
8469         /**
8470          * The end of this sack block is smaller than
8471          * our rsm i.e.:
8472          *  rsm ---                 |-----|
8473          *  end                     |--|
8474          */
8475         if ((rsm->r_flags & RACK_ACKED) == 0) {
8476                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8477                 if (prev && (prev->r_flags & RACK_ACKED)) {
8478                         /**
8479                          * Goal, we want the right remainder of rsm to shrink
8480                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8481                          * We want to expand prev to go all the way
8482                          * to prev->r_end <- end.
8483                          * so in the tree we have before:
8484                          *   prev     |--------|         (acked)
8485                          *   rsm               |-------| (non-acked)
8486                          *   sackblk           |-|
8487                          * We churn it so we end up with
8488                          *   prev     |----------|       (acked)
8489                          *   rsm                 |-----| (non-acked)
8490                          *   nrsm              |-| (temporary)
8491                          */
8492                         nrsm = &stack_map;
8493                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8494                         prev->r_end = end;
8495                         rsm->r_start = end;
8496                         /* Now adjust nrsm (stack copy) to be
8497                          * the one that is the small
8498                          * piece that was "sacked".
8499                          */
8500                         nrsm->r_end = end;
8501                         rsm->r_dupack = 0;
8502                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8503                         /*
8504                          * Now that the rsm has had its start moved forward
8505                          * lets go ahead and get its new place in the world.
8506                          */
8507                         rack_setup_offset_for_rsm(prev, rsm);
8508                         /*
8509                          * Now nrsm is our new little piece
8510                          * that is acked (which was merged
8511                          * to prev). Update the rtt and changed
8512                          * based on that. Also check for reordering.
8513                          */
8514                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8515                         if (rack->app_limited_needs_set)
8516                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8517                         changed += (nrsm->r_end - nrsm->r_start);
8518                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8519                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8520                                 counter_u64_add(rack_reorder_seen, 1);
8521                                 rack->r_ctl.rc_reorder_ts = cts;
8522                         }
8523                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8524                         rsm = prev;
8525                         counter_u64_add(rack_sack_used_prev_merge, 1);
8526                 } else {
8527                         /**
8528                          * This is the case where our previous
8529                          * block is not acked either, so we must
8530                          * split the block in two.
8531                          */
8532                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8533                         if (nrsm == NULL) {
8534                                 /* failed rrs what can we do but loose the sack info? */
8535                                 goto out;
8536                         }
8537                         /**
8538                          * In this case nrsm becomes
8539                          * nrsm->r_start = end;
8540                          * nrsm->r_end = rsm->r_end;
8541                          * which is un-acked.
8542                          * <and>
8543                          * rsm->r_end = nrsm->r_start;
8544                          * i.e. the remaining un-acked
8545                          * piece is left on the left
8546                          * hand side.
8547                          *
8548                          * So we start like this
8549                          * rsm      |----------| (not acked)
8550                          * sackblk  |---|
8551                          * build it so we have
8552                          * rsm      |---|         (acked)
8553                          * nrsm         |------|  (not acked)
8554                          */
8555                         counter_u64_add(rack_sack_splits, 1);
8556                         rack_clone_rsm(rack, nrsm, rsm, end);
8557                         rsm->r_flags &= (~RACK_HAS_FIN);
8558                         rsm->r_just_ret = 0;
8559                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8560 #ifdef INVARIANTS
8561                         if (insret != NULL) {
8562                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8563                                       nrsm, insret, rack, rsm);
8564                         }
8565 #endif
8566                         if (rsm->r_in_tmap) {
8567                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8568                                 nrsm->r_in_tmap = 1;
8569                         }
8570                         nrsm->r_dupack = 0;
8571                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8572                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8573                         changed += (rsm->r_end - rsm->r_start);
8574                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8575                         if (rsm->r_in_tmap) /* should be true */
8576                                 rack_log_sack_passed(tp, rack, rsm);
8577                         /* Is Reordering occuring? */
8578                         if (rsm->r_flags & RACK_SACK_PASSED) {
8579                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8580                                 counter_u64_add(rack_reorder_seen, 1);
8581                                 rack->r_ctl.rc_reorder_ts = cts;
8582                         }
8583                         if (rack->app_limited_needs_set)
8584                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8585                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8586                         rsm->r_flags |= RACK_ACKED;
8587                         rsm->r_flags &= ~RACK_TLP;
8588                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8589                         if (rsm->r_in_tmap) {
8590                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8591                                 rsm->r_in_tmap = 0;
8592                         }
8593                 }
8594         } else if (start != end){
8595                 /*
8596                  * The block was already acked.
8597                  */
8598                 counter_u64_add(rack_sack_skipped_acked, 1);
8599                 moved++;
8600         }
8601 out:
8602         if (rsm && (rsm->r_flags & RACK_ACKED)) {
8603                 /*
8604                  * Now can we merge where we worked
8605                  * with either the previous or
8606                  * next block?
8607                  */
8608                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8609                 while (next) {
8610                     if (next->r_flags & RACK_ACKED) {
8611                         /* yep this and next can be merged */
8612                         rsm = rack_merge_rsm(rack, rsm, next);
8613                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8614                     } else
8615                             break;
8616                 }
8617                 /* Now what about the previous? */
8618                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8619                 while (prev) {
8620                     if (prev->r_flags & RACK_ACKED) {
8621                         /* yep the previous and this can be merged */
8622                         rsm = rack_merge_rsm(rack, prev, rsm);
8623                         prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8624                     } else
8625                             break;
8626                 }
8627         }
8628         if (used_ref == 0) {
8629                 counter_u64_add(rack_sack_proc_all, 1);
8630         } else {
8631                 counter_u64_add(rack_sack_proc_short, 1);
8632         }
8633         /* Save off the next one for quick reference. */
8634         if (rsm)
8635                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8636         else
8637                 nrsm = NULL;
8638         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8639         /* Pass back the moved. */
8640         *moved_two = moved;
8641         return (changed);
8642 }
8643
8644 static void inline
8645 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8646 {
8647         struct rack_sendmap *tmap;
8648
8649         tmap = NULL;
8650         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8651                 /* Its no longer sacked, mark it so */
8652                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8653 #ifdef INVARIANTS
8654                 if (rsm->r_in_tmap) {
8655                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8656                               rack, rsm, rsm->r_flags);
8657                 }
8658 #endif
8659                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8660                 /* Rebuild it into our tmap */
8661                 if (tmap == NULL) {
8662                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8663                         tmap = rsm;
8664                 } else {
8665                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8666                         tmap = rsm;
8667                 }
8668                 tmap->r_in_tmap = 1;
8669                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8670         }
8671         /*
8672          * Now lets possibly clear the sack filter so we start
8673          * recognizing sacks that cover this area.
8674          */
8675         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8676
8677 }
8678
8679 static void
8680 rack_do_decay(struct tcp_rack *rack)
8681 {
8682         struct timeval res;
8683
8684 #define timersub(tvp, uvp, vvp)                                         \
8685         do {                                                            \
8686                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8687                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8688                 if ((vvp)->tv_usec < 0) {                               \
8689                         (vvp)->tv_sec--;                                \
8690                         (vvp)->tv_usec += 1000000;                      \
8691                 }                                                       \
8692         } while (0)
8693
8694         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8695 #undef timersub
8696
8697         rack->r_ctl.input_pkt++;
8698         if ((rack->rc_in_persist) ||
8699             (res.tv_sec >= 1) ||
8700             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8701                 /*
8702                  * Check for decay of non-SAD,
8703                  * we want all SAD detection metrics to
8704                  * decay 1/4 per second (or more) passed.
8705                  */
8706                 uint32_t pkt_delta;
8707
8708                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8709                 /* Update our saved tracking values */
8710                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8711                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8712                 /* Now do we escape without decay? */
8713 #ifdef NETFLIX_EXP_DETECTION
8714                 if (rack->rc_in_persist ||
8715                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8716                     (pkt_delta < tcp_sad_low_pps)){
8717                         /*
8718                          * We don't decay idle connections
8719                          * or ones that have a low input pps.
8720                          */
8721                         return;
8722                 }
8723                 /* Decay the counters */
8724                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8725                                                         tcp_sad_decay_val);
8726                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8727                                                          tcp_sad_decay_val);
8728                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8729                                                                tcp_sad_decay_val);
8730                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8731                                                                 tcp_sad_decay_val);
8732 #endif
8733         }
8734 }
8735
8736 static void
8737 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8738 {
8739         struct rack_sendmap *rsm, *rm;
8740
8741         /*
8742          * The ACK point is advancing to th_ack, we must drop off
8743          * the packets in the rack log and calculate any eligble
8744          * RTT's.
8745          */
8746         rack->r_wanted_output = 1;
8747 more:
8748         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8749         if (rsm == NULL) {
8750                 if ((th_ack - 1) == tp->iss) {
8751                         /*
8752                          * For the SYN incoming case we will not
8753                          * have called tcp_output for the sending of
8754                          * the SYN, so there will be no map. All
8755                          * other cases should probably be a panic.
8756                          */
8757                         return;
8758                 }
8759                 if (tp->t_flags & TF_SENTFIN) {
8760                         /* if we sent a FIN we often will not have map */
8761                         return;
8762                 }
8763 #ifdef INVARIANTS
8764                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8765                       tp,
8766                       tp->t_state, th_ack, rack,
8767                       tp->snd_una, tp->snd_max, tp->snd_nxt);
8768 #endif
8769                 return;
8770         }
8771         if (SEQ_LT(th_ack, rsm->r_start)) {
8772                 /* Huh map is missing this */
8773 #ifdef INVARIANTS
8774                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8775                        rsm->r_start,
8776                        th_ack, tp->t_state, rack->r_state);
8777 #endif
8778                 return;
8779         }
8780         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8781         /* Now do we consume the whole thing? */
8782         if (SEQ_GEQ(th_ack, rsm->r_end)) {
8783                 /* Its all consumed. */
8784                 uint32_t left;
8785                 uint8_t newly_acked;
8786
8787                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8788                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8789                 rsm->r_rtr_bytes = 0;
8790                 /* Record the time of highest cumack sent */
8791                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8792                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8793 #ifdef INVARIANTS
8794                 if (rm != rsm) {
8795                         panic("removing head in rack:%p rsm:%p rm:%p",
8796                               rack, rsm, rm);
8797                 }
8798 #endif
8799                 if (rsm->r_in_tmap) {
8800                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8801                         rsm->r_in_tmap = 0;
8802                 }
8803                 newly_acked = 1;
8804                 if (rsm->r_flags & RACK_ACKED) {
8805                         /*
8806                          * It was acked on the scoreboard -- remove
8807                          * it from total
8808                          */
8809                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8810                         newly_acked = 0;
8811                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
8812                         /*
8813                          * There are segments ACKED on the
8814                          * scoreboard further up. We are seeing
8815                          * reordering.
8816                          */
8817                         rsm->r_flags &= ~RACK_SACK_PASSED;
8818                         counter_u64_add(rack_reorder_seen, 1);
8819                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8820                         rsm->r_flags |= RACK_ACKED;
8821                         rack->r_ctl.rc_reorder_ts = cts;
8822                         if (rack->r_ent_rec_ns) {
8823                                 /*
8824                                  * We have sent no more, and we saw an sack
8825                                  * then ack arrive.
8826                                  */
8827                                 rack->r_might_revert = 1;
8828                         }
8829                 }
8830                 if ((rsm->r_flags & RACK_TO_REXT) &&
8831                     (tp->t_flags & TF_RCVD_TSTMP) &&
8832                     (to->to_flags & TOF_TS) &&
8833                     (tp->t_flags & TF_PREVVALID)) {
8834                         /*
8835                          * We can use the timestamp to see
8836                          * if this retransmission was from the
8837                          * first transmit. If so we made a mistake.
8838                          */
8839                         tp->t_flags &= ~TF_PREVVALID;
8840                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
8841                                 /* The first transmit is what this ack is for */
8842                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack);
8843                         }
8844                 }
8845                 left = th_ack - rsm->r_end;
8846                 if (rack->app_limited_needs_set && newly_acked)
8847                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
8848                 /* Free back to zone */
8849                 rack_free(rack, rsm);
8850                 if (left) {
8851                         goto more;
8852                 }
8853                 /* Check for reneging */
8854                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8855                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
8856                         /*
8857                          * The peer has moved snd_una up to
8858                          * the edge of this send, i.e. one
8859                          * that it had previously acked. The only
8860                          * way that can be true if the peer threw
8861                          * away data (space issues) that it had
8862                          * previously sacked (else it would have
8863                          * given us snd_una up to (rsm->r_end).
8864                          * We need to undo the acked markings here.
8865                          *
8866                          * Note we have to look to make sure th_ack is
8867                          * our rsm->r_start in case we get an old ack
8868                          * where th_ack is behind snd_una.
8869                          */
8870                         rack_peer_reneges(rack, rsm, th_ack);
8871                 }
8872                 return;
8873         }
8874         if (rsm->r_flags & RACK_ACKED) {
8875                 /*
8876                  * It was acked on the scoreboard -- remove it from
8877                  * total for the part being cum-acked.
8878                  */
8879                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
8880         }
8881         /*
8882          * Clear the dup ack count for
8883          * the piece that remains.
8884          */
8885         rsm->r_dupack = 0;
8886         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8887         if (rsm->r_rtr_bytes) {
8888                 /*
8889                  * It was retransmitted adjust the
8890                  * sack holes for what was acked.
8891                  */
8892                 int ack_am;
8893
8894                 ack_am = (th_ack - rsm->r_start);
8895                 if (ack_am >= rsm->r_rtr_bytes) {
8896                         rack->r_ctl.rc_holes_rxt -= ack_am;
8897                         rsm->r_rtr_bytes -= ack_am;
8898                 }
8899         }
8900         /*
8901          * Update where the piece starts and record
8902          * the time of send of highest cumack sent.
8903          */
8904         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8905         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
8906         /* Now we need to move our offset forward too */
8907         if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
8908                 /* Fix up the orig_m_len and possibly the mbuf offset */
8909                 rack_adjust_orig_mlen(rsm);
8910         }
8911         rsm->soff += (th_ack - rsm->r_start);
8912         rsm->r_start = th_ack;
8913         /* Now do we need to move the mbuf fwd too? */
8914         if (rsm->m) {
8915                 while (rsm->soff >= rsm->m->m_len) {
8916                         rsm->soff -= rsm->m->m_len;
8917                         rsm->m = rsm->m->m_next;
8918                         KASSERT((rsm->m != NULL),
8919                                 (" nrsm:%p hit at soff:%u null m",
8920                                  rsm, rsm->soff));
8921                 }
8922                 rsm->orig_m_len = rsm->m->m_len;
8923         }
8924         if (rack->app_limited_needs_set)
8925                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
8926 }
8927
8928 static void
8929 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
8930 {
8931         struct rack_sendmap *rsm;
8932         int sack_pass_fnd = 0;
8933
8934         if (rack->r_might_revert) {
8935                 /*
8936                  * Ok we have reordering, have not sent anything, we
8937                  * might want to revert the congestion state if nothing
8938                  * further has SACK_PASSED on it. Lets check.
8939                  *
8940                  * We also get here when we have DSACKs come in for
8941                  * all the data that we FR'd. Note that a rxt or tlp
8942                  * timer clears this from happening.
8943                  */
8944
8945                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
8946                         if (rsm->r_flags & RACK_SACK_PASSED) {
8947                                 sack_pass_fnd = 1;
8948                                 break;
8949                         }
8950                 }
8951                 if (sack_pass_fnd == 0) {
8952                         /*
8953                          * We went into recovery
8954                          * incorrectly due to reordering!
8955                          */
8956                         int orig_cwnd;
8957
8958                         rack->r_ent_rec_ns = 0;
8959                         orig_cwnd = tp->snd_cwnd;
8960                         tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec;
8961                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
8962                         tp->snd_recover = tp->snd_una;
8963                         rack_log_to_prr(rack, 14, orig_cwnd);
8964                         EXIT_RECOVERY(tp->t_flags);
8965                 }
8966                 rack->r_might_revert = 0;
8967         }
8968 }
8969
8970 #ifdef NETFLIX_EXP_DETECTION
8971 static void
8972 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
8973 {
8974         if ((rack->do_detection || tcp_force_detection) &&
8975             tcp_sack_to_ack_thresh &&
8976             tcp_sack_to_move_thresh &&
8977             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
8978                 /*
8979                  * We have thresholds set to find
8980                  * possible attackers and disable sack.
8981                  * Check them.
8982                  */
8983                 uint64_t ackratio, moveratio, movetotal;
8984
8985                 /* Log detecting */
8986                 rack_log_sad(rack, 1);
8987                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
8988                 ackratio *= (uint64_t)(1000);
8989                 if (rack->r_ctl.ack_count)
8990                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
8991                 else {
8992                         /* We really should not hit here */
8993                         ackratio = 1000;
8994                 }
8995                 if ((rack->sack_attack_disable == 0) &&
8996                     (ackratio > rack_highest_sack_thresh_seen))
8997                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
8998                 movetotal = rack->r_ctl.sack_moved_extra;
8999                 movetotal += rack->r_ctl.sack_noextra_move;
9000                 moveratio = rack->r_ctl.sack_moved_extra;
9001                 moveratio *= (uint64_t)1000;
9002                 if (movetotal)
9003                         moveratio /= movetotal;
9004                 else {
9005                         /* No moves, thats pretty good */
9006                         moveratio = 0;
9007                 }
9008                 if ((rack->sack_attack_disable == 0) &&
9009                     (moveratio > rack_highest_move_thresh_seen))
9010                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
9011                 if (rack->sack_attack_disable == 0) {
9012                         if ((ackratio > tcp_sack_to_ack_thresh) &&
9013                             (moveratio > tcp_sack_to_move_thresh)) {
9014                                 /* Disable sack processing */
9015                                 rack->sack_attack_disable = 1;
9016                                 if (rack->r_rep_attack == 0) {
9017                                         rack->r_rep_attack = 1;
9018                                         counter_u64_add(rack_sack_attacks_detected, 1);
9019                                 }
9020                                 if (tcp_attack_on_turns_on_logging) {
9021                                         /*
9022                                          * Turn on logging, used for debugging
9023                                          * false positives.
9024                                          */
9025                                         rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9026                                 }
9027                                 /* Clamp the cwnd at flight size */
9028                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9029                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9030                                 rack_log_sad(rack, 2);
9031                         }
9032                 } else {
9033                         /* We are sack-disabled check for false positives */
9034                         if ((ackratio <= tcp_restoral_thresh) ||
9035                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9036                                 rack->sack_attack_disable = 0;
9037                                 rack_log_sad(rack, 3);
9038                                 /* Restart counting */
9039                                 rack->r_ctl.sack_count = 0;
9040                                 rack->r_ctl.sack_moved_extra = 0;
9041                                 rack->r_ctl.sack_noextra_move = 1;
9042                                 rack->r_ctl.ack_count = max(1,
9043                                       (bytes_this_ack / segsiz));
9044
9045                                 if (rack->r_rep_reverse == 0) {
9046                                         rack->r_rep_reverse = 1;
9047                                         counter_u64_add(rack_sack_attacks_reversed, 1);
9048                                 }
9049                                 /* Restore the cwnd */
9050                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9051                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9052                         }
9053                 }
9054         }
9055 }
9056 #endif
9057
9058 static void
9059 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9060 {
9061
9062         uint32_t am;
9063
9064         if (SEQ_GT(end, start))
9065                 am = end - start;
9066         else
9067                 am = 0;
9068         /*
9069          * We keep track of how many DSACK blocks we get
9070          * after a recovery incident.
9071          */
9072         rack->r_ctl.dsack_byte_cnt += am;
9073         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9074             rack->r_ctl.retran_during_recovery &&
9075             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9076                 /*
9077                  * False recovery most likely culprit is reordering. If
9078                  * nothing else is missing we need to revert.
9079                  */
9080                 rack->r_might_revert = 1;
9081                 rack_handle_might_revert(rack->rc_tp, rack);
9082                 rack->r_might_revert = 0;
9083                 rack->r_ctl.retran_during_recovery = 0;
9084                 rack->r_ctl.dsack_byte_cnt = 0;
9085         }
9086 }
9087
9088 static void
9089 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9090 {
9091         /* Deal with changed and PRR here (in recovery only) */
9092         uint32_t pipe, snd_una;
9093
9094         rack->r_ctl.rc_prr_delivered += changed;
9095
9096         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9097                 /*
9098                  * It is all outstanding, we are application limited
9099                  * and thus we don't need more room to send anything.
9100                  * Note we use tp->snd_una here and not th_ack because
9101                  * the data as yet not been cut from the sb.
9102                  */
9103                 rack->r_ctl.rc_prr_sndcnt = 0;
9104                 return;
9105         }
9106         /* Compute prr_sndcnt */
9107         if (SEQ_GT(tp->snd_una, th_ack)) {
9108                 snd_una = tp->snd_una;
9109         } else {
9110                 snd_una = th_ack;
9111         }
9112         pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9113         if (pipe > tp->snd_ssthresh) {
9114                 long sndcnt;
9115
9116                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9117                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9118                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9119                 else {
9120                         rack->r_ctl.rc_prr_sndcnt = 0;
9121                         rack_log_to_prr(rack, 9, 0);
9122                         sndcnt = 0;
9123                 }
9124                 sndcnt++;
9125                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9126                         sndcnt -= rack->r_ctl.rc_prr_out;
9127                 else
9128                         sndcnt = 0;
9129                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9130                 rack_log_to_prr(rack, 10, 0);
9131         } else {
9132                 uint32_t limit;
9133
9134                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9135                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9136                 else
9137                         limit = 0;
9138                 if (changed > limit)
9139                         limit = changed;
9140                 limit += ctf_fixed_maxseg(tp);
9141                 if (tp->snd_ssthresh > pipe) {
9142                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9143                         rack_log_to_prr(rack, 11, 0);
9144                 } else {
9145                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9146                         rack_log_to_prr(rack, 12, 0);
9147                 }
9148         }
9149 }
9150
9151 static void
9152 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9153 {
9154         uint32_t changed;
9155         struct tcp_rack *rack;
9156         struct rack_sendmap *rsm;
9157         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9158         register uint32_t th_ack;
9159         int32_t i, j, k, num_sack_blks = 0;
9160         uint32_t cts, acked, ack_point, sack_changed = 0;
9161         int loop_start = 0, moved_two = 0;
9162         uint32_t tsused;
9163
9164
9165         INP_WLOCK_ASSERT(tp->t_inpcb);
9166         if (th->th_flags & TH_RST) {
9167                 /* We don't log resets */
9168                 return;
9169         }
9170         rack = (struct tcp_rack *)tp->t_fb_ptr;
9171         cts = tcp_get_usecs(NULL);
9172         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9173         changed = 0;
9174         th_ack = th->th_ack;
9175         if (rack->sack_attack_disable == 0)
9176                 rack_do_decay(rack);
9177         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9178                 /*
9179                  * You only get credit for
9180                  * MSS and greater (and you get extra
9181                  * credit for larger cum-ack moves).
9182                  */
9183                 int ac;
9184
9185                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9186                 rack->r_ctl.ack_count += ac;
9187                 counter_u64_add(rack_ack_total, ac);
9188         }
9189         if (rack->r_ctl.ack_count > 0xfff00000) {
9190                 /*
9191                  * reduce the number to keep us under
9192                  * a uint32_t.
9193                  */
9194                 rack->r_ctl.ack_count /= 2;
9195                 rack->r_ctl.sack_count /= 2;
9196         }
9197         if (SEQ_GT(th_ack, tp->snd_una)) {
9198                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9199                 tp->t_acktime = ticks;
9200         }
9201         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9202                 changed = th_ack - rsm->r_start;
9203         if (changed) {
9204                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9205         }
9206         if ((to->to_flags & TOF_SACK) == 0) {
9207                 /* We are done nothing left and no sack. */
9208                 rack_handle_might_revert(tp, rack);
9209                 /*
9210                  * For cases where we struck a dup-ack
9211                  * with no SACK, add to the changes so
9212                  * PRR will work right.
9213                  */
9214                 if (dup_ack_struck && (changed == 0)) {
9215                         changed += ctf_fixed_maxseg(rack->rc_tp);
9216                 }
9217                 goto out;
9218         }
9219         /* Sack block processing */
9220         if (SEQ_GT(th_ack, tp->snd_una))
9221                 ack_point = th_ack;
9222         else
9223                 ack_point = tp->snd_una;
9224         for (i = 0; i < to->to_nsacks; i++) {
9225                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9226                       &sack, sizeof(sack));
9227                 sack.start = ntohl(sack.start);
9228                 sack.end = ntohl(sack.end);
9229                 if (SEQ_GT(sack.end, sack.start) &&
9230                     SEQ_GT(sack.start, ack_point) &&
9231                     SEQ_LT(sack.start, tp->snd_max) &&
9232                     SEQ_GT(sack.end, ack_point) &&
9233                     SEQ_LEQ(sack.end, tp->snd_max)) {
9234                         sack_blocks[num_sack_blks] = sack;
9235                         num_sack_blks++;
9236 #ifdef NETFLIX_STATS
9237                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9238                            SEQ_LEQ(sack.end, th_ack)) {
9239                         /*
9240                          * Its a D-SACK block.
9241                          */
9242                         tcp_record_dsack(sack.start, sack.end);
9243 #endif
9244                         rack_note_dsack(rack, sack.start, sack.end);
9245                 }
9246         }
9247         /*
9248          * Sort the SACK blocks so we can update the rack scoreboard with
9249          * just one pass.
9250          */
9251         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9252                                          num_sack_blks, th->th_ack);
9253         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9254         if (num_sack_blks == 0) {
9255                 /* Nothing to sack (DSACKs?) */
9256                 goto out_with_totals;
9257         }
9258         if (num_sack_blks < 2) {
9259                 /* Only one, we don't need to sort */
9260                 goto do_sack_work;
9261         }
9262         /* Sort the sacks */
9263         for (i = 0; i < num_sack_blks; i++) {
9264                 for (j = i + 1; j < num_sack_blks; j++) {
9265                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9266                                 sack = sack_blocks[i];
9267                                 sack_blocks[i] = sack_blocks[j];
9268                                 sack_blocks[j] = sack;
9269                         }
9270                 }
9271         }
9272         /*
9273          * Now are any of the sack block ends the same (yes some
9274          * implementations send these)?
9275          */
9276 again:
9277         if (num_sack_blks == 0)
9278                 goto out_with_totals;
9279         if (num_sack_blks > 1) {
9280                 for (i = 0; i < num_sack_blks; i++) {
9281                         for (j = i + 1; j < num_sack_blks; j++) {
9282                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9283                                         /*
9284                                          * Ok these two have the same end we
9285                                          * want the smallest end and then
9286                                          * throw away the larger and start
9287                                          * again.
9288                                          */
9289                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9290                                                 /*
9291                                                  * The second block covers
9292                                                  * more area use that
9293                                                  */
9294                                                 sack_blocks[i].start = sack_blocks[j].start;
9295                                         }
9296                                         /*
9297                                          * Now collapse out the dup-sack and
9298                                          * lower the count
9299                                          */
9300                                         for (k = (j + 1); k < num_sack_blks; k++) {
9301                                                 sack_blocks[j].start = sack_blocks[k].start;
9302                                                 sack_blocks[j].end = sack_blocks[k].end;
9303                                                 j++;
9304                                         }
9305                                         num_sack_blks--;
9306                                         goto again;
9307                                 }
9308                         }
9309                 }
9310         }
9311 do_sack_work:
9312         /*
9313          * First lets look to see if
9314          * we have retransmitted and
9315          * can use the transmit next?
9316          */
9317         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9318         if (rsm &&
9319             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9320             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9321                 /*
9322                  * We probably did the FR and the next
9323                  * SACK in continues as we would expect.
9324                  */
9325                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9326                 if (acked) {
9327                         rack->r_wanted_output = 1;
9328                         changed += acked;
9329                         sack_changed += acked;
9330                 }
9331                 if (num_sack_blks == 1) {
9332                         /*
9333                          * This is what we would expect from
9334                          * a normal implementation to happen
9335                          * after we have retransmitted the FR,
9336                          * i.e the sack-filter pushes down
9337                          * to 1 block and the next to be retransmitted
9338                          * is the sequence in the sack block (has more
9339                          * are acked). Count this as ACK'd data to boost
9340                          * up the chances of recovering any false positives.
9341                          */
9342                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9343                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9344                         counter_u64_add(rack_express_sack, 1);
9345                         if (rack->r_ctl.ack_count > 0xfff00000) {
9346                                 /*
9347                                  * reduce the number to keep us under
9348                                  * a uint32_t.
9349                                  */
9350                                 rack->r_ctl.ack_count /= 2;
9351                                 rack->r_ctl.sack_count /= 2;
9352                         }
9353                         goto out_with_totals;
9354                 } else {
9355                         /*
9356                          * Start the loop through the
9357                          * rest of blocks, past the first block.
9358                          */
9359                         moved_two = 0;
9360                         loop_start = 1;
9361                 }
9362         }
9363         /* Its a sack of some sort */
9364         rack->r_ctl.sack_count++;
9365         if (rack->r_ctl.sack_count > 0xfff00000) {
9366                 /*
9367                  * reduce the number to keep us under
9368                  * a uint32_t.
9369                  */
9370                 rack->r_ctl.ack_count /= 2;
9371                 rack->r_ctl.sack_count /= 2;
9372         }
9373         counter_u64_add(rack_sack_total, 1);
9374         if (rack->sack_attack_disable) {
9375                 /* An attacker disablement is in place */
9376                 if (num_sack_blks > 1) {
9377                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9378                         rack->r_ctl.sack_moved_extra++;
9379                         counter_u64_add(rack_move_some, 1);
9380                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9381                                 rack->r_ctl.sack_moved_extra /= 2;
9382                                 rack->r_ctl.sack_noextra_move /= 2;
9383                         }
9384                 }
9385                 goto out;
9386         }
9387         rsm = rack->r_ctl.rc_sacklast;
9388         for (i = loop_start; i < num_sack_blks; i++) {
9389                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9390                 if (acked) {
9391                         rack->r_wanted_output = 1;
9392                         changed += acked;
9393                         sack_changed += acked;
9394                 }
9395                 if (moved_two) {
9396                         /*
9397                          * If we did not get a SACK for at least a MSS and
9398                          * had to move at all, or if we moved more than our
9399                          * threshold, it counts against the "extra" move.
9400                          */
9401                         rack->r_ctl.sack_moved_extra += moved_two;
9402                         counter_u64_add(rack_move_some, 1);
9403                 } else {
9404                         /*
9405                          * else we did not have to move
9406                          * any more than we would expect.
9407                          */
9408                         rack->r_ctl.sack_noextra_move++;
9409                         counter_u64_add(rack_move_none, 1);
9410                 }
9411                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9412                         /*
9413                          * If the SACK was not a full MSS then
9414                          * we add to sack_count the number of
9415                          * MSS's (or possibly more than
9416                          * a MSS if its a TSO send) we had to skip by.
9417                          */
9418                         rack->r_ctl.sack_count += moved_two;
9419                         counter_u64_add(rack_sack_total, moved_two);
9420                 }
9421                 /*
9422                  * Now we need to setup for the next
9423                  * round. First we make sure we won't
9424                  * exceed the size of our uint32_t on
9425                  * the various counts, and then clear out
9426                  * moved_two.
9427                  */
9428                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9429                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9430                         rack->r_ctl.sack_moved_extra /= 2;
9431                         rack->r_ctl.sack_noextra_move /= 2;
9432                 }
9433                 if (rack->r_ctl.sack_count > 0xfff00000) {
9434                         rack->r_ctl.ack_count /= 2;
9435                         rack->r_ctl.sack_count /= 2;
9436                 }
9437                 moved_two = 0;
9438         }
9439 out_with_totals:
9440         if (num_sack_blks > 1) {
9441                 /*
9442                  * You get an extra stroke if
9443                  * you have more than one sack-blk, this
9444                  * could be where we are skipping forward
9445                  * and the sack-filter is still working, or
9446                  * it could be an attacker constantly
9447                  * moving us.
9448                  */
9449                 rack->r_ctl.sack_moved_extra++;
9450                 counter_u64_add(rack_move_some, 1);
9451         }
9452 out:
9453 #ifdef NETFLIX_EXP_DETECTION
9454         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9455 #endif
9456         if (changed) {
9457                 /* Something changed cancel the rack timer */
9458                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9459         }
9460         tsused = tcp_get_usecs(NULL);
9461         rsm = tcp_rack_output(tp, rack, tsused);
9462         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9463             rsm) {
9464                 /* Enter recovery */
9465                 rack->r_ctl.rc_rsm_start = rsm->r_start;
9466                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
9467                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
9468                 entered_recovery = 1;
9469                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9470                 /*
9471                  * When we enter recovery we need to assure we send
9472                  * one packet.
9473                  */
9474                 if (rack->rack_no_prr == 0) {
9475                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9476                         rack_log_to_prr(rack, 8, 0);
9477                 }
9478                 rack->r_timer_override = 1;
9479                 rack->r_early = 0;
9480                 rack->r_ctl.rc_agg_early = 0;
9481         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9482                    rsm &&
9483                    (rack->r_rr_config == 3)) {
9484                 /*
9485                  * Assure we can output and we get no
9486                  * remembered pace time except the retransmit.
9487                  */
9488                 rack->r_timer_override = 1;
9489                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9490                 rack->r_ctl.rc_resend = rsm;
9491         }
9492         if (IN_FASTRECOVERY(tp->t_flags) &&
9493             (rack->rack_no_prr == 0) &&
9494             (entered_recovery == 0)) {
9495                 rack_update_prr(tp, rack, changed, th_ack);
9496                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9497                      ((rack->rc_inp->inp_in_hpts == 0) &&
9498                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9499                         /*
9500                          * If you are pacing output you don't want
9501                          * to override.
9502                          */
9503                         rack->r_early = 0;
9504                         rack->r_ctl.rc_agg_early = 0;
9505                         rack->r_timer_override = 1;
9506                 }
9507         }
9508 }
9509
9510 static void
9511 rack_strike_dupack(struct tcp_rack *rack)
9512 {
9513         struct rack_sendmap *rsm;
9514
9515         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9516         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9517                 rsm = TAILQ_NEXT(rsm, r_tnext);
9518         }
9519         if (rsm && (rsm->r_dupack < 0xff)) {
9520                 rsm->r_dupack++;
9521                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9522                         struct timeval tv;
9523                         uint32_t cts;
9524                         /*
9525                          * Here we see if we need to retransmit. For
9526                          * a SACK type connection if enough time has passed
9527                          * we will get a return of the rsm. For a non-sack
9528                          * connection we will get the rsm returned if the
9529                          * dupack value is 3 or more.
9530                          */
9531                         cts = tcp_get_usecs(&tv);
9532                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9533                         if (rack->r_ctl.rc_resend != NULL) {
9534                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9535                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9536                                                          rack->rc_tp->snd_una);
9537                                 }
9538                                 rack->r_wanted_output = 1;
9539                                 rack->r_timer_override = 1;
9540                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9541                         }
9542                 } else {
9543                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9544                 }
9545         }
9546 }
9547
9548 static void
9549 rack_check_bottom_drag(struct tcpcb *tp,
9550                        struct tcp_rack *rack,
9551                        struct socket *so, int32_t acked)
9552 {
9553         uint32_t segsiz, minseg;
9554
9555         segsiz = ctf_fixed_maxseg(tp);
9556         minseg = segsiz;
9557
9558         if (tp->snd_max == tp->snd_una) {
9559                 /*
9560                  * We are doing dynamic pacing and we are way
9561                  * under. Basically everything got acked while
9562                  * we were still waiting on the pacer to expire.
9563                  *
9564                  * This means we need to boost the b/w in
9565                  * addition to any earlier boosting of
9566                  * the multipler.
9567                  */
9568                 rack->rc_dragged_bottom = 1;
9569                 rack_validate_multipliers_at_or_above100(rack);
9570                 /*
9571                  * Lets use the segment bytes acked plus
9572                  * the lowest RTT seen as the basis to
9573                  * form a b/w estimate. This will be off
9574                  * due to the fact that the true estimate
9575                  * should be around 1/2 the time of the RTT
9576                  * but we can settle for that.
9577                  */
9578                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9579                     acked) {
9580                         uint64_t bw, calc_bw, rtt;
9581
9582                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9583                         if (rtt == 0) {
9584                                 /* no us sample is there a ms one? */
9585                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9586                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9587                                 } else {
9588                                         goto no_measurement;
9589                                 }
9590                         }
9591                         bw = acked;
9592                         calc_bw = bw * 1000000;
9593                         calc_bw /= rtt;
9594                         if (rack->r_ctl.last_max_bw &&
9595                             (rack->r_ctl.last_max_bw < calc_bw)) {
9596                                 /*
9597                                  * If we have a last calculated max bw
9598                                  * enforce it.
9599                                  */
9600                                 calc_bw = rack->r_ctl.last_max_bw;
9601                         }
9602                         /* now plop it in */
9603                         if (rack->rc_gp_filled == 0) {
9604                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9605                                         /*
9606                                          * If we have no measurement
9607                                          * don't let us set in more than
9608                                          * 1.2Mbps. If we are still too
9609                                          * low after pacing with this we
9610                                          * will hopefully have a max b/w
9611                                          * available to sanity check things.
9612                                          */
9613                                         calc_bw = ONE_POINT_TWO_MEG;
9614                                 }
9615                                 rack->r_ctl.rc_rtt_diff = 0;
9616                                 rack->r_ctl.gp_bw = calc_bw;
9617                                 rack->rc_gp_filled = 1;
9618                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9619                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9620                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9621                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9622                                 rack->r_ctl.rc_rtt_diff = 0;
9623                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9624                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9625                                 rack->r_ctl.gp_bw = calc_bw;
9626                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9627                         } else
9628                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9629                         if ((rack->gp_ready == 0) &&
9630                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9631                                 /* We have enough measurements now */
9632                                 rack->gp_ready = 1;
9633                                 rack_set_cc_pacing(rack);
9634                                 if (rack->defer_options)
9635                                         rack_apply_deferred_options(rack);
9636                         }
9637                         /*
9638                          * For acks over 1mss we do a extra boost to simulate
9639                          * where we would get 2 acks (we want 110 for the mul).
9640                          */
9641                         if (acked > segsiz)
9642                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9643                 } else {
9644                         /*
9645                          * zero rtt possibly?, settle for just an old increase.
9646                          */
9647 no_measurement:
9648                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9649                 }
9650         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9651                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9652                                                minseg)) &&
9653                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9654                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9655                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9656                     (segsiz * rack_req_segs))) {
9657                 /*
9658                  * We are doing dynamic GP pacing and
9659                  * we have everything except 1MSS or less
9660                  * bytes left out. We are still pacing away.
9661                  * And there is data that could be sent, This
9662                  * means we are inserting delayed ack time in
9663                  * our measurements because we are pacing too slow.
9664                  */
9665                 rack_validate_multipliers_at_or_above100(rack);
9666                 rack->rc_dragged_bottom = 1;
9667                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9668         }
9669 }
9670
9671
9672
9673 static void
9674 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9675 {
9676         /*
9677          * The fast output path is enabled and we
9678          * have moved the cumack forward. Lets see if
9679          * we can expand forward the fast path length by
9680          * that amount. What we would ideally like to
9681          * do is increase the number of bytes in the
9682          * fast path block (left_to_send) by the
9683          * acked amount. However we have to gate that
9684          * by two factors:
9685          * 1) The amount outstanding and the rwnd of the peer
9686          *    (i.e. we don't want to exceed the rwnd of the peer).
9687          *    <and>
9688          * 2) The amount of data left in the socket buffer (i.e.
9689          *    we can't send beyond what is in the buffer).
9690          *
9691          * Note that this does not take into account any increase
9692          * in the cwnd. We will only extend the fast path by
9693          * what was acked.
9694          */
9695         uint32_t new_total, gating_val;
9696
9697         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9698         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9699                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9700         if (new_total <= gating_val) {
9701                 /* We can increase left_to_send by the acked amount */
9702                 counter_u64_add(rack_extended_rfo, 1);
9703                 rack->r_ctl.fsb.left_to_send = new_total;
9704                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9705                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
9706                          rack, rack->r_ctl.fsb.left_to_send,
9707                          sbavail(&rack->rc_inp->inp_socket->so_snd),
9708                          (tp->snd_max - tp->snd_una)));
9709
9710         }
9711 }
9712
9713 static void
9714 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9715 {
9716         /*
9717          * Here any sendmap entry that points to the
9718          * beginning mbuf must be adjusted to the correct
9719          * offset. This must be called with:
9720          * 1) The socket buffer locked
9721          * 2) snd_una adjusted to its new postion.
9722          *
9723          * Note that (2) implies rack_ack_received has also
9724          * been called.
9725          *
9726          * We grab the first mbuf in the socket buffer and
9727          * then go through the front of the sendmap, recalculating
9728          * the stored offset for any sendmap entry that has
9729          * that mbuf. We must use the sb functions to do this
9730          * since its possible an add was done has well as
9731          * the subtraction we may have just completed. This should
9732          * not be a penalty though, since we just referenced the sb
9733          * to go in and trim off the mbufs that we freed (of course
9734          * there will be a penalty for the sendmap references though).
9735          */
9736         struct mbuf *m;
9737         struct rack_sendmap *rsm;
9738
9739         SOCKBUF_LOCK_ASSERT(sb);
9740         m = sb->sb_mb;
9741         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9742         if ((rsm == NULL) || (m == NULL)) {
9743                 /* Nothing outstanding */
9744                 return;
9745         }
9746         while (rsm->m && (rsm->m == m)) {
9747                 /* one to adjust */
9748 #ifdef INVARIANTS
9749                 struct mbuf *tm;
9750                 uint32_t soff;
9751
9752                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9753                 if (rsm->orig_m_len != m->m_len) {
9754                         rack_adjust_orig_mlen(rsm);
9755                 }
9756                 if (rsm->soff != soff) {
9757                         /*
9758                          * This is not a fatal error, we anticipate it
9759                          * might happen (the else code), so we count it here
9760                          * so that under invariant we can see that it really
9761                          * does happen.
9762                          */
9763                         counter_u64_add(rack_adjust_map_bw, 1);
9764                 }
9765                 rsm->m = tm;
9766                 rsm->soff = soff;
9767                 if (tm)
9768                         rsm->orig_m_len = rsm->m->m_len;
9769                 else
9770                         rsm->orig_m_len = 0;
9771 #else
9772                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
9773                 if (rsm->m)
9774                         rsm->orig_m_len = rsm->m->m_len;
9775                 else
9776                         rsm->orig_m_len = 0;
9777 #endif
9778                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
9779                               rsm);
9780                 if (rsm == NULL)
9781                         break;
9782         }
9783 }
9784
9785 /*
9786  * Return value of 1, we do not need to call rack_process_data().
9787  * return value of 0, rack_process_data can be called.
9788  * For ret_val if its 0 the TCP is locked, if its non-zero
9789  * its unlocked and probably unsafe to touch the TCB.
9790  */
9791 static int
9792 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9793     struct tcpcb *tp, struct tcpopt *to,
9794     uint32_t tiwin, int32_t tlen,
9795     int32_t * ofia, int32_t thflags, int32_t *ret_val)
9796 {
9797         int32_t ourfinisacked = 0;
9798         int32_t nsegs, acked_amount;
9799         int32_t acked;
9800         struct mbuf *mfree;
9801         struct tcp_rack *rack;
9802         int32_t under_pacing = 0;
9803         int32_t recovery = 0;
9804
9805         rack = (struct tcp_rack *)tp->t_fb_ptr;
9806         if (SEQ_GT(th->th_ack, tp->snd_max)) {
9807                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
9808                                       &rack->r_ctl.challenge_ack_ts,
9809                                       &rack->r_ctl.challenge_ack_cnt);
9810                 rack->r_wanted_output = 1;
9811                 return (1);
9812         }
9813         if (rack->gp_ready &&
9814             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
9815                 under_pacing = 1;
9816         }
9817         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
9818                 int in_rec, dup_ack_struck = 0;
9819
9820                 in_rec = IN_FASTRECOVERY(tp->t_flags);
9821                 if (rack->rc_in_persist) {
9822                         tp->t_rxtshift = 0;
9823                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9824                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9825                 }
9826                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) {
9827                         rack_strike_dupack(rack);
9828                         dup_ack_struck = 1;
9829                 }
9830                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
9831         }
9832         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
9833                 /*
9834                  * Old ack, behind (or duplicate to) the last one rcv'd
9835                  * Note: We mark reordering is occuring if its
9836                  * less than and we have not closed our window.
9837                  */
9838                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
9839                         counter_u64_add(rack_reorder_seen, 1);
9840                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9841                 }
9842                 return (0);
9843         }
9844         /*
9845          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
9846          * something we sent.
9847          */
9848         if (tp->t_flags & TF_NEEDSYN) {
9849                 /*
9850                  * T/TCP: Connection was half-synchronized, and our SYN has
9851                  * been ACK'd (so connection is now fully synchronized).  Go
9852                  * to non-starred state, increment snd_una for ACK of SYN,
9853                  * and check if we can do window scaling.
9854                  */
9855                 tp->t_flags &= ~TF_NEEDSYN;
9856                 tp->snd_una++;
9857                 /* Do window scaling? */
9858                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9859                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9860                         tp->rcv_scale = tp->request_r_scale;
9861                         /* Send window already scaled. */
9862                 }
9863         }
9864         nsegs = max(1, m->m_pkthdr.lro_nsegs);
9865         INP_WLOCK_ASSERT(tp->t_inpcb);
9866
9867         acked = BYTES_THIS_ACK(tp, th);
9868         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
9869         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
9870         /*
9871          * If we just performed our first retransmit, and the ACK arrives
9872          * within our recovery window, then it was a mistake to do the
9873          * retransmit in the first place.  Recover our original cwnd and
9874          * ssthresh, and proceed to transmit where we left off.
9875          */
9876         if ((tp->t_flags & TF_PREVVALID) &&
9877             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
9878                 tp->t_flags &= ~TF_PREVVALID;
9879                 if (tp->t_rxtshift == 1 &&
9880                     (int)(ticks - tp->t_badrxtwin) < 0)
9881                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
9882         }
9883         if (acked) {
9884                 /* assure we are not backed off */
9885                 tp->t_rxtshift = 0;
9886                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9887                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9888                 rack->rc_tlp_in_progress = 0;
9889                 rack->r_ctl.rc_tlp_cnt_out = 0;
9890                 /*
9891                  * If it is the RXT timer we want to
9892                  * stop it, so we can restart a TLP.
9893                  */
9894                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
9895                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9896 #ifdef NETFLIX_HTTP_LOGGING
9897                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
9898 #endif
9899         }
9900         /*
9901          * If we have a timestamp reply, update smoothed round trip time. If
9902          * no timestamp is present but transmit timer is running and timed
9903          * sequence number was acked, update smoothed round trip time. Since
9904          * we now have an rtt measurement, cancel the timer backoff (cf.,
9905          * Phil Karn's retransmit alg.). Recompute the initial retransmit
9906          * timer.
9907          *
9908          * Some boxes send broken timestamp replies during the SYN+ACK
9909          * phase, ignore timestamps of 0 or we could calculate a huge RTT
9910          * and blow up the retransmit timer.
9911          */
9912         /*
9913          * If all outstanding data is acked, stop retransmit timer and
9914          * remember to restart (more output or persist). If there is more
9915          * data to be acked, restart retransmit timer, using current
9916          * (possibly backed-off) value.
9917          */
9918         if (acked == 0) {
9919                 if (ofia)
9920                         *ofia = ourfinisacked;
9921                 return (0);
9922         }
9923         if (IN_RECOVERY(tp->t_flags)) {
9924                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
9925                     (SEQ_LT(th->th_ack, tp->snd_max))) {
9926                         tcp_rack_partialack(tp);
9927                 } else {
9928                         rack_post_recovery(tp, th->th_ack);
9929                         recovery = 1;
9930                 }
9931         }
9932         /*
9933          * Let the congestion control algorithm update congestion control
9934          * related information. This typically means increasing the
9935          * congestion window.
9936          */
9937         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
9938         SOCKBUF_LOCK(&so->so_snd);
9939         acked_amount = min(acked, (int)sbavail(&so->so_snd));
9940         tp->snd_wnd -= acked_amount;
9941         mfree = sbcut_locked(&so->so_snd, acked_amount);
9942         if ((sbused(&so->so_snd) == 0) &&
9943             (acked > acked_amount) &&
9944             (tp->t_state >= TCPS_FIN_WAIT_1) &&
9945             (tp->t_flags & TF_SENTFIN)) {
9946                 /*
9947                  * We must be sure our fin
9948                  * was sent and acked (we can be
9949                  * in FIN_WAIT_1 without having
9950                  * sent the fin).
9951                  */
9952                 ourfinisacked = 1;
9953         }
9954         tp->snd_una = th->th_ack;
9955         if (acked_amount && sbavail(&so->so_snd))
9956                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
9957         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
9958         /* NB: sowwakeup_locked() does an implicit unlock. */
9959         sowwakeup_locked(so);
9960         m_freem(mfree);
9961         if (SEQ_GT(tp->snd_una, tp->snd_recover))
9962                 tp->snd_recover = tp->snd_una;
9963
9964         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
9965                 tp->snd_nxt = tp->snd_una;
9966         }
9967         if (under_pacing &&
9968             (rack->use_fixed_rate == 0) &&
9969             (rack->in_probe_rtt == 0) &&
9970             rack->rc_gp_dyn_mul &&
9971             rack->rc_always_pace) {
9972                 /* Check if we are dragging bottom */
9973                 rack_check_bottom_drag(tp, rack, so, acked);
9974         }
9975         if (tp->snd_una == tp->snd_max) {
9976                 /* Nothing left outstanding */
9977                 tp->t_flags &= ~TF_PREVVALID;
9978                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
9979                 rack->r_ctl.retran_during_recovery = 0;
9980                 rack->r_ctl.dsack_byte_cnt = 0;
9981                 if (rack->r_ctl.rc_went_idle_time == 0)
9982                         rack->r_ctl.rc_went_idle_time = 1;
9983                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
9984                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
9985                         tp->t_acktime = 0;
9986                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9987                 /* Set need output so persist might get set */
9988                 rack->r_wanted_output = 1;
9989                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
9990                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
9991                     (sbavail(&so->so_snd) == 0) &&
9992                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
9993                         /*
9994                          * The socket was gone and the
9995                          * peer sent data (now or in the past), time to
9996                          * reset him.
9997                          */
9998                         *ret_val = 1;
9999                         /* tcp_close will kill the inp pre-log the Reset */
10000                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10001                         tp = tcp_close(tp);
10002                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10003                         return (1);
10004                 }
10005         }
10006         if (ofia)
10007                 *ofia = ourfinisacked;
10008         return (0);
10009 }
10010
10011 static void
10012 rack_collapsed_window(struct tcp_rack *rack)
10013 {
10014         /*
10015          * Now we must walk the
10016          * send map and divide the
10017          * ones left stranded. These
10018          * guys can't cause us to abort
10019          * the connection and are really
10020          * "unsent". However if a buggy
10021          * client actually did keep some
10022          * of the data i.e. collapsed the win
10023          * and refused to ack and then opened
10024          * the win and acked that data. We would
10025          * get into an ack war, the simplier
10026          * method then of just pretending we
10027          * did not send those segments something
10028          * won't work.
10029          */
10030         struct rack_sendmap *rsm, *nrsm, fe, *insret;
10031         tcp_seq max_seq;
10032
10033         max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10034         memset(&fe, 0, sizeof(fe));
10035         fe.r_start = max_seq;
10036         /* Find the first seq past or at maxseq */
10037         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10038         if (rsm == NULL) {
10039                 /* Nothing to do strange */
10040                 rack->rc_has_collapsed = 0;
10041                 return;
10042         }
10043         /*
10044          * Now do we need to split at
10045          * the collapse point?
10046          */
10047         if (SEQ_GT(max_seq, rsm->r_start)) {
10048                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10049                 if (nrsm == NULL) {
10050                         /* We can't get a rsm, mark all? */
10051                         nrsm = rsm;
10052                         goto no_split;
10053                 }
10054                 /* Clone it */
10055                 rack_clone_rsm(rack, nrsm, rsm, max_seq);
10056                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10057 #ifdef INVARIANTS
10058                 if (insret != NULL) {
10059                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10060                               nrsm, insret, rack, rsm);
10061                 }
10062 #endif
10063                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
10064                 if (rsm->r_in_tmap) {
10065                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10066                         nrsm->r_in_tmap = 1;
10067                 }
10068                 /*
10069                  * Set in the new RSM as the
10070                  * collapsed starting point
10071                  */
10072                 rsm = nrsm;
10073         }
10074 no_split:
10075         counter_u64_add(rack_collapsed_win, 1);
10076         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10077                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
10078         }
10079         rack->rc_has_collapsed = 1;
10080 }
10081
10082 static void
10083 rack_un_collapse_window(struct tcp_rack *rack)
10084 {
10085         struct rack_sendmap *rsm;
10086
10087         RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
10088                 if (rsm->r_flags & RACK_RWND_COLLAPSED)
10089                         rsm->r_flags &= ~RACK_RWND_COLLAPSED;
10090                 else
10091                         break;
10092         }
10093         rack->rc_has_collapsed = 0;
10094 }
10095
10096 static void
10097 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10098                         int32_t tlen, int32_t tfo_syn)
10099 {
10100         if (DELAY_ACK(tp, tlen) || tfo_syn) {
10101                 if (rack->rc_dack_mode &&
10102                     (tlen > 500) &&
10103                     (rack->rc_dack_toggle == 1)) {
10104                         goto no_delayed_ack;
10105                 }
10106                 rack_timer_cancel(tp, rack,
10107                                   rack->r_ctl.rc_rcvtime, __LINE__);
10108                 tp->t_flags |= TF_DELACK;
10109         } else {
10110 no_delayed_ack:
10111                 rack->r_wanted_output = 1;
10112                 tp->t_flags |= TF_ACKNOW;
10113                 if (rack->rc_dack_mode) {
10114                         if (tp->t_flags & TF_DELACK)
10115                                 rack->rc_dack_toggle = 1;
10116                         else
10117                                 rack->rc_dack_toggle = 0;
10118                 }
10119         }
10120 }
10121
10122 static void
10123 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10124 {
10125         /*
10126          * If fast output is in progress, lets validate that
10127          * the new window did not shrink on us and make it
10128          * so fast output should end.
10129          */
10130         if (rack->r_fast_output) {
10131                 uint32_t out;
10132
10133                 /*
10134                  * Calculate what we will send if left as is
10135                  * and compare that to our send window.
10136                  */
10137                 out = ctf_outstanding(tp);
10138                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10139                         /* ok we have an issue */
10140                         if (out >= tp->snd_wnd) {
10141                                 /* Turn off fast output the window is met or collapsed */
10142                                 rack->r_fast_output = 0;
10143                         } else {
10144                                 /* we have some room left */
10145                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10146                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10147                                         /* If not at least 1 full segment never mind */
10148                                         rack->r_fast_output = 0;
10149                                 }
10150                         }
10151                 }
10152         }
10153 }
10154
10155
10156 /*
10157  * Return value of 1, the TCB is unlocked and most
10158  * likely gone, return value of 0, the TCP is still
10159  * locked.
10160  */
10161 static int
10162 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10163     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10164     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10165 {
10166         /*
10167          * Update window information. Don't look at window if no ACK: TAC's
10168          * send garbage on first SYN.
10169          */
10170         int32_t nsegs;
10171         int32_t tfo_syn;
10172         struct tcp_rack *rack;
10173
10174         rack = (struct tcp_rack *)tp->t_fb_ptr;
10175         INP_WLOCK_ASSERT(tp->t_inpcb);
10176         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10177         if ((thflags & TH_ACK) &&
10178             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10179             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10180             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10181                 /* keep track of pure window updates */
10182                 if (tlen == 0 &&
10183                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10184                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10185                 tp->snd_wnd = tiwin;
10186                 rack_validate_fo_sendwin_up(tp, rack);
10187                 tp->snd_wl1 = th->th_seq;
10188                 tp->snd_wl2 = th->th_ack;
10189                 if (tp->snd_wnd > tp->max_sndwnd)
10190                         tp->max_sndwnd = tp->snd_wnd;
10191                 rack->r_wanted_output = 1;
10192         } else if (thflags & TH_ACK) {
10193                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10194                         tp->snd_wnd = tiwin;
10195                         rack_validate_fo_sendwin_up(tp, rack);
10196                         tp->snd_wl1 = th->th_seq;
10197                         tp->snd_wl2 = th->th_ack;
10198                 }
10199         }
10200         if (tp->snd_wnd < ctf_outstanding(tp))
10201                 /* The peer collapsed the window */
10202                 rack_collapsed_window(rack);
10203         else if (rack->rc_has_collapsed)
10204                 rack_un_collapse_window(rack);
10205         /* Was persist timer active and now we have window space? */
10206         if ((rack->rc_in_persist != 0) &&
10207             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10208                                 rack->r_ctl.rc_pace_min_segs))) {
10209                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10210                 tp->snd_nxt = tp->snd_max;
10211                 /* Make sure we output to start the timer */
10212                 rack->r_wanted_output = 1;
10213         }
10214         /* Do we enter persists? */
10215         if ((rack->rc_in_persist == 0) &&
10216             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10217             TCPS_HAVEESTABLISHED(tp->t_state) &&
10218             (tp->snd_max == tp->snd_una) &&
10219             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10220             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10221                 /*
10222                  * Here the rwnd is less than
10223                  * the pacing size, we are established,
10224                  * nothing is outstanding, and there is
10225                  * data to send. Enter persists.
10226                  */
10227                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10228         }
10229         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10230                 m_freem(m);
10231                 return (0);
10232         }
10233         /*
10234          * don't process the URG bit, ignore them drag
10235          * along the up.
10236          */
10237         tp->rcv_up = tp->rcv_nxt;
10238         INP_WLOCK_ASSERT(tp->t_inpcb);
10239
10240         /*
10241          * Process the segment text, merging it into the TCP sequencing
10242          * queue, and arranging for acknowledgment of receipt if necessary.
10243          * This process logically involves adjusting tp->rcv_wnd as data is
10244          * presented to the user (this happens in tcp_usrreq.c, case
10245          * PRU_RCVD).  If a FIN has already been received on this connection
10246          * then we just ignore the text.
10247          */
10248         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10249                    IS_FASTOPEN(tp->t_flags));
10250         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10251             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10252                 tcp_seq save_start = th->th_seq;
10253                 tcp_seq save_rnxt  = tp->rcv_nxt;
10254                 int     save_tlen  = tlen;
10255
10256                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10257                 /*
10258                  * Insert segment which includes th into TCP reassembly
10259                  * queue with control block tp.  Set thflags to whether
10260                  * reassembly now includes a segment with FIN.  This handles
10261                  * the common case inline (segment is the next to be
10262                  * received on an established connection, and the queue is
10263                  * empty), avoiding linkage into and removal from the queue
10264                  * and repetition of various conversions. Set DELACK for
10265                  * segments received in order, but ack immediately when
10266                  * segments are out of order (so fast retransmit can work).
10267                  */
10268                 if (th->th_seq == tp->rcv_nxt &&
10269                     SEGQ_EMPTY(tp) &&
10270                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10271                     tfo_syn)) {
10272 #ifdef NETFLIX_SB_LIMITS
10273                         u_int mcnt, appended;
10274
10275                         if (so->so_rcv.sb_shlim) {
10276                                 mcnt = m_memcnt(m);
10277                                 appended = 0;
10278                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10279                                     CFO_NOSLEEP, NULL) == false) {
10280                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10281                                         m_freem(m);
10282                                         return (0);
10283                                 }
10284                         }
10285 #endif
10286                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10287                         tp->rcv_nxt += tlen;
10288                         if (tlen &&
10289                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10290                             (tp->t_fbyte_in == 0)) {
10291                                 tp->t_fbyte_in = ticks;
10292                                 if (tp->t_fbyte_in == 0)
10293                                         tp->t_fbyte_in = 1;
10294                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10295                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10296                         }
10297                         thflags = th->th_flags & TH_FIN;
10298                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10299                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10300                         SOCKBUF_LOCK(&so->so_rcv);
10301                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10302                                 m_freem(m);
10303                         } else
10304 #ifdef NETFLIX_SB_LIMITS
10305                                 appended =
10306 #endif
10307                                         sbappendstream_locked(&so->so_rcv, m, 0);
10308
10309                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10310                         /* NB: sorwakeup_locked() does an implicit unlock. */
10311                         sorwakeup_locked(so);
10312 #ifdef NETFLIX_SB_LIMITS
10313                         if (so->so_rcv.sb_shlim && appended != mcnt)
10314                                 counter_fo_release(so->so_rcv.sb_shlim,
10315                                     mcnt - appended);
10316 #endif
10317                 } else {
10318                         /*
10319                          * XXX: Due to the header drop above "th" is
10320                          * theoretically invalid by now.  Fortunately
10321                          * m_adj() doesn't actually frees any mbufs when
10322                          * trimming from the head.
10323                          */
10324                         tcp_seq temp = save_start;
10325
10326                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10327                         tp->t_flags |= TF_ACKNOW;
10328                         if (tp->t_flags & TF_WAKESOR) {
10329                                 tp->t_flags &= ~TF_WAKESOR;
10330                                 /* NB: sorwakeup_locked() does an implicit unlock. */
10331                                 sorwakeup_locked(so);
10332                         }
10333                 }
10334                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10335                     (save_tlen > 0) &&
10336                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10337                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10338                                 /*
10339                                  * DSACK actually handled in the fastpath
10340                                  * above.
10341                                  */
10342                                 RACK_OPTS_INC(tcp_sack_path_1);
10343                                 tcp_update_sack_list(tp, save_start,
10344                                     save_start + save_tlen);
10345                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10346                                 if ((tp->rcv_numsacks >= 1) &&
10347                                     (tp->sackblks[0].end == save_start)) {
10348                                         /*
10349                                          * Partial overlap, recorded at todrop
10350                                          * above.
10351                                          */
10352                                         RACK_OPTS_INC(tcp_sack_path_2a);
10353                                         tcp_update_sack_list(tp,
10354                                             tp->sackblks[0].start,
10355                                             tp->sackblks[0].end);
10356                                 } else {
10357                                         RACK_OPTS_INC(tcp_sack_path_2b);
10358                                         tcp_update_dsack_list(tp, save_start,
10359                                             save_start + save_tlen);
10360                                 }
10361                         } else if (tlen >= save_tlen) {
10362                                 /* Update of sackblks. */
10363                                 RACK_OPTS_INC(tcp_sack_path_3);
10364                                 tcp_update_dsack_list(tp, save_start,
10365                                     save_start + save_tlen);
10366                         } else if (tlen > 0) {
10367                                 RACK_OPTS_INC(tcp_sack_path_4);
10368                                 tcp_update_dsack_list(tp, save_start,
10369                                     save_start + tlen);
10370                         }
10371                 }
10372         } else {
10373                 m_freem(m);
10374                 thflags &= ~TH_FIN;
10375         }
10376
10377         /*
10378          * If FIN is received ACK the FIN and let the user know that the
10379          * connection is closing.
10380          */
10381         if (thflags & TH_FIN) {
10382                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10383                         /* The socket upcall is handled by socantrcvmore. */
10384                         socantrcvmore(so);
10385                         /*
10386                          * If connection is half-synchronized (ie NEEDSYN
10387                          * flag on) then delay ACK, so it may be piggybacked
10388                          * when SYN is sent. Otherwise, since we received a
10389                          * FIN then no more input can be expected, send ACK
10390                          * now.
10391                          */
10392                         if (tp->t_flags & TF_NEEDSYN) {
10393                                 rack_timer_cancel(tp, rack,
10394                                     rack->r_ctl.rc_rcvtime, __LINE__);
10395                                 tp->t_flags |= TF_DELACK;
10396                         } else {
10397                                 tp->t_flags |= TF_ACKNOW;
10398                         }
10399                         tp->rcv_nxt++;
10400                 }
10401                 switch (tp->t_state) {
10402                         /*
10403                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10404                          * CLOSE_WAIT state.
10405                          */
10406                 case TCPS_SYN_RECEIVED:
10407                         tp->t_starttime = ticks;
10408                         /* FALLTHROUGH */
10409                 case TCPS_ESTABLISHED:
10410                         rack_timer_cancel(tp, rack,
10411                             rack->r_ctl.rc_rcvtime, __LINE__);
10412                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10413                         break;
10414
10415                         /*
10416                          * If still in FIN_WAIT_1 STATE FIN has not been
10417                          * acked so enter the CLOSING state.
10418                          */
10419                 case TCPS_FIN_WAIT_1:
10420                         rack_timer_cancel(tp, rack,
10421                             rack->r_ctl.rc_rcvtime, __LINE__);
10422                         tcp_state_change(tp, TCPS_CLOSING);
10423                         break;
10424
10425                         /*
10426                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10427                          * starting the time-wait timer, turning off the
10428                          * other standard timers.
10429                          */
10430                 case TCPS_FIN_WAIT_2:
10431                         rack_timer_cancel(tp, rack,
10432                             rack->r_ctl.rc_rcvtime, __LINE__);
10433                         tcp_twstart(tp);
10434                         return (1);
10435                 }
10436         }
10437         /*
10438          * Return any desired output.
10439          */
10440         if ((tp->t_flags & TF_ACKNOW) ||
10441             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10442                 rack->r_wanted_output = 1;
10443         }
10444         INP_WLOCK_ASSERT(tp->t_inpcb);
10445         return (0);
10446 }
10447
10448 /*
10449  * Here nothing is really faster, its just that we
10450  * have broken out the fast-data path also just like
10451  * the fast-ack.
10452  */
10453 static int
10454 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10455     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10456     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10457 {
10458         int32_t nsegs;
10459         int32_t newsize = 0;    /* automatic sockbuf scaling */
10460         struct tcp_rack *rack;
10461 #ifdef NETFLIX_SB_LIMITS
10462         u_int mcnt, appended;
10463 #endif
10464 #ifdef TCPDEBUG
10465         /*
10466          * The size of tcp_saveipgen must be the size of the max ip header,
10467          * now IPv6.
10468          */
10469         u_char tcp_saveipgen[IP6_HDR_LEN];
10470         struct tcphdr tcp_savetcp;
10471         short ostate = 0;
10472
10473 #endif
10474         /*
10475          * If last ACK falls within this segment's sequence numbers, record
10476          * the timestamp. NOTE that the test is modified according to the
10477          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10478          */
10479         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10480                 return (0);
10481         }
10482         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10483                 return (0);
10484         }
10485         if (tiwin && tiwin != tp->snd_wnd) {
10486                 return (0);
10487         }
10488         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10489                 return (0);
10490         }
10491         if (__predict_false((to->to_flags & TOF_TS) &&
10492             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10493                 return (0);
10494         }
10495         if (__predict_false((th->th_ack != tp->snd_una))) {
10496                 return (0);
10497         }
10498         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10499                 return (0);
10500         }
10501         if ((to->to_flags & TOF_TS) != 0 &&
10502             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10503                 tp->ts_recent_age = tcp_ts_getticks();
10504                 tp->ts_recent = to->to_tsval;
10505         }
10506         rack = (struct tcp_rack *)tp->t_fb_ptr;
10507         /*
10508          * This is a pure, in-sequence data packet with nothing on the
10509          * reassembly queue and we have enough buffer space to take it.
10510          */
10511         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10512
10513 #ifdef NETFLIX_SB_LIMITS
10514         if (so->so_rcv.sb_shlim) {
10515                 mcnt = m_memcnt(m);
10516                 appended = 0;
10517                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10518                     CFO_NOSLEEP, NULL) == false) {
10519                         counter_u64_add(tcp_sb_shlim_fails, 1);
10520                         m_freem(m);
10521                         return (1);
10522                 }
10523         }
10524 #endif
10525         /* Clean receiver SACK report if present */
10526         if (tp->rcv_numsacks)
10527                 tcp_clean_sackreport(tp);
10528         KMOD_TCPSTAT_INC(tcps_preddat);
10529         tp->rcv_nxt += tlen;
10530         if (tlen &&
10531             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10532             (tp->t_fbyte_in == 0)) {
10533                 tp->t_fbyte_in = ticks;
10534                 if (tp->t_fbyte_in == 0)
10535                         tp->t_fbyte_in = 1;
10536                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10537                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10538         }
10539         /*
10540          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10541          */
10542         tp->snd_wl1 = th->th_seq;
10543         /*
10544          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10545          */
10546         tp->rcv_up = tp->rcv_nxt;
10547         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10548         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10549 #ifdef TCPDEBUG
10550         if (so->so_options & SO_DEBUG)
10551                 tcp_trace(TA_INPUT, ostate, tp,
10552                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
10553 #endif
10554         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10555
10556         /* Add data to socket buffer. */
10557         SOCKBUF_LOCK(&so->so_rcv);
10558         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10559                 m_freem(m);
10560         } else {
10561                 /*
10562                  * Set new socket buffer size. Give up when limit is
10563                  * reached.
10564                  */
10565                 if (newsize)
10566                         if (!sbreserve_locked(&so->so_rcv,
10567                             newsize, so, NULL))
10568                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10569                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10570 #ifdef NETFLIX_SB_LIMITS
10571                 appended =
10572 #endif
10573                         sbappendstream_locked(&so->so_rcv, m, 0);
10574                 ctf_calc_rwin(so, tp);
10575         }
10576         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10577         /* NB: sorwakeup_locked() does an implicit unlock. */
10578         sorwakeup_locked(so);
10579 #ifdef NETFLIX_SB_LIMITS
10580         if (so->so_rcv.sb_shlim && mcnt != appended)
10581                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10582 #endif
10583         rack_handle_delayed_ack(tp, rack, tlen, 0);
10584         if (tp->snd_una == tp->snd_max)
10585                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10586         return (1);
10587 }
10588
10589 /*
10590  * This subfunction is used to try to highly optimize the
10591  * fast path. We again allow window updates that are
10592  * in sequence to remain in the fast-path. We also add
10593  * in the __predict's to attempt to help the compiler.
10594  * Note that if we return a 0, then we can *not* process
10595  * it and the caller should push the packet into the
10596  * slow-path.
10597  */
10598 static int
10599 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10600     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10601     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10602 {
10603         int32_t acked;
10604         int32_t nsegs;
10605 #ifdef TCPDEBUG
10606         /*
10607          * The size of tcp_saveipgen must be the size of the max ip header,
10608          * now IPv6.
10609          */
10610         u_char tcp_saveipgen[IP6_HDR_LEN];
10611         struct tcphdr tcp_savetcp;
10612         short ostate = 0;
10613 #endif
10614         int32_t under_pacing = 0;
10615         struct tcp_rack *rack;
10616
10617         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10618                 /* Old ack, behind (or duplicate to) the last one rcv'd */
10619                 return (0);
10620         }
10621         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10622                 /* Above what we have sent? */
10623                 return (0);
10624         }
10625         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10626                 /* We are retransmitting */
10627                 return (0);
10628         }
10629         if (__predict_false(tiwin == 0)) {
10630                 /* zero window */
10631                 return (0);
10632         }
10633         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10634                 /* We need a SYN or a FIN, unlikely.. */
10635                 return (0);
10636         }
10637         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10638                 /* Timestamp is behind .. old ack with seq wrap? */
10639                 return (0);
10640         }
10641         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10642                 /* Still recovering */
10643                 return (0);
10644         }
10645         rack = (struct tcp_rack *)tp->t_fb_ptr;
10646         if (rack->r_ctl.rc_sacked) {
10647                 /* We have sack holes on our scoreboard */
10648                 return (0);
10649         }
10650         /* Ok if we reach here, we can process a fast-ack */
10651         if (rack->gp_ready &&
10652             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10653                 under_pacing = 1;
10654         }
10655         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10656         rack_log_ack(tp, to, th, 0, 0);
10657         /* Did the window get updated? */
10658         if (tiwin != tp->snd_wnd) {
10659                 tp->snd_wnd = tiwin;
10660                 rack_validate_fo_sendwin_up(tp, rack);
10661                 tp->snd_wl1 = th->th_seq;
10662                 if (tp->snd_wnd > tp->max_sndwnd)
10663                         tp->max_sndwnd = tp->snd_wnd;
10664         }
10665         /* Do we exit persists? */
10666         if ((rack->rc_in_persist != 0) &&
10667             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10668                                rack->r_ctl.rc_pace_min_segs))) {
10669                 rack_exit_persist(tp, rack, cts);
10670         }
10671         /* Do we enter persists? */
10672         if ((rack->rc_in_persist == 0) &&
10673             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10674             TCPS_HAVEESTABLISHED(tp->t_state) &&
10675             (tp->snd_max == tp->snd_una) &&
10676             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10677             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10678                 /*
10679                  * Here the rwnd is less than
10680                  * the pacing size, we are established,
10681                  * nothing is outstanding, and there is
10682                  * data to send. Enter persists.
10683                  */
10684                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10685         }
10686         /*
10687          * If last ACK falls within this segment's sequence numbers, record
10688          * the timestamp. NOTE that the test is modified according to the
10689          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10690          */
10691         if ((to->to_flags & TOF_TS) != 0 &&
10692             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10693                 tp->ts_recent_age = tcp_ts_getticks();
10694                 tp->ts_recent = to->to_tsval;
10695         }
10696         /*
10697          * This is a pure ack for outstanding data.
10698          */
10699         KMOD_TCPSTAT_INC(tcps_predack);
10700
10701         /*
10702          * "bad retransmit" recovery.
10703          */
10704         if ((tp->t_flags & TF_PREVVALID) &&
10705             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10706                 tp->t_flags &= ~TF_PREVVALID;
10707                 if (tp->t_rxtshift == 1 &&
10708                     (int)(ticks - tp->t_badrxtwin) < 0)
10709                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10710         }
10711         /*
10712          * Recalculate the transmit timer / rtt.
10713          *
10714          * Some boxes send broken timestamp replies during the SYN+ACK
10715          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10716          * and blow up the retransmit timer.
10717          */
10718         acked = BYTES_THIS_ACK(tp, th);
10719
10720 #ifdef TCP_HHOOK
10721         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10722         hhook_run_tcp_est_in(tp, th, to);
10723 #endif
10724         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10725         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10726         if (acked) {
10727                 struct mbuf *mfree;
10728
10729                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
10730                 SOCKBUF_LOCK(&so->so_snd);
10731                 mfree = sbcut_locked(&so->so_snd, acked);
10732                 tp->snd_una = th->th_ack;
10733                 /* Note we want to hold the sb lock through the sendmap adjust */
10734                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10735                 /* Wake up the socket if we have room to write more */
10736                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10737                 sowwakeup_locked(so);
10738                 m_freem(mfree);
10739                 tp->t_rxtshift = 0;
10740                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10741                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10742                 rack->rc_tlp_in_progress = 0;
10743                 rack->r_ctl.rc_tlp_cnt_out = 0;
10744                 /*
10745                  * If it is the RXT timer we want to
10746                  * stop it, so we can restart a TLP.
10747                  */
10748                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10749                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10750 #ifdef NETFLIX_HTTP_LOGGING
10751                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10752 #endif
10753         }
10754         /*
10755          * Let the congestion control algorithm update congestion control
10756          * related information. This typically means increasing the
10757          * congestion window.
10758          */
10759         if (tp->snd_wnd < ctf_outstanding(tp)) {
10760                 /* The peer collapsed the window */
10761                 rack_collapsed_window(rack);
10762         } else if (rack->rc_has_collapsed)
10763                 rack_un_collapse_window(rack);
10764
10765         /*
10766          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
10767          */
10768         tp->snd_wl2 = th->th_ack;
10769         tp->t_dupacks = 0;
10770         m_freem(m);
10771         /* ND6_HINT(tp);         *//* Some progress has been made. */
10772
10773         /*
10774          * If all outstanding data are acked, stop retransmit timer,
10775          * otherwise restart timer using current (possibly backed-off)
10776          * value. If process is waiting for space, wakeup/selwakeup/signal.
10777          * If data are ready to send, let tcp_output decide between more
10778          * output or persist.
10779          */
10780 #ifdef TCPDEBUG
10781         if (so->so_options & SO_DEBUG)
10782                 tcp_trace(TA_INPUT, ostate, tp,
10783                     (void *)tcp_saveipgen,
10784                     &tcp_savetcp, 0);
10785 #endif
10786         if (under_pacing &&
10787             (rack->use_fixed_rate == 0) &&
10788             (rack->in_probe_rtt == 0) &&
10789             rack->rc_gp_dyn_mul &&
10790             rack->rc_always_pace) {
10791                 /* Check if we are dragging bottom */
10792                 rack_check_bottom_drag(tp, rack, so, acked);
10793         }
10794         if (tp->snd_una == tp->snd_max) {
10795                 tp->t_flags &= ~TF_PREVVALID;
10796                 rack->r_ctl.retran_during_recovery = 0;
10797                 rack->r_ctl.dsack_byte_cnt = 0;
10798                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10799                 if (rack->r_ctl.rc_went_idle_time == 0)
10800                         rack->r_ctl.rc_went_idle_time = 1;
10801                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10802                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10803                         tp->t_acktime = 0;
10804                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10805         }
10806         if (acked && rack->r_fast_output)
10807                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
10808         if (sbavail(&so->so_snd)) {
10809                 rack->r_wanted_output = 1;
10810         }
10811         return (1);
10812 }
10813
10814 /*
10815  * Return value of 1, the TCB is unlocked and most
10816  * likely gone, return value of 0, the TCP is still
10817  * locked.
10818  */
10819 static int
10820 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
10821     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10822     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10823 {
10824         int32_t ret_val = 0;
10825         int32_t todrop;
10826         int32_t ourfinisacked = 0;
10827         struct tcp_rack *rack;
10828
10829         ctf_calc_rwin(so, tp);
10830         /*
10831          * If the state is SYN_SENT: if seg contains an ACK, but not for our
10832          * SYN, drop the input. if seg contains a RST, then drop the
10833          * connection. if seg does not contain SYN, then drop it. Otherwise
10834          * this is an acceptable SYN segment initialize tp->rcv_nxt and
10835          * tp->irs if seg contains ack then advance tp->snd_una if seg
10836          * contains an ECE and ECN support is enabled, the stream is ECN
10837          * capable. if SYN has been acked change to ESTABLISHED else
10838          * SYN_RCVD state arrange for segment to be acked (eventually)
10839          * continue processing rest of data/controls.
10840          */
10841         if ((thflags & TH_ACK) &&
10842             (SEQ_LEQ(th->th_ack, tp->iss) ||
10843             SEQ_GT(th->th_ack, tp->snd_max))) {
10844                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10845                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10846                 return (1);
10847         }
10848         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
10849                 TCP_PROBE5(connect__refused, NULL, tp,
10850                     mtod(m, const char *), tp, th);
10851                 tp = tcp_drop(tp, ECONNREFUSED);
10852                 ctf_do_drop(m, tp);
10853                 return (1);
10854         }
10855         if (thflags & TH_RST) {
10856                 ctf_do_drop(m, tp);
10857                 return (1);
10858         }
10859         if (!(thflags & TH_SYN)) {
10860                 ctf_do_drop(m, tp);
10861                 return (1);
10862         }
10863         tp->irs = th->th_seq;
10864         tcp_rcvseqinit(tp);
10865         rack = (struct tcp_rack *)tp->t_fb_ptr;
10866         if (thflags & TH_ACK) {
10867                 int tfo_partial = 0;
10868
10869                 KMOD_TCPSTAT_INC(tcps_connects);
10870                 soisconnected(so);
10871 #ifdef MAC
10872                 mac_socketpeer_set_from_mbuf(m, so);
10873 #endif
10874                 /* Do window scaling on this connection? */
10875                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10876                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10877                         tp->rcv_scale = tp->request_r_scale;
10878                 }
10879                 tp->rcv_adv += min(tp->rcv_wnd,
10880                     TCP_MAXWIN << tp->rcv_scale);
10881                 /*
10882                  * If not all the data that was sent in the TFO SYN
10883                  * has been acked, resend the remainder right away.
10884                  */
10885                 if (IS_FASTOPEN(tp->t_flags) &&
10886                     (tp->snd_una != tp->snd_max)) {
10887                         tp->snd_nxt = th->th_ack;
10888                         tfo_partial = 1;
10889                 }
10890                 /*
10891                  * If there's data, delay ACK; if there's also a FIN ACKNOW
10892                  * will be turned on later.
10893                  */
10894                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
10895                         rack_timer_cancel(tp, rack,
10896                                           rack->r_ctl.rc_rcvtime, __LINE__);
10897                         tp->t_flags |= TF_DELACK;
10898                 } else {
10899                         rack->r_wanted_output = 1;
10900                         tp->t_flags |= TF_ACKNOW;
10901                         rack->rc_dack_toggle = 0;
10902                 }
10903                 if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
10904                     (V_tcp_do_ecn == 1)) {
10905                         tp->t_flags2 |= TF2_ECN_PERMIT;
10906                         KMOD_TCPSTAT_INC(tcps_ecn_shs);
10907                 }
10908                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
10909                         /*
10910                          * We advance snd_una for the
10911                          * fast open case. If th_ack is
10912                          * acknowledging data beyond
10913                          * snd_una we can't just call
10914                          * ack-processing since the
10915                          * data stream in our send-map
10916                          * will start at snd_una + 1 (one
10917                          * beyond the SYN). If its just
10918                          * equal we don't need to do that
10919                          * and there is no send_map.
10920                          */
10921                         tp->snd_una++;
10922                 }
10923                 /*
10924                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
10925                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
10926                  */
10927                 tp->t_starttime = ticks;
10928                 if (tp->t_flags & TF_NEEDFIN) {
10929                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
10930                         tp->t_flags &= ~TF_NEEDFIN;
10931                         thflags &= ~TH_SYN;
10932                 } else {
10933                         tcp_state_change(tp, TCPS_ESTABLISHED);
10934                         TCP_PROBE5(connect__established, NULL, tp,
10935                             mtod(m, const char *), tp, th);
10936                         rack_cc_conn_init(tp);
10937                 }
10938         } else {
10939                 /*
10940                  * Received initial SYN in SYN-SENT[*] state => simultaneous
10941                  * open.  If segment contains CC option and there is a
10942                  * cached CC, apply TAO test. If it succeeds, connection is *
10943                  * half-synchronized. Otherwise, do 3-way handshake:
10944                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
10945                  * there was no CC option, clear cached CC value.
10946                  */
10947                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
10948                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
10949         }
10950         INP_WLOCK_ASSERT(tp->t_inpcb);
10951         /*
10952          * Advance th->th_seq to correspond to first data byte. If data,
10953          * trim to stay within window, dropping FIN if necessary.
10954          */
10955         th->th_seq++;
10956         if (tlen > tp->rcv_wnd) {
10957                 todrop = tlen - tp->rcv_wnd;
10958                 m_adj(m, -todrop);
10959                 tlen = tp->rcv_wnd;
10960                 thflags &= ~TH_FIN;
10961                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
10962                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
10963         }
10964         tp->snd_wl1 = th->th_seq - 1;
10965         tp->rcv_up = th->th_seq;
10966         /*
10967          * Client side of transaction: already sent SYN and data. If the
10968          * remote host used T/TCP to validate the SYN, our data will be
10969          * ACK'd; if so, enter normal data segment processing in the middle
10970          * of step 5, ack processing. Otherwise, goto step 6.
10971          */
10972         if (thflags & TH_ACK) {
10973                 /* For syn-sent we need to possibly update the rtt */
10974                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
10975                         uint32_t t, mcts;
10976
10977                         mcts = tcp_ts_getticks();
10978                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
10979                         if (!tp->t_rttlow || tp->t_rttlow > t)
10980                                 tp->t_rttlow = t;
10981                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
10982                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
10983                         tcp_rack_xmit_timer_commit(rack, tp);
10984                 }
10985                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
10986                         return (ret_val);
10987                 /* We may have changed to FIN_WAIT_1 above */
10988                 if (tp->t_state == TCPS_FIN_WAIT_1) {
10989                         /*
10990                          * In FIN_WAIT_1 STATE in addition to the processing
10991                          * for the ESTABLISHED state if our FIN is now
10992                          * acknowledged then enter FIN_WAIT_2.
10993                          */
10994                         if (ourfinisacked) {
10995                                 /*
10996                                  * If we can't receive any more data, then
10997                                  * closing user can proceed. Starting the
10998                                  * timer is contrary to the specification,
10999                                  * but if we don't get a FIN we'll hang
11000                                  * forever.
11001                                  *
11002                                  * XXXjl: we should release the tp also, and
11003                                  * use a compressed state.
11004                                  */
11005                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11006                                         soisdisconnected(so);
11007                                         tcp_timer_activate(tp, TT_2MSL,
11008                                             (tcp_fast_finwait2_recycle ?
11009                                             tcp_finwait2_timeout :
11010                                             TP_MAXIDLE(tp)));
11011                                 }
11012                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11013                         }
11014                 }
11015         }
11016         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11017            tiwin, thflags, nxt_pkt));
11018 }
11019
11020 /*
11021  * Return value of 1, the TCB is unlocked and most
11022  * likely gone, return value of 0, the TCP is still
11023  * locked.
11024  */
11025 static int
11026 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11027     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11028     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11029 {
11030         struct tcp_rack *rack;
11031         int32_t ret_val = 0;
11032         int32_t ourfinisacked = 0;
11033
11034         ctf_calc_rwin(so, tp);
11035         if ((thflags & TH_ACK) &&
11036             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11037             SEQ_GT(th->th_ack, tp->snd_max))) {
11038                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11039                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11040                 return (1);
11041         }
11042         rack = (struct tcp_rack *)tp->t_fb_ptr;
11043         if (IS_FASTOPEN(tp->t_flags)) {
11044                 /*
11045                  * When a TFO connection is in SYN_RECEIVED, the
11046                  * only valid packets are the initial SYN, a
11047                  * retransmit/copy of the initial SYN (possibly with
11048                  * a subset of the original data), a valid ACK, a
11049                  * FIN, or a RST.
11050                  */
11051                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11052                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11053                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11054                         return (1);
11055                 } else if (thflags & TH_SYN) {
11056                         /* non-initial SYN is ignored */
11057                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11058                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11059                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11060                                 ctf_do_drop(m, NULL);
11061                                 return (0);
11062                         }
11063                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11064                         ctf_do_drop(m, NULL);
11065                         return (0);
11066                 }
11067         }
11068         if ((thflags & TH_RST) ||
11069             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11070                 return (ctf_process_rst(m, th, so, tp));
11071         /*
11072          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11073          * it's less than ts_recent, drop it.
11074          */
11075         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11076             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11077                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11078                         return (ret_val);
11079         }
11080         /*
11081          * In the SYN-RECEIVED state, validate that the packet belongs to
11082          * this connection before trimming the data to fit the receive
11083          * window.  Check the sequence number versus IRS since we know the
11084          * sequence numbers haven't wrapped.  This is a partial fix for the
11085          * "LAND" DoS attack.
11086          */
11087         if (SEQ_LT(th->th_seq, tp->irs)) {
11088                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11089                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11090                 return (1);
11091         }
11092         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11093                               &rack->r_ctl.challenge_ack_ts,
11094                               &rack->r_ctl.challenge_ack_cnt)) {
11095                 return (ret_val);
11096         }
11097         /*
11098          * If last ACK falls within this segment's sequence numbers, record
11099          * its timestamp. NOTE: 1) That the test incorporates suggestions
11100          * from the latest proposal of the tcplw@cray.com list (Braden
11101          * 1993/04/26). 2) That updating only on newer timestamps interferes
11102          * with our earlier PAWS tests, so this check should be solely
11103          * predicated on the sequence space of this segment. 3) That we
11104          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11105          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11106          * SEG.Len, This modified check allows us to overcome RFC1323's
11107          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11108          * p.869. In such cases, we can still calculate the RTT correctly
11109          * when RCV.NXT == Last.ACK.Sent.
11110          */
11111         if ((to->to_flags & TOF_TS) != 0 &&
11112             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11113             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11114             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11115                 tp->ts_recent_age = tcp_ts_getticks();
11116                 tp->ts_recent = to->to_tsval;
11117         }
11118         tp->snd_wnd = tiwin;
11119         rack_validate_fo_sendwin_up(tp, rack);
11120         /*
11121          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11122          * is on (half-synchronized state), then queue data for later
11123          * processing; else drop segment and return.
11124          */
11125         if ((thflags & TH_ACK) == 0) {
11126                 if (IS_FASTOPEN(tp->t_flags)) {
11127                         rack_cc_conn_init(tp);
11128                 }
11129                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11130                     tiwin, thflags, nxt_pkt));
11131         }
11132         KMOD_TCPSTAT_INC(tcps_connects);
11133         soisconnected(so);
11134         /* Do window scaling? */
11135         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11136             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11137                 tp->rcv_scale = tp->request_r_scale;
11138         }
11139         /*
11140          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11141          * FIN-WAIT-1
11142          */
11143         tp->t_starttime = ticks;
11144         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11145                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11146                 tp->t_tfo_pending = NULL;
11147         }
11148         if (tp->t_flags & TF_NEEDFIN) {
11149                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11150                 tp->t_flags &= ~TF_NEEDFIN;
11151         } else {
11152                 tcp_state_change(tp, TCPS_ESTABLISHED);
11153                 TCP_PROBE5(accept__established, NULL, tp,
11154                     mtod(m, const char *), tp, th);
11155                 /*
11156                  * TFO connections call cc_conn_init() during SYN
11157                  * processing.  Calling it again here for such connections
11158                  * is not harmless as it would undo the snd_cwnd reduction
11159                  * that occurs when a TFO SYN|ACK is retransmitted.
11160                  */
11161                 if (!IS_FASTOPEN(tp->t_flags))
11162                         rack_cc_conn_init(tp);
11163         }
11164         /*
11165          * Account for the ACK of our SYN prior to
11166          * regular ACK processing below, except for
11167          * simultaneous SYN, which is handled later.
11168          */
11169         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11170                 tp->snd_una++;
11171         /*
11172          * If segment contains data or ACK, will call tcp_reass() later; if
11173          * not, do so now to pass queued data to user.
11174          */
11175         if (tlen == 0 && (thflags & TH_FIN) == 0) {
11176                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11177                     (struct mbuf *)0);
11178                 if (tp->t_flags & TF_WAKESOR) {
11179                         tp->t_flags &= ~TF_WAKESOR;
11180                         /* NB: sorwakeup_locked() does an implicit unlock. */
11181                         sorwakeup_locked(so);
11182                 }
11183         }
11184         tp->snd_wl1 = th->th_seq - 1;
11185         /* For syn-recv we need to possibly update the rtt */
11186         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11187                 uint32_t t, mcts;
11188
11189                 mcts = tcp_ts_getticks();
11190                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11191                 if (!tp->t_rttlow || tp->t_rttlow > t)
11192                         tp->t_rttlow = t;
11193                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11194                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11195                 tcp_rack_xmit_timer_commit(rack, tp);
11196         }
11197         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11198                 return (ret_val);
11199         }
11200         if (tp->t_state == TCPS_FIN_WAIT_1) {
11201                 /* We could have went to FIN_WAIT_1 (or EST) above */
11202                 /*
11203                  * In FIN_WAIT_1 STATE in addition to the processing for the
11204                  * ESTABLISHED state if our FIN is now acknowledged then
11205                  * enter FIN_WAIT_2.
11206                  */
11207                 if (ourfinisacked) {
11208                         /*
11209                          * If we can't receive any more data, then closing
11210                          * user can proceed. Starting the timer is contrary
11211                          * to the specification, but if we don't get a FIN
11212                          * we'll hang forever.
11213                          *
11214                          * XXXjl: we should release the tp also, and use a
11215                          * compressed state.
11216                          */
11217                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11218                                 soisdisconnected(so);
11219                                 tcp_timer_activate(tp, TT_2MSL,
11220                                     (tcp_fast_finwait2_recycle ?
11221                                     tcp_finwait2_timeout :
11222                                     TP_MAXIDLE(tp)));
11223                         }
11224                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11225                 }
11226         }
11227         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11228             tiwin, thflags, nxt_pkt));
11229 }
11230
11231 /*
11232  * Return value of 1, the TCB is unlocked and most
11233  * likely gone, return value of 0, the TCP is still
11234  * locked.
11235  */
11236 static int
11237 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11238     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11239     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11240 {
11241         int32_t ret_val = 0;
11242         struct tcp_rack *rack;
11243
11244         /*
11245          * Header prediction: check for the two common cases of a
11246          * uni-directional data xfer.  If the packet has no control flags,
11247          * is in-sequence, the window didn't change and we're not
11248          * retransmitting, it's a candidate.  If the length is zero and the
11249          * ack moved forward, we're the sender side of the xfer.  Just free
11250          * the data acked & wake any higher level process that was blocked
11251          * waiting for space.  If the length is non-zero and the ack didn't
11252          * move, we're the receiver side.  If we're getting packets in-order
11253          * (the reassembly queue is empty), add the data toc The socket
11254          * buffer and note that we need a delayed ack. Make sure that the
11255          * hidden state-flags are also off. Since we check for
11256          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11257          */
11258         rack = (struct tcp_rack *)tp->t_fb_ptr;
11259         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11260             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11261             __predict_true(SEGQ_EMPTY(tp)) &&
11262             __predict_true(th->th_seq == tp->rcv_nxt)) {
11263                 if (tlen == 0) {
11264                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11265                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11266                                 return (0);
11267                         }
11268                 } else {
11269                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11270                             tiwin, nxt_pkt, iptos)) {
11271                                 return (0);
11272                         }
11273                 }
11274         }
11275         ctf_calc_rwin(so, tp);
11276
11277         if ((thflags & TH_RST) ||
11278             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11279                 return (ctf_process_rst(m, th, so, tp));
11280
11281         /*
11282          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11283          * synchronized state.
11284          */
11285         if (thflags & TH_SYN) {
11286                 ctf_challenge_ack(m, th, tp, &ret_val);
11287                 return (ret_val);
11288         }
11289         /*
11290          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11291          * it's less than ts_recent, drop it.
11292          */
11293         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11294             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11295                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11296                         return (ret_val);
11297         }
11298         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11299                               &rack->r_ctl.challenge_ack_ts,
11300                               &rack->r_ctl.challenge_ack_cnt)) {
11301                 return (ret_val);
11302         }
11303         /*
11304          * If last ACK falls within this segment's sequence numbers, record
11305          * its timestamp. NOTE: 1) That the test incorporates suggestions
11306          * from the latest proposal of the tcplw@cray.com list (Braden
11307          * 1993/04/26). 2) That updating only on newer timestamps interferes
11308          * with our earlier PAWS tests, so this check should be solely
11309          * predicated on the sequence space of this segment. 3) That we
11310          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11311          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11312          * SEG.Len, This modified check allows us to overcome RFC1323's
11313          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11314          * p.869. In such cases, we can still calculate the RTT correctly
11315          * when RCV.NXT == Last.ACK.Sent.
11316          */
11317         if ((to->to_flags & TOF_TS) != 0 &&
11318             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11319             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11320             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11321                 tp->ts_recent_age = tcp_ts_getticks();
11322                 tp->ts_recent = to->to_tsval;
11323         }
11324         /*
11325          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11326          * is on (half-synchronized state), then queue data for later
11327          * processing; else drop segment and return.
11328          */
11329         if ((thflags & TH_ACK) == 0) {
11330                 if (tp->t_flags & TF_NEEDSYN) {
11331                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11332                             tiwin, thflags, nxt_pkt));
11333
11334                 } else if (tp->t_flags & TF_ACKNOW) {
11335                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11336                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11337                         return (ret_val);
11338                 } else {
11339                         ctf_do_drop(m, NULL);
11340                         return (0);
11341                 }
11342         }
11343         /*
11344          * Ack processing.
11345          */
11346         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11347                 return (ret_val);
11348         }
11349         if (sbavail(&so->so_snd)) {
11350                 if (ctf_progress_timeout_check(tp, true)) {
11351                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11352                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11353                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11354                         return (1);
11355                 }
11356         }
11357         /* State changes only happen in rack_process_data() */
11358         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11359             tiwin, thflags, nxt_pkt));
11360 }
11361
11362 /*
11363  * Return value of 1, the TCB is unlocked and most
11364  * likely gone, return value of 0, the TCP is still
11365  * locked.
11366  */
11367 static int
11368 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11369     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11370     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11371 {
11372         int32_t ret_val = 0;
11373         struct tcp_rack *rack;
11374
11375         rack = (struct tcp_rack *)tp->t_fb_ptr;
11376         ctf_calc_rwin(so, tp);
11377         if ((thflags & TH_RST) ||
11378             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11379                 return (ctf_process_rst(m, th, so, tp));
11380         /*
11381          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11382          * synchronized state.
11383          */
11384         if (thflags & TH_SYN) {
11385                 ctf_challenge_ack(m, th, tp, &ret_val);
11386                 return (ret_val);
11387         }
11388         /*
11389          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11390          * it's less than ts_recent, drop it.
11391          */
11392         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11393             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11394                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11395                         return (ret_val);
11396         }
11397         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11398                               &rack->r_ctl.challenge_ack_ts,
11399                               &rack->r_ctl.challenge_ack_cnt)) {
11400                 return (ret_val);
11401         }
11402         /*
11403          * If last ACK falls within this segment's sequence numbers, record
11404          * its timestamp. NOTE: 1) That the test incorporates suggestions
11405          * from the latest proposal of the tcplw@cray.com list (Braden
11406          * 1993/04/26). 2) That updating only on newer timestamps interferes
11407          * with our earlier PAWS tests, so this check should be solely
11408          * predicated on the sequence space of this segment. 3) That we
11409          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11410          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11411          * SEG.Len, This modified check allows us to overcome RFC1323's
11412          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11413          * p.869. In such cases, we can still calculate the RTT correctly
11414          * when RCV.NXT == Last.ACK.Sent.
11415          */
11416         if ((to->to_flags & TOF_TS) != 0 &&
11417             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11418             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11419             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11420                 tp->ts_recent_age = tcp_ts_getticks();
11421                 tp->ts_recent = to->to_tsval;
11422         }
11423         /*
11424          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11425          * is on (half-synchronized state), then queue data for later
11426          * processing; else drop segment and return.
11427          */
11428         if ((thflags & TH_ACK) == 0) {
11429                 if (tp->t_flags & TF_NEEDSYN) {
11430                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11431                             tiwin, thflags, nxt_pkt));
11432
11433                 } else if (tp->t_flags & TF_ACKNOW) {
11434                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11435                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11436                         return (ret_val);
11437                 } else {
11438                         ctf_do_drop(m, NULL);
11439                         return (0);
11440                 }
11441         }
11442         /*
11443          * Ack processing.
11444          */
11445         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11446                 return (ret_val);
11447         }
11448         if (sbavail(&so->so_snd)) {
11449                 if (ctf_progress_timeout_check(tp, true)) {
11450                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11451                                                 tp, tick, PROGRESS_DROP, __LINE__);
11452                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11453                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11454                         return (1);
11455                 }
11456         }
11457         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11458             tiwin, thflags, nxt_pkt));
11459 }
11460
11461 static int
11462 rack_check_data_after_close(struct mbuf *m,
11463     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11464 {
11465         struct tcp_rack *rack;
11466
11467         rack = (struct tcp_rack *)tp->t_fb_ptr;
11468         if (rack->rc_allow_data_af_clo == 0) {
11469         close_now:
11470                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11471                 /* tcp_close will kill the inp pre-log the Reset */
11472                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11473                 tp = tcp_close(tp);
11474                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11475                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11476                 return (1);
11477         }
11478         if (sbavail(&so->so_snd) == 0)
11479                 goto close_now;
11480         /* Ok we allow data that is ignored and a followup reset */
11481         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11482         tp->rcv_nxt = th->th_seq + *tlen;
11483         tp->t_flags2 |= TF2_DROP_AF_DATA;
11484         rack->r_wanted_output = 1;
11485         *tlen = 0;
11486         return (0);
11487 }
11488
11489 /*
11490  * Return value of 1, the TCB is unlocked and most
11491  * likely gone, return value of 0, the TCP is still
11492  * locked.
11493  */
11494 static int
11495 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11496     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11497     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11498 {
11499         int32_t ret_val = 0;
11500         int32_t ourfinisacked = 0;
11501         struct tcp_rack *rack;
11502
11503         rack = (struct tcp_rack *)tp->t_fb_ptr;
11504         ctf_calc_rwin(so, tp);
11505
11506         if ((thflags & TH_RST) ||
11507             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11508                 return (ctf_process_rst(m, th, so, tp));
11509         /*
11510          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11511          * synchronized state.
11512          */
11513         if (thflags & TH_SYN) {
11514                 ctf_challenge_ack(m, th, tp, &ret_val);
11515                 return (ret_val);
11516         }
11517         /*
11518          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11519          * it's less than ts_recent, drop it.
11520          */
11521         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11522             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11523                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11524                         return (ret_val);
11525         }
11526         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11527                               &rack->r_ctl.challenge_ack_ts,
11528                               &rack->r_ctl.challenge_ack_cnt)) {
11529                 return (ret_val);
11530         }
11531         /*
11532          * If new data are received on a connection after the user processes
11533          * are gone, then RST the other end.
11534          */
11535         if ((so->so_state & SS_NOFDREF) && tlen) {
11536                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11537                         return (1);
11538         }
11539         /*
11540          * If last ACK falls within this segment's sequence numbers, record
11541          * its timestamp. NOTE: 1) That the test incorporates suggestions
11542          * from the latest proposal of the tcplw@cray.com list (Braden
11543          * 1993/04/26). 2) That updating only on newer timestamps interferes
11544          * with our earlier PAWS tests, so this check should be solely
11545          * predicated on the sequence space of this segment. 3) That we
11546          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11547          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11548          * SEG.Len, This modified check allows us to overcome RFC1323's
11549          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11550          * p.869. In such cases, we can still calculate the RTT correctly
11551          * when RCV.NXT == Last.ACK.Sent.
11552          */
11553         if ((to->to_flags & TOF_TS) != 0 &&
11554             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11555             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11556             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11557                 tp->ts_recent_age = tcp_ts_getticks();
11558                 tp->ts_recent = to->to_tsval;
11559         }
11560         /*
11561          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11562          * is on (half-synchronized state), then queue data for later
11563          * processing; else drop segment and return.
11564          */
11565         if ((thflags & TH_ACK) == 0) {
11566                 if (tp->t_flags & TF_NEEDSYN) {
11567                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11568                             tiwin, thflags, nxt_pkt));
11569                 } else if (tp->t_flags & TF_ACKNOW) {
11570                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11571                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11572                         return (ret_val);
11573                 } else {
11574                         ctf_do_drop(m, NULL);
11575                         return (0);
11576                 }
11577         }
11578         /*
11579          * Ack processing.
11580          */
11581         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11582                 return (ret_val);
11583         }
11584         if (ourfinisacked) {
11585                 /*
11586                  * If we can't receive any more data, then closing user can
11587                  * proceed. Starting the timer is contrary to the
11588                  * specification, but if we don't get a FIN we'll hang
11589                  * forever.
11590                  *
11591                  * XXXjl: we should release the tp also, and use a
11592                  * compressed state.
11593                  */
11594                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11595                         soisdisconnected(so);
11596                         tcp_timer_activate(tp, TT_2MSL,
11597                             (tcp_fast_finwait2_recycle ?
11598                             tcp_finwait2_timeout :
11599                             TP_MAXIDLE(tp)));
11600                 }
11601                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11602         }
11603         if (sbavail(&so->so_snd)) {
11604                 if (ctf_progress_timeout_check(tp, true)) {
11605                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11606                                                 tp, tick, PROGRESS_DROP, __LINE__);
11607                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11608                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11609                         return (1);
11610                 }
11611         }
11612         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11613             tiwin, thflags, nxt_pkt));
11614 }
11615
11616 /*
11617  * Return value of 1, the TCB is unlocked and most
11618  * likely gone, return value of 0, the TCP is still
11619  * locked.
11620  */
11621 static int
11622 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11623     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11624     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11625 {
11626         int32_t ret_val = 0;
11627         int32_t ourfinisacked = 0;
11628         struct tcp_rack *rack;
11629
11630         rack = (struct tcp_rack *)tp->t_fb_ptr;
11631         ctf_calc_rwin(so, tp);
11632
11633         if ((thflags & TH_RST) ||
11634             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11635                 return (ctf_process_rst(m, th, so, tp));
11636         /*
11637          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11638          * synchronized state.
11639          */
11640         if (thflags & TH_SYN) {
11641                 ctf_challenge_ack(m, th, tp, &ret_val);
11642                 return (ret_val);
11643         }
11644         /*
11645          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11646          * it's less than ts_recent, drop it.
11647          */
11648         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11649             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11650                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11651                         return (ret_val);
11652         }
11653         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11654                               &rack->r_ctl.challenge_ack_ts,
11655                               &rack->r_ctl.challenge_ack_cnt)) {
11656                 return (ret_val);
11657         }
11658         /*
11659          * If new data are received on a connection after the user processes
11660          * are gone, then RST the other end.
11661          */
11662         if ((so->so_state & SS_NOFDREF) && tlen) {
11663                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11664                         return (1);
11665         }
11666         /*
11667          * If last ACK falls within this segment's sequence numbers, record
11668          * its timestamp. NOTE: 1) That the test incorporates suggestions
11669          * from the latest proposal of the tcplw@cray.com list (Braden
11670          * 1993/04/26). 2) That updating only on newer timestamps interferes
11671          * with our earlier PAWS tests, so this check should be solely
11672          * predicated on the sequence space of this segment. 3) That we
11673          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11674          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11675          * SEG.Len, This modified check allows us to overcome RFC1323's
11676          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11677          * p.869. In such cases, we can still calculate the RTT correctly
11678          * when RCV.NXT == Last.ACK.Sent.
11679          */
11680         if ((to->to_flags & TOF_TS) != 0 &&
11681             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11682             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11683             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11684                 tp->ts_recent_age = tcp_ts_getticks();
11685                 tp->ts_recent = to->to_tsval;
11686         }
11687         /*
11688          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11689          * is on (half-synchronized state), then queue data for later
11690          * processing; else drop segment and return.
11691          */
11692         if ((thflags & TH_ACK) == 0) {
11693                 if (tp->t_flags & TF_NEEDSYN) {
11694                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11695                             tiwin, thflags, nxt_pkt));
11696                 } else if (tp->t_flags & TF_ACKNOW) {
11697                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11698                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11699                         return (ret_val);
11700                 } else {
11701                         ctf_do_drop(m, NULL);
11702                         return (0);
11703                 }
11704         }
11705         /*
11706          * Ack processing.
11707          */
11708         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11709                 return (ret_val);
11710         }
11711         if (ourfinisacked) {
11712                 tcp_twstart(tp);
11713                 m_freem(m);
11714                 return (1);
11715         }
11716         if (sbavail(&so->so_snd)) {
11717                 if (ctf_progress_timeout_check(tp, true)) {
11718                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11719                                                 tp, tick, PROGRESS_DROP, __LINE__);
11720                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11721                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11722                         return (1);
11723                 }
11724         }
11725         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11726             tiwin, thflags, nxt_pkt));
11727 }
11728
11729 /*
11730  * Return value of 1, the TCB is unlocked and most
11731  * likely gone, return value of 0, the TCP is still
11732  * locked.
11733  */
11734 static int
11735 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11736     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11737     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11738 {
11739         int32_t ret_val = 0;
11740         int32_t ourfinisacked = 0;
11741         struct tcp_rack *rack;
11742
11743         rack = (struct tcp_rack *)tp->t_fb_ptr;
11744         ctf_calc_rwin(so, tp);
11745
11746         if ((thflags & TH_RST) ||
11747             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11748                 return (ctf_process_rst(m, th, so, tp));
11749         /*
11750          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11751          * synchronized state.
11752          */
11753         if (thflags & TH_SYN) {
11754                 ctf_challenge_ack(m, th, tp, &ret_val);
11755                 return (ret_val);
11756         }
11757         /*
11758          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11759          * it's less than ts_recent, drop it.
11760          */
11761         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11762             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11763                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11764                         return (ret_val);
11765         }
11766         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11767                               &rack->r_ctl.challenge_ack_ts,
11768                               &rack->r_ctl.challenge_ack_cnt)) {
11769                 return (ret_val);
11770         }
11771         /*
11772          * If new data are received on a connection after the user processes
11773          * are gone, then RST the other end.
11774          */
11775         if ((so->so_state & SS_NOFDREF) && tlen) {
11776                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11777                         return (1);
11778         }
11779         /*
11780          * If last ACK falls within this segment's sequence numbers, record
11781          * its timestamp. NOTE: 1) That the test incorporates suggestions
11782          * from the latest proposal of the tcplw@cray.com list (Braden
11783          * 1993/04/26). 2) That updating only on newer timestamps interferes
11784          * with our earlier PAWS tests, so this check should be solely
11785          * predicated on the sequence space of this segment. 3) That we
11786          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11787          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11788          * SEG.Len, This modified check allows us to overcome RFC1323's
11789          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11790          * p.869. In such cases, we can still calculate the RTT correctly
11791          * when RCV.NXT == Last.ACK.Sent.
11792          */
11793         if ((to->to_flags & TOF_TS) != 0 &&
11794             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11795             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11796             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11797                 tp->ts_recent_age = tcp_ts_getticks();
11798                 tp->ts_recent = to->to_tsval;
11799         }
11800         /*
11801          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11802          * is on (half-synchronized state), then queue data for later
11803          * processing; else drop segment and return.
11804          */
11805         if ((thflags & TH_ACK) == 0) {
11806                 if (tp->t_flags & TF_NEEDSYN) {
11807                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11808                             tiwin, thflags, nxt_pkt));
11809                 } else if (tp->t_flags & TF_ACKNOW) {
11810                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11811                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11812                         return (ret_val);
11813                 } else {
11814                         ctf_do_drop(m, NULL);
11815                         return (0);
11816                 }
11817         }
11818         /*
11819          * case TCPS_LAST_ACK: Ack processing.
11820          */
11821         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11822                 return (ret_val);
11823         }
11824         if (ourfinisacked) {
11825                 tp = tcp_close(tp);
11826                 ctf_do_drop(m, tp);
11827                 return (1);
11828         }
11829         if (sbavail(&so->so_snd)) {
11830                 if (ctf_progress_timeout_check(tp, true)) {
11831                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11832                                                 tp, tick, PROGRESS_DROP, __LINE__);
11833                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11834                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11835                         return (1);
11836                 }
11837         }
11838         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11839             tiwin, thflags, nxt_pkt));
11840 }
11841
11842 /*
11843  * Return value of 1, the TCB is unlocked and most
11844  * likely gone, return value of 0, the TCP is still
11845  * locked.
11846  */
11847 static int
11848 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
11849     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11850     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11851 {
11852         int32_t ret_val = 0;
11853         int32_t ourfinisacked = 0;
11854         struct tcp_rack *rack;
11855
11856         rack = (struct tcp_rack *)tp->t_fb_ptr;
11857         ctf_calc_rwin(so, tp);
11858
11859         /* Reset receive buffer auto scaling when not in bulk receive mode. */
11860         if ((thflags & TH_RST) ||
11861             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11862                 return (ctf_process_rst(m, th, so, tp));
11863         /*
11864          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11865          * synchronized state.
11866          */
11867         if (thflags & TH_SYN) {
11868                 ctf_challenge_ack(m, th, tp, &ret_val);
11869                 return (ret_val);
11870         }
11871         /*
11872          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11873          * it's less than ts_recent, drop it.
11874          */
11875         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11876             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11877                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11878                         return (ret_val);
11879         }
11880         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11881                               &rack->r_ctl.challenge_ack_ts,
11882                               &rack->r_ctl.challenge_ack_cnt)) {
11883                 return (ret_val);
11884         }
11885         /*
11886          * If new data are received on a connection after the user processes
11887          * are gone, then RST the other end.
11888          */
11889         if ((so->so_state & SS_NOFDREF) &&
11890             tlen) {
11891                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11892                         return (1);
11893         }
11894         /*
11895          * If last ACK falls within this segment's sequence numbers, record
11896          * its timestamp. NOTE: 1) That the test incorporates suggestions
11897          * from the latest proposal of the tcplw@cray.com list (Braden
11898          * 1993/04/26). 2) That updating only on newer timestamps interferes
11899          * with our earlier PAWS tests, so this check should be solely
11900          * predicated on the sequence space of this segment. 3) That we
11901          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11902          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11903          * SEG.Len, This modified check allows us to overcome RFC1323's
11904          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11905          * p.869. In such cases, we can still calculate the RTT correctly
11906          * when RCV.NXT == Last.ACK.Sent.
11907          */
11908         if ((to->to_flags & TOF_TS) != 0 &&
11909             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11910             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11911             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11912                 tp->ts_recent_age = tcp_ts_getticks();
11913                 tp->ts_recent = to->to_tsval;
11914         }
11915         /*
11916          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11917          * is on (half-synchronized state), then queue data for later
11918          * processing; else drop segment and return.
11919          */
11920         if ((thflags & TH_ACK) == 0) {
11921                 if (tp->t_flags & TF_NEEDSYN) {
11922                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11923                             tiwin, thflags, nxt_pkt));
11924                 } else if (tp->t_flags & TF_ACKNOW) {
11925                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11926                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11927                         return (ret_val);
11928                 } else {
11929                         ctf_do_drop(m, NULL);
11930                         return (0);
11931                 }
11932         }
11933         /*
11934          * Ack processing.
11935          */
11936         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11937                 return (ret_val);
11938         }
11939         if (sbavail(&so->so_snd)) {
11940                 if (ctf_progress_timeout_check(tp, true)) {
11941                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11942                                                 tp, tick, PROGRESS_DROP, __LINE__);
11943                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11944                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11945                         return (1);
11946                 }
11947         }
11948         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11949             tiwin, thflags, nxt_pkt));
11950 }
11951
11952 static void inline
11953 rack_clear_rate_sample(struct tcp_rack *rack)
11954 {
11955         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
11956         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
11957         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
11958 }
11959
11960 static void
11961 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
11962 {
11963         uint64_t bw_est, rate_wanted;
11964         int chged = 0;
11965         uint32_t user_max, orig_min, orig_max;
11966
11967         orig_min = rack->r_ctl.rc_pace_min_segs;
11968         orig_max = rack->r_ctl.rc_pace_max_segs;
11969         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
11970         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
11971                 chged = 1;
11972         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
11973         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
11974                 if (user_max != rack->r_ctl.rc_pace_max_segs)
11975                         chged = 1;
11976         }
11977         if (rack->rc_force_max_seg) {
11978                 rack->r_ctl.rc_pace_max_segs = user_max;
11979         } else if (rack->use_fixed_rate) {
11980                 bw_est = rack_get_bw(rack);
11981                 if ((rack->r_ctl.crte == NULL) ||
11982                     (bw_est != rack->r_ctl.crte->rate)) {
11983                         rack->r_ctl.rc_pace_max_segs = user_max;
11984                 } else {
11985                         /* We are pacing right at the hardware rate */
11986                         uint32_t segsiz;
11987
11988                         segsiz = min(ctf_fixed_maxseg(tp),
11989                                      rack->r_ctl.rc_pace_min_segs);
11990                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
11991                                                            tp, bw_est, segsiz, 0,
11992                                                            rack->r_ctl.crte, NULL);
11993                 }
11994         } else if (rack->rc_always_pace) {
11995                 if (rack->r_ctl.gp_bw ||
11996 #ifdef NETFLIX_PEAKRATE
11997                     rack->rc_tp->t_maxpeakrate ||
11998 #endif
11999                     rack->r_ctl.init_rate) {
12000                         /* We have a rate of some sort set */
12001                         uint32_t  orig;
12002
12003                         bw_est = rack_get_bw(rack);
12004                         orig = rack->r_ctl.rc_pace_max_segs;
12005                         if (fill_override)
12006                                 rate_wanted = *fill_override;
12007                         else
12008                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12009                         if (rate_wanted) {
12010                                 /* We have something */
12011                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12012                                                                                    rate_wanted,
12013                                                                                    ctf_fixed_maxseg(rack->rc_tp));
12014                         } else
12015                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12016                         if (orig != rack->r_ctl.rc_pace_max_segs)
12017                                 chged = 1;
12018                 } else if ((rack->r_ctl.gp_bw == 0) &&
12019                            (rack->r_ctl.rc_pace_max_segs == 0)) {
12020                         /*
12021                          * If we have nothing limit us to bursting
12022                          * out IW sized pieces.
12023                          */
12024                         chged = 1;
12025                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12026                 }
12027         }
12028         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12029                 chged = 1;
12030                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12031         }
12032         if (chged)
12033                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12034 }
12035
12036
12037 static void
12038 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12039 {
12040 #ifdef INET6
12041         struct ip6_hdr *ip6 = NULL;
12042 #endif
12043 #ifdef INET
12044         struct ip *ip = NULL;
12045 #endif
12046 #if defined(INET) || defined(INET6)
12047         struct udphdr *udp = NULL;
12048 #endif
12049
12050         /* Ok lets fill in the fast block, it can only be used with no IP options! */
12051 #ifdef INET6
12052         if (rack->r_is_v6) {
12053                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12054                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12055                 if (tp->t_port) {
12056                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12057                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12058                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12059                         udp->uh_dport = tp->t_port;
12060                         rack->r_ctl.fsb.udp = udp;
12061                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12062                 } else
12063                 {
12064                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12065                         rack->r_ctl.fsb.udp = NULL;
12066                 }
12067                 tcpip_fillheaders(rack->rc_inp,
12068                                   tp->t_port,
12069                                   ip6, rack->r_ctl.fsb.th);
12070         } else
12071 #endif                          /* INET6 */
12072 #ifdef INET
12073         {
12074                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12075                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12076                 if (tp->t_port) {
12077                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12078                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12079                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12080                         udp->uh_dport = tp->t_port;
12081                         rack->r_ctl.fsb.udp = udp;
12082                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12083                 } else
12084                 {
12085                         rack->r_ctl.fsb.udp = NULL;
12086                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12087                 }
12088                 tcpip_fillheaders(rack->rc_inp,
12089                                   tp->t_port,
12090                                   ip, rack->r_ctl.fsb.th);
12091         }
12092 #endif
12093         rack->r_fsb_inited = 1;
12094 }
12095
12096 static int
12097 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12098 {
12099         /*
12100          * Allocate the larger of spaces V6 if available else just
12101          * V4 and include udphdr (overbook)
12102          */
12103 #ifdef INET6
12104         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12105 #else
12106         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12107 #endif
12108         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12109                                             M_TCPFSB, M_NOWAIT|M_ZERO);
12110         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12111                 return (ENOMEM);
12112         }
12113         rack->r_fsb_inited = 0;
12114         return (0);
12115 }
12116
12117 static int
12118 rack_init(struct tcpcb *tp)
12119 {
12120         struct tcp_rack *rack = NULL;
12121         struct rack_sendmap *insret;
12122         uint32_t iwin, snt, us_cts;
12123         int err;
12124
12125         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12126         if (tp->t_fb_ptr == NULL) {
12127                 /*
12128                  * We need to allocate memory but cant. The INP and INP_INFO
12129                  * locks and they are recusive (happens during setup. So a
12130                  * scheme to drop the locks fails :(
12131                  *
12132                  */
12133                 return (ENOMEM);
12134         }
12135         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12136
12137         rack = (struct tcp_rack *)tp->t_fb_ptr;
12138         RB_INIT(&rack->r_ctl.rc_mtree);
12139         TAILQ_INIT(&rack->r_ctl.rc_free);
12140         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12141         rack->rc_tp = tp;
12142         rack->rc_inp = tp->t_inpcb;
12143         /* Set the flag */
12144         rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12145         /* Probably not needed but lets be sure */
12146         rack_clear_rate_sample(rack);
12147         /*
12148          * Save off the default values, socket options will poke
12149          * at these if pacing is not on or we have not yet
12150          * reached where pacing is on (gp_ready/fixed enabled).
12151          * When they get set into the CC module (when gp_ready
12152          * is enabled or we enable fixed) then we will set these
12153          * values into the CC and place in here the old values
12154          * so we have a restoral. Then we will set the flag
12155          * rc_pacing_cc_set. That way whenever we turn off pacing
12156          * or switch off this stack, we will know to go restore
12157          * the saved values.
12158          */
12159         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12160         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12161         /* We want abe like behavior as well */
12162         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
12163         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12164         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12165         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12166         if (use_rack_rr)
12167                 rack->use_rack_rr = 1;
12168         if (V_tcp_delack_enabled)
12169                 tp->t_delayed_ack = 1;
12170         else
12171                 tp->t_delayed_ack = 0;
12172 #ifdef TCP_ACCOUNTING
12173         if (rack_tcp_accounting) {
12174                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12175         }
12176 #endif
12177         if (rack_enable_shared_cwnd)
12178                 rack->rack_enable_scwnd = 1;
12179         rack->rc_user_set_max_segs = rack_hptsi_segments;
12180         rack->rc_force_max_seg = 0;
12181         if (rack_use_imac_dack)
12182                 rack->rc_dack_mode = 1;
12183         TAILQ_INIT(&rack->r_ctl.opt_list);
12184         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12185         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12186         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12187         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12188         rack->r_ctl.rc_highest_us_rtt = 0;
12189         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12190         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12191         if (rack_use_cmp_acks)
12192                 rack->r_use_cmp_ack = 1;
12193         if (rack_disable_prr)
12194                 rack->rack_no_prr = 1;
12195         if (rack_gp_no_rec_chg)
12196                 rack->rc_gp_no_rec_chg = 1;
12197         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12198                 rack->rc_always_pace = 1;
12199                 if (rack->use_fixed_rate || rack->gp_ready)
12200                         rack_set_cc_pacing(rack);
12201         } else
12202                 rack->rc_always_pace = 0;
12203         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12204                 rack->r_mbuf_queue = 1;
12205         else
12206                 rack->r_mbuf_queue = 0;
12207         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12208                 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12209         else
12210                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12211         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12212         if (rack_limits_scwnd)
12213                 rack->r_limit_scw = 1;
12214         else
12215                 rack->r_limit_scw = 0;
12216         rack->rc_labc = V_tcp_abc_l_var;
12217         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12218         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12219         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12220         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12221         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12222         rack->r_ctl.rc_min_to = rack_min_to;
12223         microuptime(&rack->r_ctl.act_rcv_time);
12224         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12225         rack->r_running_late = 0;
12226         rack->r_running_early = 0;
12227         rack->rc_init_win = rack_default_init_window;
12228         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12229         if (rack_hw_up_only)
12230                 rack->r_up_only = 1;
12231         if (rack_do_dyn_mul) {
12232                 /* When dynamic adjustment is on CA needs to start at 100% */
12233                 rack->rc_gp_dyn_mul = 1;
12234                 if (rack_do_dyn_mul >= 100)
12235                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12236         } else
12237                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12238         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12239         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12240         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12241         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12242                                 rack_probertt_filter_life);
12243         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12244         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12245         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12246         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12247         rack->r_ctl.rc_time_probertt_starts = 0;
12248         /* We require at least one measurement, even if the sysctl is 0 */
12249         if (rack_req_measurements)
12250                 rack->r_ctl.req_measurements = rack_req_measurements;
12251         else
12252                 rack->r_ctl.req_measurements = 1;
12253         if (rack_enable_hw_pacing)
12254                 rack->rack_hdw_pace_ena = 1;
12255         if (rack_hw_rate_caps)
12256                 rack->r_rack_hw_rate_caps = 1;
12257         /* Do we force on detection? */
12258 #ifdef NETFLIX_EXP_DETECTION
12259         if (tcp_force_detection)
12260                 rack->do_detection = 1;
12261         else
12262 #endif
12263                 rack->do_detection = 0;
12264         if (rack_non_rxt_use_cr)
12265                 rack->rack_rec_nonrxt_use_cr = 1;
12266         err = rack_init_fsb(tp, rack);
12267         if (err) {
12268                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12269                 tp->t_fb_ptr = NULL;
12270                 return (err);
12271         }
12272         if (tp->snd_una != tp->snd_max) {
12273                 /* Create a send map for the current outstanding data */
12274                 struct rack_sendmap *rsm;
12275
12276                 rsm = rack_alloc(rack);
12277                 if (rsm == NULL) {
12278                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12279                         tp->t_fb_ptr = NULL;
12280                         return (ENOMEM);
12281                 }
12282                 rsm->r_no_rtt_allowed = 1;
12283                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12284                 rsm->r_rtr_cnt = 1;
12285                 rsm->r_rtr_bytes = 0;
12286                 if (tp->t_flags & TF_SENTFIN) {
12287                         rsm->r_end = tp->snd_max - 1;
12288                         rsm->r_flags |= RACK_HAS_FIN;
12289                 } else {
12290                         rsm->r_end = tp->snd_max;
12291                 }
12292                 if (tp->snd_una == tp->iss) {
12293                         /* The data space is one beyond snd_una */
12294                         rsm->r_flags |= RACK_HAS_SYN;
12295                         rsm->r_start = tp->iss;
12296                         rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12297                 } else
12298                         rsm->r_start = tp->snd_una;
12299                 rsm->r_dupack = 0;
12300                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12301                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12302                         if (rsm->m)
12303                                 rsm->orig_m_len = rsm->m->m_len;
12304                         else
12305                                 rsm->orig_m_len = 0;
12306                 } else {
12307                         /*
12308                          * This can happen if we have a stand-alone FIN or
12309                          *  SYN.
12310                          */
12311                         rsm->m = NULL;
12312                         rsm->orig_m_len = 0;
12313                         rsm->soff = 0;
12314                 }
12315                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12316 #ifdef INVARIANTS
12317                 if (insret != NULL) {
12318                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12319                               insret, rack, rsm);
12320                 }
12321 #endif
12322                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12323                 rsm->r_in_tmap = 1;
12324         }
12325         /*
12326          * Timers in Rack are kept in microseconds so lets
12327          * convert any initial incoming variables
12328          * from ticks into usecs. Note that we
12329          * also change the values of t_srtt and t_rttvar, if
12330          * they are non-zero. They are kept with a 5
12331          * bit decimal so we have to carefully convert
12332          * these to get the full precision.
12333          */
12334         rack_convert_rtts(tp);
12335         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12336         if (rack_def_profile)
12337                 rack_set_profile(rack, rack_def_profile);
12338         /* Cancel the GP measurement in progress */
12339         tp->t_flags &= ~TF_GPUTINPROG;
12340         if (SEQ_GT(tp->snd_max, tp->iss))
12341                 snt = tp->snd_max - tp->iss;
12342         else
12343                 snt = 0;
12344         iwin = rc_init_window(rack);
12345         if (snt < iwin) {
12346                 /* We are not past the initial window
12347                  * so we need to make sure cwnd is
12348                  * correct.
12349                  */
12350                 if (tp->snd_cwnd < iwin)
12351                         tp->snd_cwnd = iwin;
12352                 /*
12353                  * If we are within the initial window
12354                  * we want ssthresh to be unlimited. Setting
12355                  * it to the rwnd (which the default stack does
12356                  * and older racks) is not really a good idea
12357                  * since we want to be in SS and grow both the
12358                  * cwnd and the rwnd (via dynamic rwnd growth). If
12359                  * we set it to the rwnd then as the peer grows its
12360                  * rwnd we will be stuck in CA and never hit SS.
12361                  *
12362                  * Its far better to raise it up high (this takes the
12363                  * risk that there as been a loss already, probably
12364                  * we should have an indicator in all stacks of loss
12365                  * but we don't), but considering the normal use this
12366                  * is a risk worth taking. The consequences of not
12367                  * hitting SS are far worse than going one more time
12368                  * into it early on (before we have sent even a IW).
12369                  * It is highly unlikely that we will have had a loss
12370                  * before getting the IW out.
12371                  */
12372                 tp->snd_ssthresh = 0xffffffff;
12373         }
12374         rack_stop_all_timers(tp);
12375         /* Lets setup the fsb block */
12376         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12377         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12378                              __LINE__, RACK_RTTS_INIT);
12379         return (0);
12380 }
12381
12382 static int
12383 rack_handoff_ok(struct tcpcb *tp)
12384 {
12385         if ((tp->t_state == TCPS_CLOSED) ||
12386             (tp->t_state == TCPS_LISTEN)) {
12387                 /* Sure no problem though it may not stick */
12388                 return (0);
12389         }
12390         if ((tp->t_state == TCPS_SYN_SENT) ||
12391             (tp->t_state == TCPS_SYN_RECEIVED)) {
12392                 /*
12393                  * We really don't know if you support sack,
12394                  * you have to get to ESTAB or beyond to tell.
12395                  */
12396                 return (EAGAIN);
12397         }
12398         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12399                 /*
12400                  * Rack will only send a FIN after all data is acknowledged.
12401                  * So in this case we have more data outstanding. We can't
12402                  * switch stacks until either all data and only the FIN
12403                  * is left (in which case rack_init() now knows how
12404                  * to deal with that) <or> all is acknowledged and we
12405                  * are only left with incoming data, though why you
12406                  * would want to switch to rack after all data is acknowledged
12407                  * I have no idea (rrs)!
12408                  */
12409                 return (EAGAIN);
12410         }
12411         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12412                 return (0);
12413         }
12414         /*
12415          * If we reach here we don't do SACK on this connection so we can
12416          * never do rack.
12417          */
12418         return (EINVAL);
12419 }
12420
12421
12422 static void
12423 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12424 {
12425         int ack_cmp = 0;
12426
12427         if (tp->t_fb_ptr) {
12428                 struct tcp_rack *rack;
12429                 struct rack_sendmap *rsm, *nrsm, *rm;
12430
12431                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12432                 if (tp->t_in_pkt) {
12433                         /*
12434                          * It is unsafe to process the packets since a
12435                          * reset may be lurking in them (its rare but it
12436                          * can occur). If we were to find a RST, then we
12437                          * would end up dropping the connection and the
12438                          * INP lock, so when we return the caller (tcp_usrreq)
12439                          * will blow up when it trys to unlock the inp.
12440                          */
12441                         struct mbuf *save, *m;
12442
12443                         m = tp->t_in_pkt;
12444                         tp->t_in_pkt = NULL;
12445                         tp->t_tail_pkt = NULL;
12446                         while (m) {
12447                                 save = m->m_nextpkt;
12448                                 m->m_nextpkt = NULL;
12449                                 m_freem(m);
12450                                 m = save;
12451                         }
12452                         if ((tp->t_inpcb) &&
12453                             (tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP))
12454                                 ack_cmp = 1;
12455                         if (ack_cmp) {
12456                                 /* Total if we used large or small (if ack-cmp was used). */
12457                                 if (rack->rc_inp->inp_flags2 & INP_MBUF_L_ACKS)
12458                                         counter_u64_add(rack_large_ackcmp, 1);
12459                                 else
12460                                         counter_u64_add(rack_small_ackcmp, 1);
12461                         }
12462                 }
12463                 tp->t_flags &= ~TF_FORCEDATA;
12464 #ifdef NETFLIX_SHARED_CWND
12465                 if (rack->r_ctl.rc_scw) {
12466                         uint32_t limit;
12467
12468                         if (rack->r_limit_scw)
12469                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12470                         else
12471                                 limit = 0;
12472                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12473                                                   rack->r_ctl.rc_scw_index,
12474                                                   limit);
12475                         rack->r_ctl.rc_scw = NULL;
12476                 }
12477 #endif
12478                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12479                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12480                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12481                         rack->r_ctl.fsb.th = NULL;
12482                 }
12483                 /* Convert back to ticks, with  */
12484                 if (tp->t_srtt > 1) {
12485                         uint32_t val, frac;
12486
12487                         val = USEC_2_TICKS(tp->t_srtt);
12488                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12489                         tp->t_srtt = val << TCP_RTT_SHIFT;
12490                         /*
12491                          * frac is the fractional part here is left
12492                          * over from converting to hz and shifting.
12493                          * We need to convert this to the 5 bit
12494                          * remainder.
12495                          */
12496                         if (frac) {
12497                                 if (hz == 1000) {
12498                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12499                                 } else {
12500                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12501                                 }
12502                                 tp->t_srtt += frac;
12503                         }
12504                 }
12505                 if (tp->t_rttvar) {
12506                         uint32_t val, frac;
12507
12508                         val = USEC_2_TICKS(tp->t_rttvar);
12509                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12510                         tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12511                         /*
12512                          * frac is the fractional part here is left
12513                          * over from converting to hz and shifting.
12514                          * We need to convert this to the 5 bit
12515                          * remainder.
12516                          */
12517                         if (frac) {
12518                                 if (hz == 1000) {
12519                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12520                                 } else {
12521                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12522                                 }
12523                                 tp->t_rttvar += frac;
12524                         }
12525                 }
12526                 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12527                 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12528                 if (rack->rc_always_pace) {
12529                         tcp_decrement_paced_conn();
12530                         rack_undo_cc_pacing(rack);
12531                         rack->rc_always_pace = 0;
12532                 }
12533                 /* Clean up any options if they were not applied */
12534                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12535                         struct deferred_opt_list *dol;
12536
12537                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12538                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12539                         free(dol, M_TCPDO);
12540                 }
12541                 /* rack does not use force data but other stacks may clear it */
12542                 if (rack->r_ctl.crte != NULL) {
12543                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12544                         rack->rack_hdrw_pacing = 0;
12545                         rack->r_ctl.crte = NULL;
12546                 }
12547 #ifdef TCP_BLACKBOX
12548                 tcp_log_flowend(tp);
12549 #endif
12550                 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12551                         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12552 #ifdef INVARIANTS
12553                         if (rm != rsm) {
12554                                 panic("At fini, rack:%p rsm:%p rm:%p",
12555                                       rack, rsm, rm);
12556                         }
12557 #endif
12558                         uma_zfree(rack_zone, rsm);
12559                 }
12560                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12561                 while (rsm) {
12562                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12563                         uma_zfree(rack_zone, rsm);
12564                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12565                 }
12566                 rack->rc_free_cnt = 0;
12567                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12568                 tp->t_fb_ptr = NULL;
12569         }
12570         if (tp->t_inpcb) {
12571                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12572                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12573                 tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12574                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
12575                 /* Cancel the GP measurement in progress */
12576                 tp->t_flags &= ~TF_GPUTINPROG;
12577                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
12578         }
12579         /* Make sure snd_nxt is correctly set */
12580         tp->snd_nxt = tp->snd_max;
12581 }
12582
12583 static void
12584 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12585 {
12586         if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12587                 rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12588         }
12589         switch (tp->t_state) {
12590         case TCPS_SYN_SENT:
12591                 rack->r_state = TCPS_SYN_SENT;
12592                 rack->r_substate = rack_do_syn_sent;
12593                 break;
12594         case TCPS_SYN_RECEIVED:
12595                 rack->r_state = TCPS_SYN_RECEIVED;
12596                 rack->r_substate = rack_do_syn_recv;
12597                 break;
12598         case TCPS_ESTABLISHED:
12599                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12600                 rack->r_state = TCPS_ESTABLISHED;
12601                 rack->r_substate = rack_do_established;
12602                 break;
12603         case TCPS_CLOSE_WAIT:
12604                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12605                 rack->r_state = TCPS_CLOSE_WAIT;
12606                 rack->r_substate = rack_do_close_wait;
12607                 break;
12608         case TCPS_FIN_WAIT_1:
12609                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12610                 rack->r_state = TCPS_FIN_WAIT_1;
12611                 rack->r_substate = rack_do_fin_wait_1;
12612                 break;
12613         case TCPS_CLOSING:
12614                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12615                 rack->r_state = TCPS_CLOSING;
12616                 rack->r_substate = rack_do_closing;
12617                 break;
12618         case TCPS_LAST_ACK:
12619                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12620                 rack->r_state = TCPS_LAST_ACK;
12621                 rack->r_substate = rack_do_lastack;
12622                 break;
12623         case TCPS_FIN_WAIT_2:
12624                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12625                 rack->r_state = TCPS_FIN_WAIT_2;
12626                 rack->r_substate = rack_do_fin_wait_2;
12627                 break;
12628         case TCPS_LISTEN:
12629         case TCPS_CLOSED:
12630         case TCPS_TIME_WAIT:
12631         default:
12632                 break;
12633         };
12634         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12635                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12636
12637 }
12638
12639 static void
12640 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12641 {
12642         /*
12643          * We received an ack, and then did not
12644          * call send or were bounced out due to the
12645          * hpts was running. Now a timer is up as well, is
12646          * it the right timer?
12647          */
12648         struct rack_sendmap *rsm;
12649         int tmr_up;
12650
12651         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12652         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12653                 return;
12654         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12655         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12656             (tmr_up == PACE_TMR_RXT)) {
12657                 /* Should be an RXT */
12658                 return;
12659         }
12660         if (rsm == NULL) {
12661                 /* Nothing outstanding? */
12662                 if (tp->t_flags & TF_DELACK) {
12663                         if (tmr_up == PACE_TMR_DELACK)
12664                                 /* We are supposed to have delayed ack up and we do */
12665                                 return;
12666                 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12667                         /*
12668                          * if we hit enobufs then we would expect the possiblity
12669                          * of nothing outstanding and the RXT up (and the hptsi timer).
12670                          */
12671                         return;
12672                 } else if (((V_tcp_always_keepalive ||
12673                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12674                             (tp->t_state <= TCPS_CLOSING)) &&
12675                            (tmr_up == PACE_TMR_KEEP) &&
12676                            (tp->snd_max == tp->snd_una)) {
12677                         /* We should have keep alive up and we do */
12678                         return;
12679                 }
12680         }
12681         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12682                    ((tmr_up == PACE_TMR_TLP) ||
12683                     (tmr_up == PACE_TMR_RACK) ||
12684                     (tmr_up == PACE_TMR_RXT))) {
12685                 /*
12686                  * Either a Rack, TLP or RXT is fine if  we
12687                  * have outstanding data.
12688                  */
12689                 return;
12690         } else if (tmr_up == PACE_TMR_DELACK) {
12691                 /*
12692                  * If the delayed ack was going to go off
12693                  * before the rtx/tlp/rack timer were going to
12694                  * expire, then that would be the timer in control.
12695                  * Note we don't check the time here trusting the
12696                  * code is correct.
12697                  */
12698                 return;
12699         }
12700         /*
12701          * Ok the timer originally started is not what we want now.
12702          * We will force the hpts to be stopped if any, and restart
12703          * with the slot set to what was in the saved slot.
12704          */
12705         if (rack->rc_inp->inp_in_hpts) {
12706                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12707                         uint32_t us_cts;
12708
12709                         us_cts = tcp_get_usecs(NULL);
12710                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12711                                 rack->r_early = 1;
12712                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12713                         }
12714                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12715                 }
12716                 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
12717         }
12718         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12719         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12720 }
12721
12722
12723 static void
12724 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)
12725 {
12726         tp->snd_wnd = tiwin;
12727         rack_validate_fo_sendwin_up(tp, rack);
12728         tp->snd_wl1 = seq;
12729         tp->snd_wl2 = ack;
12730         if (tp->snd_wnd > tp->max_sndwnd)
12731                 tp->max_sndwnd = tp->snd_wnd;
12732         if (tp->snd_wnd < (tp->snd_max - high_seq)) {
12733                 /* The peer collapsed the window */
12734                 rack_collapsed_window(rack);
12735         } else if (rack->rc_has_collapsed)
12736                 rack_un_collapse_window(rack);
12737         /* Do we exit persists? */
12738         if ((rack->rc_in_persist != 0) &&
12739             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12740                                 rack->r_ctl.rc_pace_min_segs))) {
12741                 rack_exit_persist(tp, rack, cts);
12742         }
12743         /* Do we enter persists? */
12744         if ((rack->rc_in_persist == 0) &&
12745             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12746             TCPS_HAVEESTABLISHED(tp->t_state) &&
12747             (tp->snd_max == tp->snd_una) &&
12748             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
12749             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
12750                 /*
12751                  * Here the rwnd is less than
12752                  * the pacing size, we are established,
12753                  * nothing is outstanding, and there is
12754                  * data to send. Enter persists.
12755                  */
12756                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12757         }
12758 }
12759
12760 static void
12761 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
12762 {
12763
12764         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
12765                 union tcp_log_stackspecific log;
12766                 struct timeval ltv;
12767                 char tcp_hdr_buf[60];
12768                 struct tcphdr *th;
12769                 struct timespec ts;
12770                 uint32_t orig_snd_una;
12771                 uint8_t xx = 0;
12772
12773 #ifdef NETFLIX_HTTP_LOGGING
12774                 struct http_sendfile_track *http_req;
12775
12776                 if (SEQ_GT(ae->ack, tp->snd_una)) {
12777                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
12778                 } else {
12779                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
12780                 }
12781 #endif
12782                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
12783                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
12784                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
12785                 if (rack->rack_no_prr == 0)
12786                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
12787                 else
12788                         log.u_bbr.flex1 = 0;
12789                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
12790                 log.u_bbr.use_lt_bw <<= 1;
12791                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
12792                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
12793                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
12794                 log.u_bbr.pkts_out = tp->t_maxseg;
12795                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
12796                 log.u_bbr.flex7 = 1;
12797                 log.u_bbr.lost = ae->flags;
12798                 log.u_bbr.cwnd_gain = ackval;
12799                 log.u_bbr.pacing_gain = 0x2;
12800                 if (ae->flags & TSTMP_HDWR) {
12801                         /* Record the hardware timestamp if present */
12802                         log.u_bbr.flex3 = M_TSTMP;
12803                         ts.tv_sec = ae->timestamp / 1000000000;
12804                         ts.tv_nsec = ae->timestamp % 1000000000;
12805                         ltv.tv_sec = ts.tv_sec;
12806                         ltv.tv_usec = ts.tv_nsec / 1000;
12807                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
12808                 } else if (ae->flags & TSTMP_LRO) {
12809                         /* Record the LRO the arrival timestamp */
12810                         log.u_bbr.flex3 = M_TSTMP_LRO;
12811                         ts.tv_sec = ae->timestamp / 1000000000;
12812                         ts.tv_nsec = ae->timestamp % 1000000000;
12813                         ltv.tv_sec = ts.tv_sec;
12814                         ltv.tv_usec = ts.tv_nsec / 1000;
12815                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
12816                 }
12817                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
12818                 /* Log the rcv time */
12819                 log.u_bbr.delRate = ae->timestamp;
12820 #ifdef NETFLIX_HTTP_LOGGING
12821                 log.u_bbr.applimited = tp->t_http_closed;
12822                 log.u_bbr.applimited <<= 8;
12823                 log.u_bbr.applimited |= tp->t_http_open;
12824                 log.u_bbr.applimited <<= 8;
12825                 log.u_bbr.applimited |= tp->t_http_req;
12826                 if (http_req) {
12827                         /* Copy out any client req info */
12828                         /* seconds */
12829                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
12830                         /* useconds */
12831                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
12832                         log.u_bbr.rttProp = http_req->timestamp;
12833                         log.u_bbr.cur_del_rate = http_req->start;
12834                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
12835                                 log.u_bbr.flex8 |= 1;
12836                         } else {
12837                                 log.u_bbr.flex8 |= 2;
12838                                 log.u_bbr.bw_inuse = http_req->end;
12839                         }
12840                         log.u_bbr.flex6 = http_req->start_seq;
12841                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
12842                                 log.u_bbr.flex8 |= 4;
12843                                 log.u_bbr.epoch = http_req->end_seq;
12844                         }
12845                 }
12846 #endif
12847                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
12848                 th = (struct tcphdr *)tcp_hdr_buf;
12849                 th->th_seq = ae->seq;
12850                 th->th_ack = ae->ack;
12851                 th->th_win = ae->win;
12852                 /* Now fill in the ports */
12853                 th->th_sport = tp->t_inpcb->inp_fport;
12854                 th->th_dport = tp->t_inpcb->inp_lport;
12855                 th->th_flags = ae->flags & 0xff;
12856                 /* Now do we have a timestamp option? */
12857                 if (ae->flags & HAS_TSTMP) {
12858                         u_char *cp;
12859                         uint32_t val;
12860
12861                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
12862                         cp = (u_char *)(th + 1);
12863                         *cp = TCPOPT_NOP;
12864                         cp++;
12865                         *cp = TCPOPT_NOP;
12866                         cp++;
12867                         *cp = TCPOPT_TIMESTAMP;
12868                         cp++;
12869                         *cp = TCPOLEN_TIMESTAMP;
12870                         cp++;
12871                         val = htonl(ae->ts_value);
12872                         bcopy((char *)&val,
12873                               (char *)cp, sizeof(uint32_t));
12874                         val = htonl(ae->ts_echo);
12875                         bcopy((char *)&val,
12876                               (char *)(cp + 4), sizeof(uint32_t));
12877                 } else
12878                         th->th_off = (sizeof(struct tcphdr) >> 2);
12879
12880                 /*
12881                  * For sane logging we need to play a little trick.
12882                  * If the ack were fully processed we would have moved
12883                  * snd_una to high_seq, but since compressed acks are
12884                  * processed in two phases, at this point (logging) snd_una
12885                  * won't be advanced. So we would see multiple acks showing
12886                  * the advancement. We can prevent that by "pretending" that
12887                  * snd_una was advanced and then un-advancing it so that the
12888                  * logging code has the right value for tlb_snd_una.
12889                  */
12890                 if (tp->snd_una != high_seq) {
12891                         orig_snd_una = tp->snd_una;
12892                         tp->snd_una = high_seq;
12893                         xx = 1;
12894                 } else
12895                         xx = 0;
12896                 TCP_LOG_EVENTP(tp, th,
12897                                &tp->t_inpcb->inp_socket->so_rcv,
12898                                &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
12899                                0, &log, true, &ltv);
12900                 if (xx) {
12901                         tp->snd_una = orig_snd_una;
12902                 }
12903         }
12904
12905 }
12906
12907 static int
12908 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
12909 {
12910         /*
12911          * Handle a "special" compressed ack mbuf. Each incoming
12912          * ack has only four possible dispositions:
12913          *
12914          * A) It moves the cum-ack forward
12915          * B) It is behind the cum-ack.
12916          * C) It is a window-update ack.
12917          * D) It is a dup-ack.
12918          *
12919          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
12920          * in the incoming mbuf. We also need to still pay attention
12921          * to nxt_pkt since there may be another packet after this
12922          * one.
12923          */
12924 #ifdef TCP_ACCOUNTING
12925         uint64_t ts_val;
12926         uint64_t rdstc;
12927 #endif
12928         int segsiz;
12929         struct timespec ts;
12930         struct tcp_rack *rack;
12931         struct tcp_ackent *ae;
12932         uint32_t tiwin, us_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
12933         int cnt, i, did_out, ourfinisacked = 0;
12934         int win_up_req = 0;
12935         struct tcpopt to_holder, *to = NULL;
12936         int nsegs = 0;
12937         int under_pacing = 1;
12938         int recovery = 0;
12939         int idx;
12940 #ifdef TCP_ACCOUNTING
12941         sched_pin();
12942 #endif
12943         rack = (struct tcp_rack *)tp->t_fb_ptr;
12944         if (rack->gp_ready &&
12945             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
12946                 under_pacing = 0;
12947         else
12948                 under_pacing = 1;
12949
12950         if (rack->r_state != tp->t_state)
12951                 rack_set_state(tp, rack);
12952         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12953             (tp->t_flags & TF_GPUTINPROG)) {
12954                 /*
12955                  * We have a goodput in progress
12956                  * and we have entered a late state.
12957                  * Do we have enough data in the sb
12958                  * to handle the GPUT request?
12959                  */
12960                 uint32_t bytes;
12961
12962                 bytes = tp->gput_ack - tp->gput_seq;
12963                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
12964                         bytes += tp->gput_seq - tp->snd_una;
12965                 if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
12966                         /*
12967                          * There are not enough bytes in the socket
12968                          * buffer that have been sent to cover this
12969                          * measurement. Cancel it.
12970                          */
12971                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
12972                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
12973                                                    tp->gput_seq,
12974                                                    0, 0, 18, __LINE__, NULL, 0);
12975                         tp->t_flags &= ~TF_GPUTINPROG;
12976                 }
12977         }
12978         to = &to_holder;
12979         to->to_flags = 0;
12980         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
12981                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
12982         cnt = m->m_len / sizeof(struct tcp_ackent);
12983         idx = cnt / 5;
12984         if (idx >= MAX_NUM_OF_CNTS)
12985                 idx = MAX_NUM_OF_CNTS - 1;
12986         counter_u64_add(rack_proc_comp_ack[idx], 1);
12987         counter_u64_add(rack_multi_single_eq, cnt);
12988         high_seq = tp->snd_una;
12989         the_win = tp->snd_wnd;
12990         win_seq = tp->snd_wl1;
12991         win_upd_ack = tp->snd_wl2;
12992         cts = us_cts = tcp_tv_to_usectick(tv);
12993         segsiz = ctf_fixed_maxseg(tp);
12994         if ((rack->rc_gp_dyn_mul) &&
12995             (rack->use_fixed_rate == 0) &&
12996             (rack->rc_always_pace)) {
12997                 /* Check in on probertt */
12998                 rack_check_probe_rtt(rack, us_cts);
12999         }
13000         for (i = 0; i < cnt; i++) {
13001 #ifdef TCP_ACCOUNTING
13002                 ts_val = get_cyclecount();
13003 #endif
13004                 rack_clear_rate_sample(rack);
13005                 ae = ((mtod(m, struct tcp_ackent *)) + i);
13006                 /* Setup the window */
13007                 tiwin = ae->win << tp->snd_scale;
13008                 /* figure out the type of ack */
13009                 if (SEQ_LT(ae->ack, high_seq)) {
13010                         /* Case B*/
13011                         ae->ack_val_set = ACK_BEHIND;
13012                 } else if (SEQ_GT(ae->ack, high_seq)) {
13013                         /* Case A */
13014                         ae->ack_val_set = ACK_CUMACK;
13015                 } else if (tiwin == the_win) {
13016                         /* Case D */
13017                         ae->ack_val_set = ACK_DUPACK;
13018                 } else {
13019                         /* Case C */
13020                         ae->ack_val_set = ACK_RWND;
13021                 }
13022                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13023                 /* Validate timestamp */
13024                 if (ae->flags & HAS_TSTMP) {
13025                         /* Setup for a timestamp */
13026                         to->to_flags = TOF_TS;
13027                         ae->ts_echo -= tp->ts_offset;
13028                         to->to_tsecr = ae->ts_echo;
13029                         to->to_tsval = ae->ts_value;
13030                         /*
13031                          * If echoed timestamp is later than the current time, fall back to
13032                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13033                          * were used when this connection was established.
13034                          */
13035                         if (TSTMP_GT(ae->ts_echo, cts))
13036                                 ae->ts_echo = 0;
13037                         if (tp->ts_recent &&
13038                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13039                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13040 #ifdef TCP_ACCOUNTING
13041                                         rdstc = get_cyclecount();
13042                                         if (rdstc > ts_val) {
13043                                                 counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13044                                                                 (rdstc - ts_val));
13045                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13046                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13047                                                 }
13048                                         }
13049 #endif
13050                                         continue;
13051                                 }
13052                         }
13053                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13054                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13055                                 tp->ts_recent_age = tcp_ts_getticks();
13056                                 tp->ts_recent = ae->ts_value;
13057                         }
13058                 } else {
13059                         /* Setup for a no options */
13060                         to->to_flags = 0;
13061                 }
13062                 /* Update the rcv time and perform idle reduction possibly */
13063                 if  (tp->t_idle_reduce &&
13064                      (tp->snd_max == tp->snd_una) &&
13065                      ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13066                         counter_u64_add(rack_input_idle_reduces, 1);
13067                         rack_cc_after_idle(rack, tp);
13068                 }
13069                 tp->t_rcvtime = ticks;
13070                 /* Now what about ECN? */
13071                 if (tp->t_flags2 & TF2_ECN_PERMIT) {
13072                         if (ae->flags & TH_CWR) {
13073                                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13074                                 tp->t_flags |= TF_ACKNOW;
13075                         }
13076                         switch (ae->codepoint & IPTOS_ECN_MASK) {
13077                         case IPTOS_ECN_CE:
13078                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
13079                                 KMOD_TCPSTAT_INC(tcps_ecn_ce);
13080                                 break;
13081                         case IPTOS_ECN_ECT0:
13082                                 KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13083                                 break;
13084                         case IPTOS_ECN_ECT1:
13085                                 KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13086                                 break;
13087                         }
13088
13089                         /* Process a packet differently from RFC3168. */
13090                         cc_ecnpkt_handler_flags(tp, ae->flags, ae->codepoint);
13091                         /* Congestion experienced. */
13092                         if (ae->flags & TH_ECE) {
13093                                 rack_cong_signal(tp,  CC_ECN, ae->ack);
13094                         }
13095                 }
13096 #ifdef TCP_ACCOUNTING
13097                 /* Count for the specific type of ack in */
13098                 counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13099                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13100                         tp->tcp_cnt_counters[ae->ack_val_set]++;
13101                 }
13102 #endif
13103                 /*
13104                  * Note how we could move up these in the determination
13105                  * above, but we don't so that way the timestamp checks (and ECN)
13106                  * is done first before we do any processing on the ACK.
13107                  * The non-compressed path through the code has this
13108                  * weakness (noted by @jtl) that it actually does some
13109                  * processing before verifying the timestamp information.
13110                  * We don't take that path here which is why we set
13111                  * the ack_val_set first, do the timestamp and ecn
13112                  * processing, and then look at what we have setup.
13113                  */
13114                 if (ae->ack_val_set == ACK_BEHIND) {
13115                         /*
13116                          * Case B flag reordering, if window is not closed
13117                          * or it could be a keep-alive or persists
13118                          */
13119                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13120                                 counter_u64_add(rack_reorder_seen, 1);
13121                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13122                         }
13123                 } else if (ae->ack_val_set == ACK_DUPACK) {
13124                         /* Case D */
13125
13126                         rack_strike_dupack(rack);
13127                 } else if (ae->ack_val_set == ACK_RWND) {
13128                         /* Case C */
13129
13130                         win_up_req = 1;
13131                         win_upd_ack = ae->ack;
13132                         win_seq = ae->seq;
13133                         the_win = tiwin;
13134                 } else {
13135                         /* Case A */
13136
13137                         if (SEQ_GT(ae->ack, tp->snd_max)) {
13138                                 /*
13139                                  * We just send an ack since the incoming
13140                                  * ack is beyond the largest seq we sent.
13141                                  */
13142                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
13143                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13144                                         if (tp->t_flags && TF_ACKNOW)
13145                                                 rack->r_wanted_output = 1;
13146                                 }
13147                         } else {
13148                                 nsegs++;
13149                                 /* If the window changed setup to update */
13150                                 if (tiwin != tp->snd_wnd) {
13151                                         win_up_req = 1;
13152                                         win_upd_ack = ae->ack;
13153                                         win_seq = ae->seq;
13154                                         the_win = tiwin;
13155                                 }
13156 #ifdef TCP_ACCOUNTING
13157                                 /* Account for the acks */
13158                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13159                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13160                                 }
13161                                 counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13162                                                 (((ae->ack - high_seq) + segsiz - 1) / segsiz));
13163 #endif
13164                                 high_seq = ae->ack;
13165                                 /* Setup our act_rcv_time */
13166                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13167                                         ts.tv_sec = ae->timestamp / 1000000000;
13168                                         ts.tv_nsec = ae->timestamp % 1000000000;
13169                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13170                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13171                                 } else {
13172                                         rack->r_ctl.act_rcv_time = *tv;
13173                                 }
13174                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13175                         }
13176                 }
13177                 /* And lets be sure to commit the rtt measurements for this ack */
13178                 tcp_rack_xmit_timer_commit(rack, tp);
13179 #ifdef TCP_ACCOUNTING
13180                 rdstc = get_cyclecount();
13181                 if (rdstc > ts_val) {
13182                         counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13183                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13184                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13185                                 if (ae->ack_val_set == ACK_CUMACK)
13186                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13187                         }
13188                 }
13189 #endif
13190         }
13191 #ifdef TCP_ACCOUNTING
13192         ts_val = get_cyclecount();
13193 #endif
13194         acked_amount = acked = (high_seq - tp->snd_una);
13195         if (win_up_req) {
13196                 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13197         }
13198         if (acked) {
13199                 if (rack->sack_attack_disable == 0)
13200                         rack_do_decay(rack);
13201                 if (acked >= segsiz) {
13202                         /*
13203                          * You only get credit for
13204                          * MSS and greater (and you get extra
13205                          * credit for larger cum-ack moves).
13206                          */
13207                         int ac;
13208
13209                         ac = acked / segsiz;
13210                         rack->r_ctl.ack_count += ac;
13211                         counter_u64_add(rack_ack_total, ac);
13212                 }
13213                 if (rack->r_ctl.ack_count > 0xfff00000) {
13214                         /*
13215                          * reduce the number to keep us under
13216                          * a uint32_t.
13217                          */
13218                         rack->r_ctl.ack_count /= 2;
13219                         rack->r_ctl.sack_count /= 2;
13220                 }
13221                 if (tp->t_flags & TF_NEEDSYN) {
13222                         /*
13223                          * T/TCP: Connection was half-synchronized, and our SYN has
13224                          * been ACK'd (so connection is now fully synchronized).  Go
13225                          * to non-starred state, increment snd_una for ACK of SYN,
13226                          * and check if we can do window scaling.
13227                          */
13228                         tp->t_flags &= ~TF_NEEDSYN;
13229                         tp->snd_una++;
13230                         acked_amount = acked = (high_seq - tp->snd_una);
13231                 }
13232                 if (acked > sbavail(&so->so_snd))
13233                         acked_amount = sbavail(&so->so_snd);
13234 #ifdef NETFLIX_EXP_DETECTION
13235                 /*
13236                  * We only care on a cum-ack move if we are in a sack-disabled
13237                  * state. We have already added in to the ack_count, and we never
13238                  * would disable on a cum-ack move, so we only care to do the
13239                  * detection if it may "undo" it, i.e. we were in disabled already.
13240                  */
13241                 if (rack->sack_attack_disable)
13242                         rack_do_detection(tp, rack, acked_amount, segsiz);
13243 #endif
13244                 if (IN_FASTRECOVERY(tp->t_flags) &&
13245                     (rack->rack_no_prr == 0))
13246                         rack_update_prr(tp, rack, acked_amount, high_seq);
13247                 if (IN_RECOVERY(tp->t_flags)) {
13248                         if (SEQ_LT(high_seq, tp->snd_recover) &&
13249                             (SEQ_LT(high_seq, tp->snd_max))) {
13250                                 tcp_rack_partialack(tp);
13251                         } else {
13252                                 rack_post_recovery(tp, high_seq);
13253                                 recovery = 1;
13254                         }
13255                 }
13256                 /* Handle the rack-log-ack part (sendmap) */
13257                 if ((sbused(&so->so_snd) == 0) &&
13258                     (acked > acked_amount) &&
13259                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
13260                     (tp->t_flags & TF_SENTFIN)) {
13261                         /*
13262                          * We must be sure our fin
13263                          * was sent and acked (we can be
13264                          * in FIN_WAIT_1 without having
13265                          * sent the fin).
13266                          */
13267                         ourfinisacked = 1;
13268                         /*
13269                          * Lets make sure snd_una is updated
13270                          * since most likely acked_amount = 0 (it
13271                          * should be).
13272                          */
13273                         tp->snd_una = high_seq;
13274                 }
13275                 /* Did we make a RTO error? */
13276                 if ((tp->t_flags & TF_PREVVALID) &&
13277                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13278                         tp->t_flags &= ~TF_PREVVALID;
13279                         if (tp->t_rxtshift == 1 &&
13280                             (int)(ticks - tp->t_badrxtwin) < 0)
13281                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq);
13282                 }
13283                 /* Handle the data in the socket buffer */
13284                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13285                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13286                 if (acked_amount > 0) {
13287                         struct mbuf *mfree;
13288
13289                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13290                         SOCKBUF_LOCK(&so->so_snd);
13291                         mfree = sbcut_locked(&so->so_snd, acked);
13292                         tp->snd_una = high_seq;
13293                         /* Note we want to hold the sb lock through the sendmap adjust */
13294                         rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13295                         /* Wake up the socket if we have room to write more */
13296                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13297                         sowwakeup_locked(so);
13298                         m_freem(mfree);
13299                 }
13300                 /* update progress */
13301                 tp->t_acktime = ticks;
13302                 rack_log_progress_event(rack, tp, tp->t_acktime,
13303                                         PROGRESS_UPDATE, __LINE__);
13304                 /* Clear out shifts and such */
13305                 tp->t_rxtshift = 0;
13306                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13307                                    rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13308                 rack->rc_tlp_in_progress = 0;
13309                 rack->r_ctl.rc_tlp_cnt_out = 0;
13310                 /* Send recover and snd_nxt must be dragged along */
13311                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13312                         tp->snd_recover = tp->snd_una;
13313                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13314                         tp->snd_nxt = tp->snd_una;
13315                 /*
13316                  * If the RXT timer is running we want to
13317                  * stop it, so we can restart a TLP (or new RXT).
13318                  */
13319                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13320                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13321 #ifdef NETFLIX_HTTP_LOGGING
13322                 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13323 #endif
13324                 tp->snd_wl2 = high_seq;
13325                 tp->t_dupacks = 0;
13326                 if (under_pacing &&
13327                     (rack->use_fixed_rate == 0) &&
13328                     (rack->in_probe_rtt == 0) &&
13329                     rack->rc_gp_dyn_mul &&
13330                     rack->rc_always_pace) {
13331                         /* Check if we are dragging bottom */
13332                         rack_check_bottom_drag(tp, rack, so, acked);
13333                 }
13334                 if (tp->snd_una == tp->snd_max) {
13335                         tp->t_flags &= ~TF_PREVVALID;
13336                         rack->r_ctl.retran_during_recovery = 0;
13337                         rack->r_ctl.dsack_byte_cnt = 0;
13338                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13339                         if (rack->r_ctl.rc_went_idle_time == 0)
13340                                 rack->r_ctl.rc_went_idle_time = 1;
13341                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13342                         if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13343                                 tp->t_acktime = 0;
13344                         /* Set so we might enter persists... */
13345                         rack->r_wanted_output = 1;
13346                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13347                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13348                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13349                             (sbavail(&so->so_snd) == 0) &&
13350                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13351                                 /*
13352                                  * The socket was gone and the
13353                                  * peer sent data (not now in the past), time to
13354                                  * reset him.
13355                                  */
13356                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13357                                 /* tcp_close will kill the inp pre-log the Reset */
13358                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13359 #ifdef TCP_ACCOUNTING
13360                                 rdstc = get_cyclecount();
13361                                 if (rdstc > ts_val) {
13362                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13363                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13364                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13365                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13366                                         }
13367                                 }
13368 #endif
13369                                 m_freem(m);
13370                                 tp = tcp_close(tp);
13371                                 if (tp == NULL) {
13372 #ifdef TCP_ACCOUNTING
13373                                         sched_unpin();
13374 #endif
13375                                         return (1);
13376                                 }
13377                                 /*
13378                                  * We would normally do drop-with-reset which would
13379                                  * send back a reset. We can't since we don't have
13380                                  * all the needed bits. Instead lets arrange for
13381                                  * a call to tcp_output(). That way since we
13382                                  * are in the closed state we will generate a reset.
13383                                  *
13384                                  * Note if tcp_accounting is on we don't unpin since
13385                                  * we do that after the goto label.
13386                                  */
13387                                 goto send_out_a_rst;
13388                         }
13389                         if ((sbused(&so->so_snd) == 0) &&
13390                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
13391                             (tp->t_flags & TF_SENTFIN)) {
13392                                 /*
13393                                  * If we can't receive any more data, then closing user can
13394                                  * proceed. Starting the timer is contrary to the
13395                                  * specification, but if we don't get a FIN we'll hang
13396                                  * forever.
13397                                  *
13398                                  */
13399                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13400                                         soisdisconnected(so);
13401                                         tcp_timer_activate(tp, TT_2MSL,
13402                                                            (tcp_fast_finwait2_recycle ?
13403                                                             tcp_finwait2_timeout :
13404                                                             TP_MAXIDLE(tp)));
13405                                 }
13406                                 if (ourfinisacked == 0) {
13407                                         /*
13408                                          * We don't change to fin-wait-2 if we have our fin acked
13409                                          * which means we are probably in TCPS_CLOSING.
13410                                          */
13411                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13412                                 }
13413                         }
13414                 }
13415                 /* Wake up the socket if we have room to write more */
13416                 if (sbavail(&so->so_snd)) {
13417                         rack->r_wanted_output = 1;
13418                         if (ctf_progress_timeout_check(tp, true)) {
13419                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13420                                                         tp, tick, PROGRESS_DROP, __LINE__);
13421                                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
13422                                 /*
13423                                  * We cheat here and don't send a RST, we should send one
13424                                  * when the pacer drops the connection.
13425                                  */
13426 #ifdef TCP_ACCOUNTING
13427                                 rdstc = get_cyclecount();
13428                                 if (rdstc > ts_val) {
13429                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13430                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13431                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13432                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13433                                         }
13434                                 }
13435                                 sched_unpin();
13436 #endif
13437                                 INP_WUNLOCK(rack->rc_inp);
13438                                 m_freem(m);
13439                                 return (1);
13440                         }
13441                 }
13442                 if (ourfinisacked) {
13443                         switch(tp->t_state) {
13444                         case TCPS_CLOSING:
13445 #ifdef TCP_ACCOUNTING
13446                                 rdstc = get_cyclecount();
13447                                 if (rdstc > ts_val) {
13448                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13449                                                         (rdstc - ts_val));
13450                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13451                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13452                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13453                                         }
13454                                 }
13455                                 sched_unpin();
13456 #endif
13457                                 tcp_twstart(tp);
13458                                 m_freem(m);
13459                                 return (1);
13460                                 break;
13461                         case TCPS_LAST_ACK:
13462 #ifdef TCP_ACCOUNTING
13463                                 rdstc = get_cyclecount();
13464                                 if (rdstc > ts_val) {
13465                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13466                                                         (rdstc - ts_val));
13467                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13468                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13469                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13470                                         }
13471                                 }
13472                                 sched_unpin();
13473 #endif
13474                                 tp = tcp_close(tp);
13475                                 ctf_do_drop(m, tp);
13476                                 return (1);
13477                                 break;
13478                         case TCPS_FIN_WAIT_1:
13479 #ifdef TCP_ACCOUNTING
13480                                 rdstc = get_cyclecount();
13481                                 if (rdstc > ts_val) {
13482                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13483                                                         (rdstc - ts_val));
13484                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13485                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13486                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13487                                         }
13488                                 }
13489 #endif
13490                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13491                                         soisdisconnected(so);
13492                                         tcp_timer_activate(tp, TT_2MSL,
13493                                                            (tcp_fast_finwait2_recycle ?
13494                                                             tcp_finwait2_timeout :
13495                                                             TP_MAXIDLE(tp)));
13496                                 }
13497                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13498                                 break;
13499                         default:
13500                                 break;
13501                         }
13502                 }
13503                 if (rack->r_fast_output) {
13504                         /*
13505                          * We re doing fast output.. can we expand that?
13506                          */
13507                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13508                 }
13509 #ifdef TCP_ACCOUNTING
13510                 rdstc = get_cyclecount();
13511                 if (rdstc > ts_val) {
13512                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13513                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13514                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13515                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13516                         }
13517                 }
13518
13519         } else if (win_up_req) {
13520                 rdstc = get_cyclecount();
13521                 if (rdstc > ts_val) {
13522                         counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13523                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13524                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13525                         }
13526                 }
13527 #endif
13528         }
13529         /* Now is there a next packet, if so we are done */
13530         m_freem(m);
13531         did_out = 0;
13532         if (nxt_pkt) {
13533 #ifdef TCP_ACCOUNTING
13534                 sched_unpin();
13535 #endif
13536                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13537                 return (0);
13538         }
13539         rack_handle_might_revert(tp, rack);
13540         ctf_calc_rwin(so, tp);
13541         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13542         send_out_a_rst:
13543                 (void)tp->t_fb->tfb_tcp_output(tp);
13544                 did_out = 1;
13545         }
13546         rack_free_trim(rack);
13547 #ifdef TCP_ACCOUNTING
13548         sched_unpin();
13549 #endif
13550         rack_timer_audit(tp, rack, &so->so_snd);
13551         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13552         return (0);
13553 }
13554
13555
13556 static int
13557 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13558     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13559     int32_t nxt_pkt, struct timeval *tv)
13560 {
13561 #ifdef TCP_ACCOUNTING
13562         uint64_t ts_val;
13563 #endif
13564         int32_t thflags, retval, did_out = 0;
13565         int32_t way_out = 0;
13566         uint32_t cts;
13567         uint32_t tiwin;
13568         struct timespec ts;
13569         struct tcpopt to;
13570         struct tcp_rack *rack;
13571         struct rack_sendmap *rsm;
13572         int32_t prev_state = 0;
13573 #ifdef TCP_ACCOUNTING
13574         int ack_val_set = 0xf;
13575 #endif
13576         int nsegs;
13577         uint32_t us_cts;
13578         /*
13579          * tv passed from common code is from either M_TSTMP_LRO or
13580          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13581          */
13582         rack = (struct tcp_rack *)tp->t_fb_ptr;
13583         cts = tcp_tv_to_usectick(tv);
13584         if (m->m_flags & M_ACKCMP) {
13585                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13586         }
13587         if (m->m_flags & M_ACKCMP) {
13588                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13589         }
13590         nsegs = m->m_pkthdr.lro_nsegs;
13591         counter_u64_add(rack_proc_non_comp_ack, 1);
13592         thflags = th->th_flags;
13593 #ifdef TCP_ACCOUNTING
13594         sched_pin();
13595         if (thflags & TH_ACK)
13596                 ts_val = get_cyclecount();
13597 #endif
13598         if ((m->m_flags & M_TSTMP) ||
13599             (m->m_flags & M_TSTMP_LRO)) {
13600                 mbuf_tstmp2timespec(m, &ts);
13601                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13602                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13603         } else
13604                 rack->r_ctl.act_rcv_time = *tv;
13605         kern_prefetch(rack, &prev_state);
13606         prev_state = 0;
13607         /*
13608          * Unscale the window into a 32-bit value. For the SYN_SENT state
13609          * the scale is zero.
13610          */
13611         tiwin = th->th_win << tp->snd_scale;
13612         /*
13613          * Parse options on any incoming segment.
13614          */
13615         memset(&to, 0, sizeof(to));
13616         tcp_dooptions(&to, (u_char *)(th + 1),
13617             (th->th_off << 2) - sizeof(struct tcphdr),
13618             (thflags & TH_SYN) ? TO_SYN : 0);
13619 #ifdef TCP_ACCOUNTING
13620         if (thflags & TH_ACK) {
13621                 /*
13622                  * We have a tradeoff here. We can either do what we are
13623                  * doing i.e. pinning to this CPU and then doing the accounting
13624                  * <or> we could do a critical enter, setup the rdtsc and cpu
13625                  * as in below, and then validate we are on the same CPU on
13626                  * exit. I have choosen to not do the critical enter since
13627                  * that often will gain you a context switch, and instead lock
13628                  * us (line above this if) to the same CPU with sched_pin(). This
13629                  * means we may be context switched out for a higher priority
13630                  * interupt but we won't be moved to another CPU.
13631                  *
13632                  * If this occurs (which it won't very often since we most likely
13633                  * are running this code in interupt context and only a higher
13634                  * priority will bump us ... clock?) we will falsely add in
13635                  * to the time the interupt processing time plus the ack processing
13636                  * time. This is ok since its a rare event.
13637                  */
13638                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
13639                                                     ctf_fixed_maxseg(tp));
13640         }
13641 #endif
13642         NET_EPOCH_ASSERT();
13643         INP_WLOCK_ASSERT(tp->t_inpcb);
13644         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
13645             __func__));
13646         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
13647             __func__));
13648         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13649             (tp->t_flags & TF_GPUTINPROG)) {
13650                 /*
13651                  * We have a goodput in progress
13652                  * and we have entered a late state.
13653                  * Do we have enough data in the sb
13654                  * to handle the GPUT request?
13655                  */
13656                 uint32_t bytes;
13657
13658                 bytes = tp->gput_ack - tp->gput_seq;
13659                 if (SEQ_GT(tp->gput_seq, tp->snd_una))
13660                         bytes += tp->gput_seq - tp->snd_una;
13661                 if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
13662                         /*
13663                          * There are not enough bytes in the socket
13664                          * buffer that have been sent to cover this
13665                          * measurement. Cancel it.
13666                          */
13667                         rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13668                                                    rack->r_ctl.rc_gp_srtt /*flex1*/,
13669                                                    tp->gput_seq,
13670                                                    0, 0, 18, __LINE__, NULL, 0);
13671                         tp->t_flags &= ~TF_GPUTINPROG;
13672                 }
13673         }
13674         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13675                 union tcp_log_stackspecific log;
13676                 struct timeval ltv;
13677 #ifdef NETFLIX_HTTP_LOGGING
13678                 struct http_sendfile_track *http_req;
13679
13680                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
13681                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
13682                 } else {
13683                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
13684                 }
13685 #endif
13686                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13687                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13688                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13689                 if (rack->rack_no_prr == 0)
13690                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13691                 else
13692                         log.u_bbr.flex1 = 0;
13693                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13694                 log.u_bbr.use_lt_bw <<= 1;
13695                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13696                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13697                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13698                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
13699                 log.u_bbr.flex3 = m->m_flags;
13700                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13701                 log.u_bbr.lost = thflags;
13702                 log.u_bbr.pacing_gain = 0x1;
13703 #ifdef TCP_ACCOUNTING
13704                 log.u_bbr.cwnd_gain = ack_val_set;
13705 #endif
13706                 log.u_bbr.flex7 = 2;
13707                 if (m->m_flags & M_TSTMP) {
13708                         /* Record the hardware timestamp if present */
13709                         mbuf_tstmp2timespec(m, &ts);
13710                         ltv.tv_sec = ts.tv_sec;
13711                         ltv.tv_usec = ts.tv_nsec / 1000;
13712                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13713                 } else if (m->m_flags & M_TSTMP_LRO) {
13714                         /* Record the LRO the arrival timestamp */
13715                         mbuf_tstmp2timespec(m, &ts);
13716                         ltv.tv_sec = ts.tv_sec;
13717                         ltv.tv_usec = ts.tv_nsec / 1000;
13718                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13719                 }
13720                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13721                 /* Log the rcv time */
13722                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
13723 #ifdef NETFLIX_HTTP_LOGGING
13724                 log.u_bbr.applimited = tp->t_http_closed;
13725                 log.u_bbr.applimited <<= 8;
13726                 log.u_bbr.applimited |= tp->t_http_open;
13727                 log.u_bbr.applimited <<= 8;
13728                 log.u_bbr.applimited |= tp->t_http_req;
13729                 if (http_req) {
13730                         /* Copy out any client req info */
13731                         /* seconds */
13732                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13733                         /* useconds */
13734                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13735                         log.u_bbr.rttProp = http_req->timestamp;
13736                         log.u_bbr.cur_del_rate = http_req->start;
13737                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13738                                 log.u_bbr.flex8 |= 1;
13739                         } else {
13740                                 log.u_bbr.flex8 |= 2;
13741                                 log.u_bbr.bw_inuse = http_req->end;
13742                         }
13743                         log.u_bbr.flex6 = http_req->start_seq;
13744                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13745                                 log.u_bbr.flex8 |= 4;
13746                                 log.u_bbr.epoch = http_req->end_seq;
13747                         }
13748                 }
13749 #endif
13750                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
13751                     tlen, &log, true, &ltv);
13752         }
13753         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
13754                 way_out = 4;
13755                 retval = 0;
13756                 m_freem(m);
13757                 goto done_with_input;
13758         }
13759         /*
13760          * If a segment with the ACK-bit set arrives in the SYN-SENT state
13761          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
13762          */
13763         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
13764             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
13765                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13766                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13767 #ifdef TCP_ACCOUNTING
13768                 sched_unpin();
13769 #endif
13770                 return (1);
13771         }
13772
13773         /*
13774          * Parse options on any incoming segment.
13775          */
13776         tcp_dooptions(&to, (u_char *)(th + 1),
13777             (th->th_off << 2) - sizeof(struct tcphdr),
13778             (thflags & TH_SYN) ? TO_SYN : 0);
13779
13780         /*
13781          * If timestamps were negotiated during SYN/ACK and a
13782          * segment without a timestamp is received, silently drop
13783          * the segment, unless it is a RST segment or missing timestamps are
13784          * tolerated.
13785          * See section 3.2 of RFC 7323.
13786          */
13787         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
13788             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
13789                 way_out = 5;
13790                 retval = 0;
13791                 m_freem(m);
13792                 goto done_with_input;
13793         }
13794
13795         /*
13796          * Segment received on connection. Reset idle time and keep-alive
13797          * timer. XXX: This should be done after segment validation to
13798          * ignore broken/spoofed segs.
13799          */
13800         if  (tp->t_idle_reduce &&
13801              (tp->snd_max == tp->snd_una) &&
13802              ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13803                 counter_u64_add(rack_input_idle_reduces, 1);
13804                 rack_cc_after_idle(rack, tp);
13805         }
13806         tp->t_rcvtime = ticks;
13807 #ifdef STATS
13808         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
13809 #endif
13810         if (tiwin > rack->r_ctl.rc_high_rwnd)
13811                 rack->r_ctl.rc_high_rwnd = tiwin;
13812         /*
13813          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
13814          * this to occur after we've validated the segment.
13815          */
13816         if (tp->t_flags2 & TF2_ECN_PERMIT) {
13817                 if (thflags & TH_CWR) {
13818                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13819                         tp->t_flags |= TF_ACKNOW;
13820                 }
13821                 switch (iptos & IPTOS_ECN_MASK) {
13822                 case IPTOS_ECN_CE:
13823                         tp->t_flags2 |= TF2_ECN_SND_ECE;
13824                         KMOD_TCPSTAT_INC(tcps_ecn_ce);
13825                         break;
13826                 case IPTOS_ECN_ECT0:
13827                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13828                         break;
13829                 case IPTOS_ECN_ECT1:
13830                         KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13831                         break;
13832                 }
13833
13834                 /* Process a packet differently from RFC3168. */
13835                 cc_ecnpkt_handler(tp, th, iptos);
13836
13837                 /* Congestion experienced. */
13838                 if (thflags & TH_ECE) {
13839                         rack_cong_signal(tp, CC_ECN, th->th_ack);
13840                 }
13841         }
13842
13843         /*
13844          * If echoed timestamp is later than the current time, fall back to
13845          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13846          * were used when this connection was established.
13847          */
13848         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
13849                 to.to_tsecr -= tp->ts_offset;
13850                 if (TSTMP_GT(to.to_tsecr, cts))
13851                         to.to_tsecr = 0;
13852         }
13853
13854         /*
13855          * If its the first time in we need to take care of options and
13856          * verify we can do SACK for rack!
13857          */
13858         if (rack->r_state == 0) {
13859                 /* Should be init'd by rack_init() */
13860                 KASSERT(rack->rc_inp != NULL,
13861                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
13862                 if (rack->rc_inp == NULL) {
13863                         rack->rc_inp = tp->t_inpcb;
13864                 }
13865
13866                 /*
13867                  * Process options only when we get SYN/ACK back. The SYN
13868                  * case for incoming connections is handled in tcp_syncache.
13869                  * According to RFC1323 the window field in a SYN (i.e., a
13870                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
13871                  * this is traditional behavior, may need to be cleaned up.
13872                  */
13873                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
13874                         /* Handle parallel SYN for ECN */
13875                         if (!(thflags & TH_ACK) &&
13876                             ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
13877                             ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
13878                                 tp->t_flags2 |= TF2_ECN_PERMIT;
13879                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
13880                                 TCPSTAT_INC(tcps_ecn_shs);
13881                         }
13882                         if ((to.to_flags & TOF_SCALE) &&
13883                             (tp->t_flags & TF_REQ_SCALE)) {
13884                                 tp->t_flags |= TF_RCVD_SCALE;
13885                                 tp->snd_scale = to.to_wscale;
13886                         } else
13887                                 tp->t_flags &= ~TF_REQ_SCALE;
13888                         /*
13889                          * Initial send window.  It will be updated with the
13890                          * next incoming segment to the scaled value.
13891                          */
13892                         tp->snd_wnd = th->th_win;
13893                         rack_validate_fo_sendwin_up(tp, rack);
13894                         if ((to.to_flags & TOF_TS) &&
13895                             (tp->t_flags & TF_REQ_TSTMP)) {
13896                                 tp->t_flags |= TF_RCVD_TSTMP;
13897                                 tp->ts_recent = to.to_tsval;
13898                                 tp->ts_recent_age = cts;
13899                         } else
13900                                 tp->t_flags &= ~TF_REQ_TSTMP;
13901                         if (to.to_flags & TOF_MSS) {
13902                                 tcp_mss(tp, to.to_mss);
13903                         }
13904                         if ((tp->t_flags & TF_SACK_PERMIT) &&
13905                             (to.to_flags & TOF_SACKPERM) == 0)
13906                                 tp->t_flags &= ~TF_SACK_PERMIT;
13907                         if (IS_FASTOPEN(tp->t_flags)) {
13908                                 if (to.to_flags & TOF_FASTOPEN) {
13909                                         uint16_t mss;
13910
13911                                         if (to.to_flags & TOF_MSS)
13912                                                 mss = to.to_mss;
13913                                         else
13914                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
13915                                                         mss = TCP6_MSS;
13916                                                 else
13917                                                         mss = TCP_MSS;
13918                                         tcp_fastopen_update_cache(tp, mss,
13919                                             to.to_tfo_len, to.to_tfo_cookie);
13920                                 } else
13921                                         tcp_fastopen_disable_path(tp);
13922                         }
13923                 }
13924                 /*
13925                  * At this point we are at the initial call. Here we decide
13926                  * if we are doing RACK or not. We do this by seeing if
13927                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
13928                  * The code now does do dup-ack counting so if you don't
13929                  * switch back you won't get rack & TLP, but you will still
13930                  * get this stack.
13931                  */
13932
13933                 if ((rack_sack_not_required == 0) &&
13934                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
13935                         tcp_switch_back_to_default(tp);
13936                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
13937                             tlen, iptos);
13938 #ifdef TCP_ACCOUNTING
13939                         sched_unpin();
13940 #endif
13941                         return (1);
13942                 }
13943                 tcp_set_hpts(tp->t_inpcb);
13944                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
13945         }
13946         if (thflags & TH_FIN)
13947                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
13948         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13949         if ((rack->rc_gp_dyn_mul) &&
13950             (rack->use_fixed_rate == 0) &&
13951             (rack->rc_always_pace)) {
13952                 /* Check in on probertt */
13953                 rack_check_probe_rtt(rack, us_cts);
13954         }
13955         rack_clear_rate_sample(rack);
13956         if (rack->forced_ack) {
13957                 uint32_t us_rtt;
13958
13959                 /*
13960                  * A persist or keep-alive was forced out, update our
13961                  * min rtt time. Note we do not worry about lost
13962                  * retransmissions since KEEP-ALIVES and persists
13963                  * are usually way long on times of sending (though
13964                  * if we were really paranoid or worried we could
13965                  * at least use timestamps if available to validate).
13966                  */
13967                 rack->forced_ack = 0;
13968                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13969                 if (us_rtt == 0)
13970                         us_rtt = 1;
13971                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13972                 tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13973         }
13974         /*
13975          * This is the one exception case where we set the rack state
13976          * always. All other times (timers etc) we must have a rack-state
13977          * set (so we assure we have done the checks above for SACK).
13978          */
13979         rack->r_ctl.rc_rcvtime = cts;
13980         if (rack->r_state != tp->t_state)
13981                 rack_set_state(tp, rack);
13982         if (SEQ_GT(th->th_ack, tp->snd_una) &&
13983             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
13984                 kern_prefetch(rsm, &prev_state);
13985         prev_state = rack->r_state;
13986         retval = (*rack->r_substate) (m, th, so,
13987             tp, &to, drop_hdrlen,
13988             tlen, tiwin, thflags, nxt_pkt, iptos);
13989 #ifdef INVARIANTS
13990         if ((retval == 0) &&
13991             (tp->t_inpcb == NULL)) {
13992                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
13993                     retval, tp, prev_state);
13994         }
13995 #endif
13996         if (retval == 0) {
13997                 /*
13998                  * If retval is 1 the tcb is unlocked and most likely the tp
13999                  * is gone.
14000                  */
14001                 INP_WLOCK_ASSERT(tp->t_inpcb);
14002                 if ((rack->rc_gp_dyn_mul) &&
14003                     (rack->rc_always_pace) &&
14004                     (rack->use_fixed_rate == 0) &&
14005                     rack->in_probe_rtt &&
14006                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
14007                         /*
14008                          * If we are going for target, lets recheck before
14009                          * we output.
14010                          */
14011                         rack_check_probe_rtt(rack, us_cts);
14012                 }
14013                 if (rack->set_pacing_done_a_iw == 0) {
14014                         /* How much has been acked? */
14015                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14016                                 /* We have enough to set in the pacing segment size */
14017                                 rack->set_pacing_done_a_iw = 1;
14018                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
14019                         }
14020                 }
14021                 tcp_rack_xmit_timer_commit(rack, tp);
14022 #ifdef TCP_ACCOUNTING
14023                 /*
14024                  * If we set the ack_val_se to what ack processing we are doing
14025                  * we also want to track how many cycles we burned. Note
14026                  * the bits after tcp_output we let be "free". This is because
14027                  * we are also tracking the tcp_output times as well. Note the
14028                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
14029                  * 0xf cannot be returned and is what we initialize it too to
14030                  * indicate we are not doing the tabulations.
14031                  */
14032                 if (ack_val_set != 0xf) {
14033                         uint64_t crtsc;
14034
14035                         crtsc = get_cyclecount();
14036                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14037                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14038                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14039                         }
14040                 }
14041 #endif
14042                 if (nxt_pkt == 0) {
14043                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14044 do_output_now:
14045                                 did_out = 1;
14046                                 (void)tp->t_fb->tfb_tcp_output(tp);
14047                         }
14048                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14049                         rack_free_trim(rack);
14050                 }
14051                 if ((nxt_pkt == 0) &&
14052                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14053                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
14054                      (tp->t_flags & TF_DELACK) ||
14055                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14056                       (tp->t_state <= TCPS_CLOSING)))) {
14057                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
14058                         if ((tp->snd_max == tp->snd_una) &&
14059                             ((tp->t_flags & TF_DELACK) == 0) &&
14060                             (rack->rc_inp->inp_in_hpts) &&
14061                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14062                                 /* keep alive not needed if we are hptsi output yet */
14063                                 ;
14064                         } else {
14065                                 int late = 0;
14066                                 if (rack->rc_inp->inp_in_hpts) {
14067                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14068                                                 us_cts = tcp_get_usecs(NULL);
14069                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14070                                                         rack->r_early = 1;
14071                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14072                                                 } else
14073                                                         late = 1;
14074                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14075                                         }
14076                                         tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
14077                                 }
14078                                 if (late && (did_out == 0)) {
14079                                         /*
14080                                          * We are late in the sending
14081                                          * and we did not call the output
14082                                          * (this probably should not happen).
14083                                          */
14084                                         goto do_output_now;
14085                                 }
14086                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14087                         }
14088                         way_out = 1;
14089                 } else if (nxt_pkt == 0) {
14090                         /* Do we have the correct timer running? */
14091                         rack_timer_audit(tp, rack, &so->so_snd);
14092                         way_out = 2;
14093                 }
14094         done_with_input:
14095                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14096                 if (did_out)
14097                         rack->r_wanted_output = 0;
14098 #ifdef INVARIANTS
14099                 if (tp->t_inpcb == NULL) {
14100                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
14101                               did_out,
14102                               retval, tp, prev_state);
14103                 }
14104 #endif
14105 #ifdef TCP_ACCOUNTING
14106         } else {
14107                 /*
14108                  * Track the time (see above).
14109                  */
14110                 if (ack_val_set != 0xf) {
14111                         uint64_t crtsc;
14112
14113                         crtsc = get_cyclecount();
14114                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14115                         /*
14116                          * Note we *DO NOT* increment the per-tcb counters since
14117                          * in the else the TP may be gone!!
14118                          */
14119                 }
14120 #endif
14121         }
14122 #ifdef TCP_ACCOUNTING
14123         sched_unpin();
14124 #endif
14125         return (retval);
14126 }
14127
14128 void
14129 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14130     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14131 {
14132         struct timeval tv;
14133
14134         /* First lets see if we have old packets */
14135         if (tp->t_in_pkt) {
14136                 if (ctf_do_queued_segments(so, tp, 1)) {
14137                         m_freem(m);
14138                         return;
14139                 }
14140         }
14141         if (m->m_flags & M_TSTMP_LRO) {
14142                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
14143                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
14144         } else {
14145                 /* Should not be should we kassert instead? */
14146                 tcp_get_usecs(&tv);
14147         }
14148         if (rack_do_segment_nounlock(m, th, so, tp,
14149                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14150                 INP_WUNLOCK(tp->t_inpcb);
14151         }
14152 }
14153
14154 struct rack_sendmap *
14155 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14156 {
14157         struct rack_sendmap *rsm = NULL;
14158         int32_t idx;
14159         uint32_t srtt = 0, thresh = 0, ts_low = 0;
14160
14161         /* Return the next guy to be re-transmitted */
14162         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14163                 return (NULL);
14164         }
14165         if (tp->t_flags & TF_SENTFIN) {
14166                 /* retran the end FIN? */
14167                 return (NULL);
14168         }
14169         /* ok lets look at this one */
14170         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14171         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14172                 goto check_it;
14173         }
14174         rsm = rack_find_lowest_rsm(rack);
14175         if (rsm == NULL) {
14176                 return (NULL);
14177         }
14178 check_it:
14179         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14180             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14181                 /*
14182                  * No sack so we automatically do the 3 strikes and
14183                  * retransmit (no rack timer would be started).
14184                  */
14185
14186                 return (rsm);
14187         }
14188         if (rsm->r_flags & RACK_ACKED) {
14189                 return (NULL);
14190         }
14191         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14192             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14193                 /* Its not yet ready */
14194                 return (NULL);
14195         }
14196         srtt = rack_grab_rtt(tp, rack);
14197         idx = rsm->r_rtr_cnt - 1;
14198         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14199         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14200         if ((tsused == ts_low) ||
14201             (TSTMP_LT(tsused, ts_low))) {
14202                 /* No time since sending */
14203                 return (NULL);
14204         }
14205         if ((tsused - ts_low) < thresh) {
14206                 /* It has not been long enough yet */
14207                 return (NULL);
14208         }
14209         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14210             ((rsm->r_flags & RACK_SACK_PASSED) &&
14211              (rack->sack_attack_disable == 0))) {
14212                 /*
14213                  * We have passed the dup-ack threshold <or>
14214                  * a SACK has indicated this is missing.
14215                  * Note that if you are a declared attacker
14216                  * it is only the dup-ack threshold that
14217                  * will cause retransmits.
14218                  */
14219                 /* log retransmit reason */
14220                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14221                 rack->r_fast_output = 0;
14222                 return (rsm);
14223         }
14224         return (NULL);
14225 }
14226
14227 static void
14228 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14229                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14230                            int line, struct rack_sendmap *rsm, uint8_t quality)
14231 {
14232         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14233                 union tcp_log_stackspecific log;
14234                 struct timeval tv;
14235
14236                 memset(&log, 0, sizeof(log));
14237                 log.u_bbr.flex1 = slot;
14238                 log.u_bbr.flex2 = len;
14239                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14240                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14241                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14242                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14243                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14244                 log.u_bbr.use_lt_bw <<= 1;
14245                 log.u_bbr.use_lt_bw |= rack->r_late;
14246                 log.u_bbr.use_lt_bw <<= 1;
14247                 log.u_bbr.use_lt_bw |= rack->r_early;
14248                 log.u_bbr.use_lt_bw <<= 1;
14249                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14250                 log.u_bbr.use_lt_bw <<= 1;
14251                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14252                 log.u_bbr.use_lt_bw <<= 1;
14253                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14254                 log.u_bbr.use_lt_bw <<= 1;
14255                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14256                 log.u_bbr.use_lt_bw <<= 1;
14257                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14258                 log.u_bbr.pkt_epoch = line;
14259                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14260                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14261                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14262                 log.u_bbr.bw_inuse = bw_est;
14263                 log.u_bbr.delRate = bw;
14264                 if (rack->r_ctl.gp_bw == 0)
14265                         log.u_bbr.cur_del_rate = 0;
14266                 else
14267                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14268                 log.u_bbr.rttProp = len_time;
14269                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14270                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14271                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14272                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14273                         /* We are in slow start */
14274                         log.u_bbr.flex7 = 1;
14275                 } else {
14276                         /* we are on congestion avoidance */
14277                         log.u_bbr.flex7 = 0;
14278                 }
14279                 log.u_bbr.flex8 = method;
14280                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14281                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14282                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14283                 log.u_bbr.cwnd_gain <<= 1;
14284                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14285                 log.u_bbr.cwnd_gain <<= 1;
14286                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14287                 log.u_bbr.bbr_substate = quality;
14288                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14289                     &rack->rc_inp->inp_socket->so_rcv,
14290                     &rack->rc_inp->inp_socket->so_snd,
14291                     BBR_LOG_HPTSI_CALC, 0,
14292                     0, &log, false, &tv);
14293         }
14294 }
14295
14296 static uint32_t
14297 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14298 {
14299         uint32_t new_tso, user_max;
14300
14301         user_max = rack->rc_user_set_max_segs * mss;
14302         if (rack->rc_force_max_seg) {
14303                 return (user_max);
14304         }
14305         if (rack->use_fixed_rate &&
14306             ((rack->r_ctl.crte == NULL) ||
14307              (bw != rack->r_ctl.crte->rate))) {
14308                 /* Use the user mss since we are not exactly matched */
14309                 return (user_max);
14310         }
14311         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14312         if (new_tso > user_max)
14313                 new_tso = user_max;
14314         return (new_tso);
14315 }
14316
14317 static int32_t
14318 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)
14319 {
14320         uint64_t lentim, fill_bw;
14321
14322         /* Lets first see if we are full, if so continue with normal rate */
14323         rack->r_via_fill_cw = 0;
14324         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14325                 return (slot);
14326         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14327                 return (slot);
14328         if (rack->r_ctl.rc_last_us_rtt == 0)
14329                 return (slot);
14330         if (rack->rc_pace_fill_if_rttin_range &&
14331             (rack->r_ctl.rc_last_us_rtt >=
14332              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14333                 /* The rtt is huge, N * smallest, lets not fill */
14334                 return (slot);
14335         }
14336         /*
14337          * first lets calculate the b/w based on the last us-rtt
14338          * and the sndwnd.
14339          */
14340         fill_bw = rack->r_ctl.cwnd_to_use;
14341         /* Take the rwnd if its smaller */
14342         if (fill_bw > rack->rc_tp->snd_wnd)
14343                 fill_bw = rack->rc_tp->snd_wnd;
14344         if (rack->r_fill_less_agg) {
14345                 /*
14346                  * Now take away the inflight (this will reduce our
14347                  * aggressiveness and yeah, if we get that much out in 1RTT
14348                  * we will have had acks come back and still be behind).
14349                  */
14350                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14351         }
14352         /* Now lets make it into a b/w */
14353         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14354         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14355         /* We are below the min b/w */
14356         if (non_paced)
14357                 *rate_wanted = fill_bw;
14358         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14359                 return (slot);
14360         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14361                 fill_bw = rack->r_ctl.bw_rate_cap;
14362         rack->r_via_fill_cw = 1;
14363         if (rack->r_rack_hw_rate_caps &&
14364             (rack->r_ctl.crte != NULL)) {
14365                 uint64_t high_rate;
14366
14367                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14368                 if (fill_bw > high_rate) {
14369                         /* We are capping bw at the highest rate table entry */
14370                         if (*rate_wanted > high_rate) {
14371                                 /* The original rate was also capped */
14372                                 rack->r_via_fill_cw = 0;
14373                         }
14374                         rack_log_hdwr_pacing(rack,
14375                                              fill_bw, high_rate, __LINE__,
14376                                              0, 3);
14377                         fill_bw = high_rate;
14378                         if (capped)
14379                                 *capped = 1;
14380                 }
14381         } else if ((rack->r_ctl.crte == NULL) &&
14382                    (rack->rack_hdrw_pacing == 0) &&
14383                    (rack->rack_hdw_pace_ena) &&
14384                    rack->r_rack_hw_rate_caps &&
14385                    (rack->rack_attempt_hdwr_pace == 0) &&
14386                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14387                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14388                 /*
14389                  * Ok we may have a first attempt that is greater than our top rate
14390                  * lets check.
14391                  */
14392                 uint64_t high_rate;
14393
14394                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14395                 if (high_rate) {
14396                         if (fill_bw > high_rate) {
14397                                 fill_bw = high_rate;
14398                                 if (capped)
14399                                         *capped = 1;
14400                         }
14401                 }
14402         }
14403         /*
14404          * Ok fill_bw holds our mythical b/w to fill the cwnd
14405          * in a rtt, what does that time wise equate too?
14406          */
14407         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14408         lentim /= fill_bw;
14409         *rate_wanted = fill_bw;
14410         if (non_paced || (lentim < slot)) {
14411                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14412                                            0, lentim, 12, __LINE__, NULL, 0);
14413                 return ((int32_t)lentim);
14414         } else
14415                 return (slot);
14416 }
14417
14418 static int32_t
14419 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14420 {
14421         int32_t slot = 0;
14422         int can_start_hw_pacing = 1;
14423         int err;
14424
14425         if (rack->rc_always_pace == 0) {
14426                 /*
14427                  * We use the most optimistic possible cwnd/srtt for
14428                  * sending calculations. This will make our
14429                  * calculation anticipate getting more through
14430                  * quicker then possible. But thats ok we don't want
14431                  * the peer to have a gap in data sending.
14432                  */
14433                 uint32_t srtt, cwnd, tr_perms = 0;
14434                 int32_t reduce = 0;
14435
14436         old_method:
14437                 /*
14438                  * We keep no precise pacing with the old method
14439                  * instead we use the pacer to mitigate bursts.
14440                  */
14441                 if (rack->r_ctl.rc_rack_min_rtt)
14442                         srtt = rack->r_ctl.rc_rack_min_rtt;
14443                 else
14444                         srtt = max(tp->t_srtt, 1);
14445                 if (rack->r_ctl.rc_rack_largest_cwnd)
14446                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14447                 else
14448                         cwnd = rack->r_ctl.cwnd_to_use;
14449                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14450                 tr_perms = (cwnd * 1000) / srtt;
14451                 if (tr_perms == 0) {
14452                         tr_perms = ctf_fixed_maxseg(tp);
14453                 }
14454                 /*
14455                  * Calculate how long this will take to drain, if
14456                  * the calculation comes out to zero, thats ok we
14457                  * will use send_a_lot to possibly spin around for
14458                  * more increasing tot_len_this_send to the point
14459                  * that its going to require a pace, or we hit the
14460                  * cwnd. Which in that case we are just waiting for
14461                  * a ACK.
14462                  */
14463                 slot = len / tr_perms;
14464                 /* Now do we reduce the time so we don't run dry? */
14465                 if (slot && rack_slot_reduction) {
14466                         reduce = (slot / rack_slot_reduction);
14467                         if (reduce < slot) {
14468                                 slot -= reduce;
14469                         } else
14470                                 slot = 0;
14471                 }
14472                 slot *= HPTS_USEC_IN_MSEC;
14473                 if (rack->rc_pace_to_cwnd) {
14474                         uint64_t rate_wanted = 0;
14475
14476                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14477                         rack->rc_ack_can_sendout_data = 1;
14478                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14479                 } else
14480                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14481         } else {
14482                 uint64_t bw_est, res, lentim, rate_wanted;
14483                 uint32_t orig_val, srtt, segs, oh;
14484                 int capped = 0;
14485                 int prev_fill;
14486
14487                 if ((rack->r_rr_config == 1) && rsm) {
14488                         return (rack->r_ctl.rc_min_to);
14489                 }
14490                 if (rack->use_fixed_rate) {
14491                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14492                 } else if ((rack->r_ctl.init_rate == 0) &&
14493 #ifdef NETFLIX_PEAKRATE
14494                            (rack->rc_tp->t_maxpeakrate == 0) &&
14495 #endif
14496                            (rack->r_ctl.gp_bw == 0)) {
14497                         /* no way to yet do an estimate */
14498                         bw_est = rate_wanted = 0;
14499                 } else {
14500                         bw_est = rack_get_bw(rack);
14501                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14502                 }
14503                 if ((bw_est == 0) || (rate_wanted == 0) ||
14504                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14505                         /*
14506                          * No way yet to make a b/w estimate or
14507                          * our raise is set incorrectly.
14508                          */
14509                         goto old_method;
14510                 }
14511                 /* We need to account for all the overheads */
14512                 segs = (len + segsiz - 1) / segsiz;
14513                 /*
14514                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
14515                  * and how much data we put in each packet. Yes this
14516                  * means we may be off if we are larger than 1500 bytes
14517                  * or smaller. But this just makes us more conservative.
14518                  */
14519                 if (rack_hw_rate_min &&
14520                     (bw_est < rack_hw_rate_min))
14521                         can_start_hw_pacing = 0;
14522                 if (ETHERNET_SEGMENT_SIZE > segsiz)
14523                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
14524                 else
14525                         oh = 0;
14526                 segs *= oh;
14527                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14528                 res = lentim / rate_wanted;
14529                 slot = (uint32_t)res;
14530                 orig_val = rack->r_ctl.rc_pace_max_segs;
14531                 if (rack->r_ctl.crte == NULL) {
14532                         /*
14533                          * Only do this if we are not hardware pacing
14534                          * since if we are doing hw-pacing below we will
14535                          * set make a call after setting up or changing
14536                          * the rate.
14537                          */
14538                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14539                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14540                         /*
14541                          * We lost our rate somehow, this can happen
14542                          * if the interface changed underneath us.
14543                          */
14544                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14545                         rack->r_ctl.crte = NULL;
14546                         /* Lets re-allow attempting to setup pacing */
14547                         rack->rack_hdrw_pacing = 0;
14548                         rack->rack_attempt_hdwr_pace = 0;
14549                         rack_log_hdwr_pacing(rack,
14550                                              rate_wanted, bw_est, __LINE__,
14551                                              0, 6);
14552                 }
14553                 /* Did we change the TSO size, if so log it */
14554                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14555                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
14556                 prev_fill = rack->r_via_fill_cw;
14557                 if ((rack->rc_pace_to_cwnd) &&
14558                     (capped == 0) &&
14559                     (rack->use_fixed_rate == 0) &&
14560                     (rack->in_probe_rtt == 0) &&
14561                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14562                         /*
14563                          * We want to pace at our rate *or* faster to
14564                          * fill the cwnd to the max if its not full.
14565                          */
14566                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14567                 }
14568                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14569                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14570                         if ((rack->rack_hdw_pace_ena) &&
14571                             (can_start_hw_pacing > 0) &&
14572                             (rack->rack_hdrw_pacing == 0) &&
14573                             (rack->rack_attempt_hdwr_pace == 0)) {
14574                                 /*
14575                                  * Lets attempt to turn on hardware pacing
14576                                  * if we can.
14577                                  */
14578                                 rack->rack_attempt_hdwr_pace = 1;
14579                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14580                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
14581                                                                        rate_wanted,
14582                                                                        RS_PACING_GEQ,
14583                                                                        &err, &rack->r_ctl.crte_prev_rate);
14584                                 if (rack->r_ctl.crte) {
14585                                         rack->rack_hdrw_pacing = 1;
14586                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14587                                                                                                  0, rack->r_ctl.crte,
14588                                                                                                  NULL);
14589                                         rack_log_hdwr_pacing(rack,
14590                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14591                                                              err, 0);
14592                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14593                                 } else {
14594                                         counter_u64_add(rack_hw_pace_init_fail, 1);
14595                                 }
14596                         } else if (rack->rack_hdrw_pacing &&
14597                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14598                                 /* Do we need to adjust our rate? */
14599                                 const struct tcp_hwrate_limit_table *nrte;
14600
14601                                 if (rack->r_up_only &&
14602                                     (rate_wanted < rack->r_ctl.crte->rate)) {
14603                                         /**
14604                                          * We have four possible states here
14605                                          * having to do with the previous time
14606                                          * and this time.
14607                                          *   previous  |  this-time
14608                                          * A)     0      |     0   -- fill_cw not in the picture
14609                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
14610                                          * C)     1      |     1   -- all rates from fill_cw
14611                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
14612                                          *
14613                                          * For case A, C and D we don't allow a drop. But for
14614                                          * case B where we now our on our steady rate we do
14615                                          * allow a drop.
14616                                          *
14617                                          */
14618                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14619                                                 goto done_w_hdwr;
14620                                 }
14621                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
14622                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14623                                         if (rack_hw_rate_to_low &&
14624                                             (bw_est < rack_hw_rate_to_low)) {
14625                                                 /*
14626                                                  * The pacing rate is too low for hardware, but
14627                                                  * do allow hardware pacing to be restarted.
14628                                                  */
14629                                                 rack_log_hdwr_pacing(rack,
14630                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
14631                                                              0, 5);
14632                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14633                                                 rack->r_ctl.crte = NULL;
14634                                                 rack->rack_attempt_hdwr_pace = 0;
14635                                                 rack->rack_hdrw_pacing = 0;
14636                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14637                                                 goto done_w_hdwr;
14638                                         }
14639                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
14640                                                                    rack->rc_tp,
14641                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
14642                                                                    rate_wanted,
14643                                                                    RS_PACING_GEQ,
14644                                                                    &err, &rack->r_ctl.crte_prev_rate);
14645                                         if (nrte == NULL) {
14646                                                 /* Lost the rate */
14647                                                 rack->rack_hdrw_pacing = 0;
14648                                                 rack->r_ctl.crte = NULL;
14649                                                 rack_log_hdwr_pacing(rack,
14650                                                                      rate_wanted, 0, __LINE__,
14651                                                                      err, 1);
14652                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14653                                                 counter_u64_add(rack_hw_pace_lost, 1);
14654                                         } else if (nrte != rack->r_ctl.crte) {
14655                                                 rack->r_ctl.crte = nrte;
14656                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
14657                                                                                                          segsiz, 0,
14658                                                                                                          rack->r_ctl.crte,
14659                                                                                                          NULL);
14660                                                 rack_log_hdwr_pacing(rack,
14661                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14662                                                                      err, 2);
14663                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
14664                                         }
14665                                 } else {
14666                                         /* We just need to adjust the segment size */
14667                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14668                                         rack_log_hdwr_pacing(rack,
14669                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14670                                                              0, 4);
14671                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14672                                 }
14673                         }
14674                 }
14675                 if ((rack->r_ctl.crte != NULL) &&
14676                     (rack->r_ctl.crte->rate == rate_wanted)) {
14677                         /*
14678                          * We need to add a extra if the rates
14679                          * are exactly matched. The idea is
14680                          * we want the software to make sure the
14681                          * queue is empty before adding more, this
14682                          * gives us N MSS extra pace times where
14683                          * N is our sysctl
14684                          */
14685                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
14686                 }
14687 done_w_hdwr:
14688                 if (rack_limit_time_with_srtt &&
14689                     (rack->use_fixed_rate == 0) &&
14690 #ifdef NETFLIX_PEAKRATE
14691                     (rack->rc_tp->t_maxpeakrate == 0) &&
14692 #endif
14693                     (rack->rack_hdrw_pacing == 0)) {
14694                         /*
14695                          * Sanity check, we do not allow the pacing delay
14696                          * to be longer than the SRTT of the path. If it is
14697                          * a slow path, then adding a packet should increase
14698                          * the RTT and compensate for this i.e. the srtt will
14699                          * be greater so the allowed pacing time will be greater.
14700                          *
14701                          * Note this restriction is not for where a peak rate
14702                          * is set, we are doing fixed pacing or hardware pacing.
14703                          */
14704                         if (rack->rc_tp->t_srtt)
14705                                 srtt = rack->rc_tp->t_srtt;
14706                         else
14707                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
14708                         if (srtt < slot) {
14709                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
14710                                 slot = srtt;
14711                         }
14712                 }
14713                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
14714         }
14715         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
14716                 /*
14717                  * If this rate is seeing enobufs when it
14718                  * goes to send then either the nic is out
14719                  * of gas or we are mis-estimating the time
14720                  * somehow and not letting the queue empty
14721                  * completely. Lets add to the pacing time.
14722                  */
14723                 int hw_boost_delay;
14724
14725                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
14726                 if (hw_boost_delay > rack_enobuf_hw_max)
14727                         hw_boost_delay = rack_enobuf_hw_max;
14728                 else if (hw_boost_delay < rack_enobuf_hw_min)
14729                         hw_boost_delay = rack_enobuf_hw_min;
14730                 slot += hw_boost_delay;
14731         }
14732         if (slot)
14733                 counter_u64_add(rack_calc_nonzero, 1);
14734         else
14735                 counter_u64_add(rack_calc_zero, 1);
14736         return (slot);
14737 }
14738
14739 static void
14740 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
14741     tcp_seq startseq, uint32_t sb_offset)
14742 {
14743         struct rack_sendmap *my_rsm = NULL;
14744         struct rack_sendmap fe;
14745
14746         if (tp->t_state < TCPS_ESTABLISHED) {
14747                 /*
14748                  * We don't start any measurements if we are
14749                  * not at least established.
14750                  */
14751                 return;
14752         }
14753         if (tp->t_state >= TCPS_FIN_WAIT_1) {
14754                 /*
14755                  * We will get no more data into the SB
14756                  * this means we need to have the data available
14757                  * before we start a measurement.
14758                  */
14759
14760                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) <
14761                     max(rc_init_window(rack),
14762                         (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
14763                         /* Nope not enough data */
14764                         return;
14765                 }
14766         }
14767         tp->t_flags |= TF_GPUTINPROG;
14768         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
14769         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
14770         tp->gput_seq = startseq;
14771         rack->app_limited_needs_set = 0;
14772         if (rack->in_probe_rtt)
14773                 rack->measure_saw_probe_rtt = 1;
14774         else if ((rack->measure_saw_probe_rtt) &&
14775                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
14776                 rack->measure_saw_probe_rtt = 0;
14777         if (rack->rc_gp_filled)
14778                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14779         else {
14780                 /* Special case initial measurement */
14781                 struct timeval tv;
14782
14783                 tp->gput_ts = tcp_get_usecs(&tv);
14784                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14785         }
14786         /*
14787          * We take a guess out into the future,
14788          * if we have no measurement and no
14789          * initial rate, we measure the first
14790          * initial-windows worth of data to
14791          * speed up getting some GP measurement and
14792          * thus start pacing.
14793          */
14794         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
14795                 rack->app_limited_needs_set = 1;
14796                 tp->gput_ack = startseq + max(rc_init_window(rack),
14797                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
14798                 rack_log_pacing_delay_calc(rack,
14799                                            tp->gput_seq,
14800                                            tp->gput_ack,
14801                                            0,
14802                                            tp->gput_ts,
14803                                            rack->r_ctl.rc_app_limited_cnt,
14804                                            9,
14805                                            __LINE__, NULL, 0);
14806                 return;
14807         }
14808         if (sb_offset) {
14809                 /*
14810                  * We are out somewhere in the sb
14811                  * can we use the already outstanding data?
14812                  */
14813                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
14814                         /*
14815                          * Yes first one is good and in this case
14816                          * the tp->gput_ts is correctly set based on
14817                          * the last ack that arrived (no need to
14818                          * set things up when an ack comes in).
14819                          */
14820                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14821                         if ((my_rsm == NULL) ||
14822                             (my_rsm->r_rtr_cnt != 1)) {
14823                                 /* retransmission? */
14824                                 goto use_latest;
14825                         }
14826                 } else {
14827                         if (rack->r_ctl.rc_first_appl == NULL) {
14828                                 /*
14829                                  * If rc_first_appl is NULL
14830                                  * then the cnt should be 0.
14831                                  * This is probably an error, maybe
14832                                  * a KASSERT would be approprate.
14833                                  */
14834                                 goto use_latest;
14835                         }
14836                         /*
14837                          * If we have a marker pointer to the last one that is
14838                          * app limited we can use that, but we need to set
14839                          * things up so that when it gets ack'ed we record
14840                          * the ack time (if its not already acked).
14841                          */
14842                         rack->app_limited_needs_set = 1;
14843                         /*
14844                          * We want to get to the rsm that is either
14845                          * next with space i.e. over 1 MSS or the one
14846                          * after that (after the app-limited).
14847                          */
14848                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14849                                          rack->r_ctl.rc_first_appl);
14850                         if (my_rsm) {
14851                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
14852                                         /* Have to use the next one */
14853                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14854                                                          my_rsm);
14855                                 else {
14856                                         /* Use after the first MSS of it is acked */
14857                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
14858                                         goto start_set;
14859                                 }
14860                         }
14861                         if ((my_rsm == NULL) ||
14862                             (my_rsm->r_rtr_cnt != 1)) {
14863                                 /*
14864                                  * Either its a retransmit or
14865                                  * the last is the app-limited one.
14866                                  */
14867                                 goto use_latest;
14868                         }
14869                 }
14870                 tp->gput_seq = my_rsm->r_start;
14871 start_set:
14872                 if (my_rsm->r_flags & RACK_ACKED) {
14873                         /*
14874                          * This one has been acked use the arrival ack time
14875                          */
14876                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14877                         rack->app_limited_needs_set = 0;
14878                 }
14879                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14880                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
14881                 rack_log_pacing_delay_calc(rack,
14882                                            tp->gput_seq,
14883                                            tp->gput_ack,
14884                                            (uint64_t)my_rsm,
14885                                            tp->gput_ts,
14886                                            rack->r_ctl.rc_app_limited_cnt,
14887                                            9,
14888                                            __LINE__, NULL, 0);
14889                 return;
14890         }
14891
14892 use_latest:
14893         /*
14894          * We don't know how long we may have been
14895          * idle or if this is the first-send. Lets
14896          * setup the flag so we will trim off
14897          * the first ack'd data so we get a true
14898          * measurement.
14899          */
14900         rack->app_limited_needs_set = 1;
14901         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
14902         /* Find this guy so we can pull the send time */
14903         fe.r_start = startseq;
14904         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
14905         if (my_rsm) {
14906                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14907                 if (my_rsm->r_flags & RACK_ACKED) {
14908                         /*
14909                          * Unlikely since its probably what was
14910                          * just transmitted (but I am paranoid).
14911                          */
14912                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14913                         rack->app_limited_needs_set = 0;
14914                 }
14915                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
14916                         /* This also is unlikely */
14917                         tp->gput_seq = my_rsm->r_start;
14918                 }
14919         } else {
14920                 /*
14921                  * TSNH unless we have some send-map limit,
14922                  * and even at that it should not be hitting
14923                  * that limit (we should have stopped sending).
14924                  */
14925                 struct timeval tv;
14926
14927                 microuptime(&tv);
14928                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14929         }
14930         rack_log_pacing_delay_calc(rack,
14931                                    tp->gput_seq,
14932                                    tp->gput_ack,
14933                                    (uint64_t)my_rsm,
14934                                    tp->gput_ts,
14935                                    rack->r_ctl.rc_app_limited_cnt,
14936                                    9, __LINE__, NULL, 0);
14937 }
14938
14939 static inline uint32_t
14940 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
14941     uint32_t avail, int32_t sb_offset)
14942 {
14943         uint32_t len;
14944         uint32_t sendwin;
14945
14946         if (tp->snd_wnd > cwnd_to_use)
14947                 sendwin = cwnd_to_use;
14948         else
14949                 sendwin = tp->snd_wnd;
14950         if (ctf_outstanding(tp) >= tp->snd_wnd) {
14951                 /* We never want to go over our peers rcv-window */
14952                 len = 0;
14953         } else {
14954                 uint32_t flight;
14955
14956                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
14957                 if (flight >= sendwin) {
14958                         /*
14959                          * We have in flight what we are allowed by cwnd (if
14960                          * it was rwnd blocking it would have hit above out
14961                          * >= tp->snd_wnd).
14962                          */
14963                         return (0);
14964                 }
14965                 len = sendwin - flight;
14966                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
14967                         /* We would send too much (beyond the rwnd) */
14968                         len = tp->snd_wnd - ctf_outstanding(tp);
14969                 }
14970                 if ((len + sb_offset) > avail) {
14971                         /*
14972                          * We don't have that much in the SB, how much is
14973                          * there?
14974                          */
14975                         len = avail - sb_offset;
14976                 }
14977         }
14978         return (len);
14979 }
14980
14981 static void
14982 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
14983              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
14984              int rsm_is_null, int optlen, int line, uint16_t mode)
14985 {
14986         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14987                 union tcp_log_stackspecific log;
14988                 struct timeval tv;
14989
14990                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14991                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
14992                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
14993                 log.u_bbr.flex1 = error;
14994                 log.u_bbr.flex2 = flags;
14995                 log.u_bbr.flex3 = rsm_is_null;
14996                 log.u_bbr.flex4 = ipoptlen;
14997                 log.u_bbr.flex5 = tp->rcv_numsacks;
14998                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
14999                 log.u_bbr.flex7 = optlen;
15000                 log.u_bbr.flex8 = rack->r_fsb_inited;
15001                 log.u_bbr.applimited = rack->r_fast_output;
15002                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15003                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15004                 log.u_bbr.cwnd_gain = mode;
15005                 log.u_bbr.pkts_out = orig_len;
15006                 log.u_bbr.lt_epoch = len;
15007                 log.u_bbr.delivered = line;
15008                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15009                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15010                 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15011                                len, &log, false, NULL, NULL, 0, &tv);
15012         }
15013 }
15014
15015
15016 static struct mbuf *
15017 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15018                    struct rack_fast_send_blk *fsb,
15019                    int32_t seglimit, int32_t segsize, int hw_tls)
15020 {
15021 #ifdef KERN_TLS
15022         struct ktls_session *tls, *ntls;
15023         struct mbuf *start;
15024 #endif
15025         struct mbuf *m, *n, **np, *smb;
15026         struct mbuf *top;
15027         int32_t off, soff;
15028         int32_t len = *plen;
15029         int32_t fragsize;
15030         int32_t len_cp = 0;
15031         uint32_t mlen, frags;
15032
15033         soff = off = the_off;
15034         smb = m = the_m;
15035         np = &top;
15036         top = NULL;
15037 #ifdef KERN_TLS
15038         if (hw_tls && (m->m_flags & M_EXTPG))
15039                 tls = m->m_epg_tls;
15040         else
15041                 tls = NULL;
15042         start = m;
15043 #endif
15044         while (len > 0) {
15045                 if (m == NULL) {
15046                         *plen = len_cp;
15047                         break;
15048                 }
15049 #ifdef KERN_TLS
15050                 if (hw_tls) {
15051                         if (m->m_flags & M_EXTPG)
15052                                 ntls = m->m_epg_tls;
15053                         else
15054                                 ntls = NULL;
15055
15056                         /*
15057                          * Avoid mixing TLS records with handshake
15058                          * data or TLS records from different
15059                          * sessions.
15060                          */
15061                         if (tls != ntls) {
15062                                 MPASS(m != start);
15063                                 *plen = len_cp;
15064                                 break;
15065                         }
15066                 }
15067 #endif
15068                 mlen = min(len, m->m_len - off);
15069                 if (seglimit) {
15070                         /*
15071                          * For M_EXTPG mbufs, add 3 segments
15072                          * + 1 in case we are crossing page boundaries
15073                          * + 2 in case the TLS hdr/trailer are used
15074                          * It is cheaper to just add the segments
15075                          * than it is to take the cache miss to look
15076                          * at the mbuf ext_pgs state in detail.
15077                          */
15078                         if (m->m_flags & M_EXTPG) {
15079                                 fragsize = min(segsize, PAGE_SIZE);
15080                                 frags = 3;
15081                         } else {
15082                                 fragsize = segsize;
15083                                 frags = 0;
15084                         }
15085
15086                         /* Break if we really can't fit anymore. */
15087                         if ((frags + 1) >= seglimit) {
15088                                 *plen = len_cp;
15089                                 break;
15090                         }
15091
15092                         /*
15093                          * Reduce size if you can't copy the whole
15094                          * mbuf. If we can't copy the whole mbuf, also
15095                          * adjust len so the loop will end after this
15096                          * mbuf.
15097                          */
15098                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15099                                 mlen = (seglimit - frags - 1) * fragsize;
15100                                 len = mlen;
15101                                 *plen = len_cp + len;
15102                         }
15103                         frags += howmany(mlen, fragsize);
15104                         if (frags == 0)
15105                                 frags++;
15106                         seglimit -= frags;
15107                         KASSERT(seglimit > 0,
15108                             ("%s: seglimit went too low", __func__));
15109                 }
15110                 n = m_get(M_NOWAIT, m->m_type);
15111                 *np = n;
15112                 if (n == NULL)
15113                         goto nospace;
15114                 n->m_len = mlen;
15115                 soff += mlen;
15116                 len_cp += n->m_len;
15117                 if (m->m_flags & (M_EXT|M_EXTPG)) {
15118                         n->m_data = m->m_data + off;
15119                         mb_dupcl(n, m);
15120                 } else {
15121                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15122                             (u_int)n->m_len);
15123                 }
15124                 len -= n->m_len;
15125                 off = 0;
15126                 m = m->m_next;
15127                 np = &n->m_next;
15128                 if (len || (soff == smb->m_len)) {
15129                         /*
15130                          * We have more so we move forward  or
15131                          * we have consumed the entire mbuf and
15132                          * len has fell to 0.
15133                          */
15134                         soff = 0;
15135                         smb = m;
15136                 }
15137
15138         }
15139         if (fsb != NULL) {
15140                 fsb->m = smb;
15141                 fsb->off = soff;
15142                 if (smb) {
15143                         /*
15144                          * Save off the size of the mbuf. We do
15145                          * this so that we can recognize when it
15146                          * has been trimmed by sbcut() as acks
15147                          * come in.
15148                          */
15149                         fsb->o_m_len = smb->m_len;
15150                 } else {
15151                         /*
15152                          * This is the case where the next mbuf went to NULL. This
15153                          * means with this copy we have sent everything in the sb.
15154                          * In theory we could clear the fast_output flag, but lets
15155                          * not since its possible that we could get more added
15156                          * and acks that call the extend function which would let
15157                          * us send more.
15158                          */
15159                         fsb->o_m_len = 0;
15160                 }
15161         }
15162         return (top);
15163 nospace:
15164         if (top)
15165                 m_freem(top);
15166         return (NULL);
15167
15168 }
15169
15170 /*
15171  * This is a copy of m_copym(), taking the TSO segment size/limit
15172  * constraints into account, and advancing the sndptr as it goes.
15173  */
15174 static struct mbuf *
15175 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15176                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15177 {
15178         struct mbuf *m, *n;
15179         int32_t soff;
15180
15181         soff = rack->r_ctl.fsb.off;
15182         m = rack->r_ctl.fsb.m;
15183         if (rack->r_ctl.fsb.o_m_len != m->m_len) {
15184                 /*
15185                  * The mbuf had the front of it chopped off by an ack
15186                  * we need to adjust the soff/off by that difference.
15187                  */
15188                 uint32_t delta;
15189
15190                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15191                 soff -= delta;
15192         }
15193         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15194         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15195         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15196                                  __FUNCTION__,
15197                                  rack, *plen, m, m->m_len));
15198         /* Save off the right location before we copy and advance */
15199         *s_soff = soff;
15200         *s_mb = rack->r_ctl.fsb.m;
15201         n = rack_fo_base_copym(m, soff, plen,
15202                                &rack->r_ctl.fsb,
15203                                seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15204         return (n);
15205 }
15206
15207 static int
15208 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15209                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15210 {
15211         /*
15212          * Enter the fast retransmit path. We are given that a sched_pin is
15213          * in place (if accounting is compliled in) and the cycle count taken
15214          * at the entry is in the ts_val. The concept her is that the rsm
15215          * now holds the mbuf offsets and such so we can directly transmit
15216          * without a lot of overhead, the len field is already set for
15217          * us to prohibit us from sending too much (usually its 1MSS).
15218          */
15219         struct ip *ip = NULL;
15220         struct udphdr *udp = NULL;
15221         struct tcphdr *th = NULL;
15222         struct mbuf *m = NULL;
15223         struct inpcb *inp;
15224         uint8_t *cpto;
15225         struct tcp_log_buffer *lgb;
15226 #ifdef TCP_ACCOUNTING
15227         uint64_t crtsc;
15228         int cnt_thru = 1;
15229 #endif
15230         struct tcpopt to;
15231         u_char opt[TCP_MAXOLEN];
15232         uint32_t hdrlen, optlen;
15233         int32_t slot, segsiz, max_val, tso = 0, error = 0, flags, ulen = 0;
15234         uint32_t us_cts;
15235         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15236         uint32_t if_hw_tsomaxsegsize;
15237
15238 #ifdef INET6
15239         struct ip6_hdr *ip6 = NULL;
15240
15241         if (rack->r_is_v6) {
15242                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15243                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15244         } else
15245 #endif                          /* INET6 */
15246         {
15247                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15248                 hdrlen = sizeof(struct tcpiphdr);
15249         }
15250         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15251                 goto failed;
15252         }
15253         if (rsm->r_flags & RACK_TLP)
15254                 doing_tlp = 1;
15255         else if (doing_tlp)
15256                 rsm->r_flags |= RACK_TLP;
15257         startseq = rsm->r_start;
15258         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15259         inp = rack->rc_inp;
15260         to.to_flags = 0;
15261         flags = tcp_outflags[tp->t_state];
15262         if (flags & (TH_SYN|TH_RST)) {
15263                 goto failed;
15264         }
15265         if (rsm->r_flags & RACK_HAS_FIN) {
15266                 /* We can't send a FIN here */
15267                 goto failed;
15268         }
15269         if (flags & TH_FIN) {
15270                 /* We never send a FIN */
15271                 flags &= ~TH_FIN;
15272         }
15273         if (tp->t_flags & TF_RCVD_TSTMP) {
15274                 to.to_tsval = ms_cts + tp->ts_offset;
15275                 to.to_tsecr = tp->ts_recent;
15276                 to.to_flags = TOF_TS;
15277         }
15278         optlen = tcp_addoptions(&to, opt);
15279         hdrlen += optlen;
15280         udp = rack->r_ctl.fsb.udp;
15281         if (udp)
15282                 hdrlen += sizeof(struct udphdr);
15283         if (rack->r_ctl.rc_pace_max_segs)
15284                 max_val = rack->r_ctl.rc_pace_max_segs;
15285         else if (rack->rc_user_set_max_segs)
15286                 max_val = rack->rc_user_set_max_segs * segsiz;
15287         else
15288                 max_val = len;
15289         if ((tp->t_flags & TF_TSO) &&
15290             V_tcp_do_tso &&
15291             (len > segsiz) &&
15292             (tp->t_port == 0))
15293                 tso = 1;
15294 #ifdef INET6
15295         if (MHLEN < hdrlen + max_linkhdr)
15296                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15297         else
15298 #endif
15299                 m = m_gethdr(M_NOWAIT, MT_DATA);
15300         if (m == NULL)
15301                 goto failed;
15302         m->m_data += max_linkhdr;
15303         m->m_len = hdrlen;
15304         th = rack->r_ctl.fsb.th;
15305         /* Establish the len to send */
15306         if (len > max_val)
15307                 len = max_val;
15308         if ((tso) && (len + optlen > tp->t_maxseg)) {
15309                 uint32_t if_hw_tsomax;
15310                 int32_t max_len;
15311
15312                 /* extract TSO information */
15313                 if_hw_tsomax = tp->t_tsomax;
15314                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15315                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15316                 /*
15317                  * Check if we should limit by maximum payload
15318                  * length:
15319                  */
15320                 if (if_hw_tsomax != 0) {
15321                         /* compute maximum TSO length */
15322                         max_len = (if_hw_tsomax - hdrlen -
15323                                    max_linkhdr);
15324                         if (max_len <= 0) {
15325                                 goto failed;
15326                         } else if (len > max_len) {
15327                                 len = max_len;
15328                         }
15329                 }
15330                 if (len <= segsiz) {
15331                         /*
15332                          * In case there are too many small fragments don't
15333                          * use TSO:
15334                          */
15335                         tso = 0;
15336                 }
15337         } else {
15338                 tso = 0;
15339         }
15340         if ((tso == 0) && (len > segsiz))
15341                 len = segsiz;
15342         us_cts = tcp_get_usecs(tv);
15343         if ((len == 0) ||
15344             (len <= MHLEN - hdrlen - max_linkhdr)) {
15345                 goto failed;
15346         }
15347         th->th_seq = htonl(rsm->r_start);
15348         th->th_ack = htonl(tp->rcv_nxt);
15349         /*
15350          * The PUSH bit should only be applied
15351          * if the full retransmission is made. If
15352          * we are sending less than this is the
15353          * left hand edge and should not have
15354          * the PUSH bit.
15355          */
15356         if ((rsm->r_flags & RACK_HAD_PUSH) &&
15357             (len == (rsm->r_end - rsm->r_start)))
15358                 flags |= TH_PUSH;
15359         th->th_flags = flags;
15360         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15361         if (th->th_win == 0) {
15362                 tp->t_sndzerowin++;
15363                 tp->t_flags |= TF_RXWIN0SENT;
15364         } else
15365                 tp->t_flags &= ~TF_RXWIN0SENT;
15366         if (rsm->r_flags & RACK_TLP) {
15367                 /*
15368                  * TLP should not count in retran count, but
15369                  * in its own bin
15370                  */
15371                 counter_u64_add(rack_tlp_retran, 1);
15372                 counter_u64_add(rack_tlp_retran_bytes, len);
15373         } else {
15374                 tp->t_sndrexmitpack++;
15375                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15376                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15377         }
15378 #ifdef STATS
15379         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15380                                  len);
15381 #endif
15382         if (rsm->m == NULL)
15383                 goto failed;
15384         if (rsm->orig_m_len != rsm->m->m_len) {
15385                 /* Fix up the orig_m_len and possibly the mbuf offset */
15386                 rack_adjust_orig_mlen(rsm);
15387         }
15388         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15389         if (len <= segsiz) {
15390                 /*
15391                  * Must have ran out of mbufs for the copy
15392                  * shorten it to no longer need tso. Lets
15393                  * not put on sendalot since we are low on
15394                  * mbufs.
15395                  */
15396                 tso = 0;
15397         }
15398         if ((m->m_next == NULL) || (len <= 0)){
15399                 goto failed;
15400         }
15401         if (udp) {
15402                 if (rack->r_is_v6)
15403                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15404                 else
15405                         ulen = hdrlen + len - sizeof(struct ip);
15406                 udp->uh_ulen = htons(ulen);
15407         }
15408         m->m_pkthdr.rcvif = (struct ifnet *)0;
15409         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15410 #ifdef INET6
15411         if (rack->r_is_v6) {
15412                 if (tp->t_port) {
15413                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15414                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15415                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15416                         th->th_sum = htons(0);
15417                         UDPSTAT_INC(udps_opackets);
15418                 } else {
15419                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15420                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15421                         th->th_sum = in6_cksum_pseudo(ip6,
15422                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15423                                                       0);
15424                 }
15425         }
15426 #endif
15427 #if defined(INET6) && defined(INET)
15428         else
15429 #endif
15430 #ifdef INET
15431         {
15432                 if (tp->t_port) {
15433                         m->m_pkthdr.csum_flags = CSUM_UDP;
15434                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15435                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15436                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15437                         th->th_sum = htons(0);
15438                         UDPSTAT_INC(udps_opackets);
15439                 } else {
15440                         m->m_pkthdr.csum_flags = CSUM_TCP;
15441                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15442                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15443                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15444                                                                         IPPROTO_TCP + len + optlen));
15445                 }
15446                 /* IP version must be set here for ipv4/ipv6 checking later */
15447                 KASSERT(ip->ip_v == IPVERSION,
15448                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15449         }
15450 #endif
15451         if (tso) {
15452                 KASSERT(len > tp->t_maxseg - optlen,
15453                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15454                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15455                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15456         }
15457 #ifdef INET6
15458         if (rack->r_is_v6) {
15459                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15460                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15461                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15462                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15463                 else
15464                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15465         }
15466 #endif
15467 #if defined(INET) && defined(INET6)
15468         else
15469 #endif
15470 #ifdef INET
15471         {
15472                 ip->ip_len = htons(m->m_pkthdr.len);
15473                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15474                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15475                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15476                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15477                                 ip->ip_off |= htons(IP_DF);
15478                         }
15479                 } else {
15480                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15481                 }
15482         }
15483 #endif
15484         /* Time to copy in our header */
15485         cpto = mtod(m, uint8_t *);
15486         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15487         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15488         if (optlen) {
15489                 bcopy(opt, th + 1, optlen);
15490                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15491         } else {
15492                 th->th_off = sizeof(struct tcphdr) >> 2;
15493         }
15494         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15495                 union tcp_log_stackspecific log;
15496
15497                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15498                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15499                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15500                 if (rack->rack_no_prr)
15501                         log.u_bbr.flex1 = 0;
15502                 else
15503                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15504                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15505                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15506                 log.u_bbr.flex4 = max_val;
15507                 log.u_bbr.flex5 = 0;
15508                 /* Save off the early/late values */
15509                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15510                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15511                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15512                 if (doing_tlp == 0)
15513                         log.u_bbr.flex8 = 1;
15514                 else
15515                         log.u_bbr.flex8 = 2;
15516                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15517                 log.u_bbr.flex7 = 55;
15518                 log.u_bbr.pkts_out = tp->t_maxseg;
15519                 log.u_bbr.timeStamp = cts;
15520                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15521                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15522                 log.u_bbr.delivered = 0;
15523                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15524                                      len, &log, false, NULL, NULL, 0, tv);
15525         } else
15526                 lgb = NULL;
15527 #ifdef INET6
15528         if (rack->r_is_v6) {
15529                 error = ip6_output(m, NULL,
15530                                    &inp->inp_route6,
15531                                    0, NULL, NULL, inp);
15532         }
15533 #endif
15534 #if defined(INET) && defined(INET6)
15535         else
15536 #endif
15537 #ifdef INET
15538         {
15539                 error = ip_output(m, NULL,
15540                                   &inp->inp_route,
15541                                   0, 0, inp);
15542         }
15543 #endif
15544         m = NULL;
15545         if (lgb) {
15546                 lgb->tlb_errno = error;
15547                 lgb = NULL;
15548         }
15549         if (error) {
15550                 goto failed;
15551         }
15552         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15553                         rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
15554         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15555                 rack->rc_tlp_in_progress = 1;
15556                 rack->r_ctl.rc_tlp_cnt_out++;
15557         }
15558         if (error == 0)
15559                 tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
15560         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15561         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15562         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15563                 rack->r_ctl.retran_during_recovery += len;
15564         {
15565                 int idx;
15566
15567                 idx = (len / segsiz) + 3;
15568                 if (idx >= TCP_MSS_ACCT_ATIMER)
15569                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15570                 else
15571                         counter_u64_add(rack_out_size[idx], 1);
15572         }
15573         if (tp->t_rtttime == 0) {
15574                 tp->t_rtttime = ticks;
15575                 tp->t_rtseq = startseq;
15576                 KMOD_TCPSTAT_INC(tcps_segstimed);
15577         }
15578         counter_u64_add(rack_fto_rsm_send, 1);
15579         if (error && (error == ENOBUFS)) {
15580                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15581                 if (rack->rc_enobuf < 0x7f)
15582                         rack->rc_enobuf++;
15583                 if (slot < (10 * HPTS_USEC_IN_MSEC))
15584                         slot = 10 * HPTS_USEC_IN_MSEC;
15585         } else
15586                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15587         if ((slot == 0) ||
15588             (rack->rc_always_pace == 0) ||
15589             (rack->r_rr_config == 1)) {
15590                 /*
15591                  * We have no pacing set or we
15592                  * are using old-style rack or
15593                  * we are overriden to use the old 1ms pacing.
15594                  */
15595                 slot = rack->r_ctl.rc_min_to;
15596         }
15597         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
15598         if (rack->r_must_retran) {
15599                 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
15600                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
15601                         /*
15602                          * We have retransmitted all we need.
15603                          */
15604                         rack->r_must_retran = 0;
15605                         rack->r_ctl.rc_out_at_rto = 0;
15606                 }
15607         }
15608 #ifdef TCP_ACCOUNTING
15609         crtsc = get_cyclecount();
15610         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15611                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15612         }
15613         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15614         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15615                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15616         }
15617         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15618         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15619                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
15620         }
15621         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
15622         sched_unpin();
15623 #endif
15624         return (0);
15625 failed:
15626         if (m)
15627                 m_free(m);
15628         return (-1);
15629 }
15630
15631 static void
15632 rack_sndbuf_autoscale(struct tcp_rack *rack)
15633 {
15634         /*
15635          * Automatic sizing of send socket buffer.  Often the send buffer
15636          * size is not optimally adjusted to the actual network conditions
15637          * at hand (delay bandwidth product).  Setting the buffer size too
15638          * small limits throughput on links with high bandwidth and high
15639          * delay (eg. trans-continental/oceanic links).  Setting the
15640          * buffer size too big consumes too much real kernel memory,
15641          * especially with many connections on busy servers.
15642          *
15643          * The criteria to step up the send buffer one notch are:
15644          *  1. receive window of remote host is larger than send buffer
15645          *     (with a fudge factor of 5/4th);
15646          *  2. send buffer is filled to 7/8th with data (so we actually
15647          *     have data to make use of it);
15648          *  3. send buffer fill has not hit maximal automatic size;
15649          *  4. our send window (slow start and cogestion controlled) is
15650          *     larger than sent but unacknowledged data in send buffer.
15651          *
15652          * Note that the rack version moves things much faster since
15653          * we want to avoid hitting cache lines in the rack_fast_output()
15654          * path so this is called much less often and thus moves
15655          * the SB forward by a percentage.
15656          */
15657         struct socket *so;
15658         struct tcpcb *tp;
15659         uint32_t sendwin, scaleup;
15660
15661         tp = rack->rc_tp;
15662         so = rack->rc_inp->inp_socket;
15663         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
15664         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
15665                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
15666                     sbused(&so->so_snd) >=
15667                     (so->so_snd.sb_hiwat / 8 * 7) &&
15668                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
15669                     sendwin >= (sbused(&so->so_snd) -
15670                     (tp->snd_nxt - tp->snd_una))) {
15671                         if (rack_autosndbuf_inc)
15672                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
15673                         else
15674                                 scaleup = V_tcp_autosndbuf_inc;
15675                         if (scaleup < V_tcp_autosndbuf_inc)
15676                                 scaleup = V_tcp_autosndbuf_inc;
15677                         scaleup += so->so_snd.sb_hiwat;
15678                         if (scaleup > V_tcp_autosndbuf_max)
15679                                 scaleup = V_tcp_autosndbuf_max;
15680                         if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
15681                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
15682                 }
15683         }
15684 }
15685
15686 static int
15687 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
15688                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
15689 {
15690         /*
15691          * Enter to do fast output. We are given that the sched_pin is
15692          * in place (if accounting is compiled in) and the cycle count taken
15693          * at entry is in place in ts_val. The idea here is that
15694          * we know how many more bytes needs to be sent (presumably either
15695          * during pacing or to fill the cwnd and that was greater than
15696          * the max-burst). We have how much to send and all the info we
15697          * need to just send.
15698          */
15699         struct ip *ip = NULL;
15700         struct udphdr *udp = NULL;
15701         struct tcphdr *th = NULL;
15702         struct mbuf *m, *s_mb;
15703         struct inpcb *inp;
15704         uint8_t *cpto;
15705         struct tcp_log_buffer *lgb;
15706 #ifdef TCP_ACCOUNTING
15707         uint64_t crtsc;
15708 #endif
15709         struct tcpopt to;
15710         u_char opt[TCP_MAXOLEN];
15711         uint32_t hdrlen, optlen;
15712         int cnt_thru = 1;
15713         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error = 0, flags, ulen = 0;
15714         uint32_t us_cts, s_soff;
15715         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15716         uint32_t if_hw_tsomaxsegsize;
15717         uint16_t add_flag = RACK_SENT_FP;
15718 #ifdef INET6
15719         struct ip6_hdr *ip6 = NULL;
15720
15721         if (rack->r_is_v6) {
15722                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15723                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15724         } else
15725 #endif                          /* INET6 */
15726         {
15727                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15728                 hdrlen = sizeof(struct tcpiphdr);
15729         }
15730         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15731                 m = NULL;
15732                 goto failed;
15733         }
15734         startseq = tp->snd_max;
15735         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15736         inp = rack->rc_inp;
15737         len = rack->r_ctl.fsb.left_to_send;
15738         to.to_flags = 0;
15739         flags = rack->r_ctl.fsb.tcp_flags;
15740         if (tp->t_flags & TF_RCVD_TSTMP) {
15741                 to.to_tsval = ms_cts + tp->ts_offset;
15742                 to.to_tsecr = tp->ts_recent;
15743                 to.to_flags = TOF_TS;
15744         }
15745         optlen = tcp_addoptions(&to, opt);
15746         hdrlen += optlen;
15747         udp = rack->r_ctl.fsb.udp;
15748         if (udp)
15749                 hdrlen += sizeof(struct udphdr);
15750         if (rack->r_ctl.rc_pace_max_segs)
15751                 max_val = rack->r_ctl.rc_pace_max_segs;
15752         else if (rack->rc_user_set_max_segs)
15753                 max_val = rack->rc_user_set_max_segs * segsiz;
15754         else
15755                 max_val = len;
15756         if ((tp->t_flags & TF_TSO) &&
15757             V_tcp_do_tso &&
15758             (len > segsiz) &&
15759             (tp->t_port == 0))
15760                 tso = 1;
15761 again:
15762 #ifdef INET6
15763         if (MHLEN < hdrlen + max_linkhdr)
15764                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15765         else
15766 #endif
15767                 m = m_gethdr(M_NOWAIT, MT_DATA);
15768         if (m == NULL)
15769                 goto failed;
15770         m->m_data += max_linkhdr;
15771         m->m_len = hdrlen;
15772         th = rack->r_ctl.fsb.th;
15773         /* Establish the len to send */
15774         if (len > max_val)
15775                 len = max_val;
15776         if ((tso) && (len + optlen > tp->t_maxseg)) {
15777                 uint32_t if_hw_tsomax;
15778                 int32_t max_len;
15779
15780                 /* extract TSO information */
15781                 if_hw_tsomax = tp->t_tsomax;
15782                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15783                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15784                 /*
15785                  * Check if we should limit by maximum payload
15786                  * length:
15787                  */
15788                 if (if_hw_tsomax != 0) {
15789                         /* compute maximum TSO length */
15790                         max_len = (if_hw_tsomax - hdrlen -
15791                                    max_linkhdr);
15792                         if (max_len <= 0) {
15793                                 goto failed;
15794                         } else if (len > max_len) {
15795                                 len = max_len;
15796                         }
15797                 }
15798                 if (len <= segsiz) {
15799                         /*
15800                          * In case there are too many small fragments don't
15801                          * use TSO:
15802                          */
15803                         tso = 0;
15804                 }
15805         } else {
15806                 tso = 0;
15807         }
15808         if ((tso == 0) && (len > segsiz))
15809                 len = segsiz;
15810         us_cts = tcp_get_usecs(tv);
15811         if ((len == 0) ||
15812             (len <= MHLEN - hdrlen - max_linkhdr)) {
15813                 goto failed;
15814         }
15815         sb_offset = tp->snd_max - tp->snd_una;
15816         th->th_seq = htonl(tp->snd_max);
15817         th->th_ack = htonl(tp->rcv_nxt);
15818         th->th_flags = flags;
15819         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15820         if (th->th_win == 0) {
15821                 tp->t_sndzerowin++;
15822                 tp->t_flags |= TF_RXWIN0SENT;
15823         } else
15824                 tp->t_flags &= ~TF_RXWIN0SENT;
15825         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
15826         KMOD_TCPSTAT_INC(tcps_sndpack);
15827         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
15828 #ifdef STATS
15829         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
15830                                  len);
15831 #endif
15832         if (rack->r_ctl.fsb.m == NULL)
15833                 goto failed;
15834
15835         /* s_mb and s_soff are saved for rack_log_output */
15836         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
15837                                     &s_mb, &s_soff);
15838         if (len <= segsiz) {
15839                 /*
15840                  * Must have ran out of mbufs for the copy
15841                  * shorten it to no longer need tso. Lets
15842                  * not put on sendalot since we are low on
15843                  * mbufs.
15844                  */
15845                 tso = 0;
15846         }
15847         if (rack->r_ctl.fsb.rfo_apply_push &&
15848             (len == rack->r_ctl.fsb.left_to_send)) {
15849                 th->th_flags |= TH_PUSH;
15850                 add_flag |= RACK_HAD_PUSH;
15851         }
15852         if ((m->m_next == NULL) || (len <= 0)){
15853                 goto failed;
15854         }
15855         if (udp) {
15856                 if (rack->r_is_v6)
15857                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15858                 else
15859                         ulen = hdrlen + len - sizeof(struct ip);
15860                 udp->uh_ulen = htons(ulen);
15861         }
15862         m->m_pkthdr.rcvif = (struct ifnet *)0;
15863         if (tp->t_state == TCPS_ESTABLISHED &&
15864             (tp->t_flags2 & TF2_ECN_PERMIT)) {
15865                 /*
15866                  * If the peer has ECN, mark data packets with ECN capable
15867                  * transmission (ECT). Ignore pure ack packets,
15868                  * retransmissions.
15869                  */
15870                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) {
15871 #ifdef INET6
15872                         if (rack->r_is_v6)
15873                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
15874                         else
15875 #endif
15876                                 ip->ip_tos |= IPTOS_ECN_ECT0;
15877                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
15878                         /*
15879                          * Reply with proper ECN notifications.
15880                          * Only set CWR on new data segments.
15881                          */
15882                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
15883                                 flags |= TH_CWR;
15884                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
15885                         }
15886                 }
15887                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
15888                         flags |= TH_ECE;
15889         }
15890         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15891 #ifdef INET6
15892         if (rack->r_is_v6) {
15893                 if (tp->t_port) {
15894                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15895                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15896                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15897                         th->th_sum = htons(0);
15898                         UDPSTAT_INC(udps_opackets);
15899                 } else {
15900                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15901                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15902                         th->th_sum = in6_cksum_pseudo(ip6,
15903                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15904                                                       0);
15905                 }
15906         }
15907 #endif
15908 #if defined(INET6) && defined(INET)
15909         else
15910 #endif
15911 #ifdef INET
15912         {
15913                 if (tp->t_port) {
15914                         m->m_pkthdr.csum_flags = CSUM_UDP;
15915                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15916                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15917                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15918                         th->th_sum = htons(0);
15919                         UDPSTAT_INC(udps_opackets);
15920                 } else {
15921                         m->m_pkthdr.csum_flags = CSUM_TCP;
15922                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15923                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15924                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15925                                                                         IPPROTO_TCP + len + optlen));
15926                 }
15927                 /* IP version must be set here for ipv4/ipv6 checking later */
15928                 KASSERT(ip->ip_v == IPVERSION,
15929                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15930         }
15931 #endif
15932         if (tso) {
15933                 KASSERT(len > tp->t_maxseg - optlen,
15934                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15935                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15936                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15937         }
15938 #ifdef INET6
15939         if (rack->r_is_v6) {
15940                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15941                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15942                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15943                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15944                 else
15945                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15946         }
15947 #endif
15948 #if defined(INET) && defined(INET6)
15949         else
15950 #endif
15951 #ifdef INET
15952         {
15953                 ip->ip_len = htons(m->m_pkthdr.len);
15954                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15955                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15956                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15957                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15958                                 ip->ip_off |= htons(IP_DF);
15959                         }
15960                 } else {
15961                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15962                 }
15963         }
15964 #endif
15965         /* Time to copy in our header */
15966         cpto = mtod(m, uint8_t *);
15967         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15968         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15969         if (optlen) {
15970                 bcopy(opt, th + 1, optlen);
15971                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15972         } else {
15973                 th->th_off = sizeof(struct tcphdr) >> 2;
15974         }
15975         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15976                 union tcp_log_stackspecific log;
15977
15978                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15979                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15980                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15981                 if (rack->rack_no_prr)
15982                         log.u_bbr.flex1 = 0;
15983                 else
15984                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15985                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15986                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15987                 log.u_bbr.flex4 = max_val;
15988                 log.u_bbr.flex5 = 0;
15989                 /* Save off the early/late values */
15990                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15991                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15992                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15993                 log.u_bbr.flex8 = 0;
15994                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15995                 log.u_bbr.flex7 = 44;
15996                 log.u_bbr.pkts_out = tp->t_maxseg;
15997                 log.u_bbr.timeStamp = cts;
15998                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15999                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16000                 log.u_bbr.delivered = 0;
16001                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16002                                      len, &log, false, NULL, NULL, 0, tv);
16003         } else
16004                 lgb = NULL;
16005 #ifdef INET6
16006         if (rack->r_is_v6) {
16007                 error = ip6_output(m, NULL,
16008                                    &inp->inp_route6,
16009                                    0, NULL, NULL, inp);
16010         }
16011 #endif
16012 #if defined(INET) && defined(INET6)
16013         else
16014 #endif
16015 #ifdef INET
16016         {
16017                 error = ip_output(m, NULL,
16018                                   &inp->inp_route,
16019                                   0, 0, inp);
16020         }
16021 #endif
16022         if (lgb) {
16023                 lgb->tlb_errno = error;
16024                 lgb = NULL;
16025         }
16026         if (error) {
16027                 *send_err = error;
16028                 m = NULL;
16029                 goto failed;
16030         }
16031         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16032                         NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16033         m = NULL;
16034         if (tp->snd_una == tp->snd_max) {
16035                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
16036                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16037                 tp->t_acktime = ticks;
16038         }
16039         if (error == 0)
16040                 tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16041
16042         rack->forced_ack = 0;   /* If we send something zap the FA flag */
16043         tot_len += len;
16044         if ((tp->t_flags & TF_GPUTINPROG) == 0)
16045                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16046         tp->snd_max += len;
16047         tp->snd_nxt = tp->snd_max;
16048         {
16049                 int idx;
16050
16051                 idx = (len / segsiz) + 3;
16052                 if (idx >= TCP_MSS_ACCT_ATIMER)
16053                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16054                 else
16055                         counter_u64_add(rack_out_size[idx], 1);
16056         }
16057         if (len <= rack->r_ctl.fsb.left_to_send)
16058                 rack->r_ctl.fsb.left_to_send -= len;
16059         else
16060                 rack->r_ctl.fsb.left_to_send = 0;
16061         if (rack->r_ctl.fsb.left_to_send < segsiz) {
16062                 rack->r_fast_output = 0;
16063                 rack->r_ctl.fsb.left_to_send = 0;
16064                 /* At the end of fast_output scale up the sb */
16065                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16066                 rack_sndbuf_autoscale(rack);
16067                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16068         }
16069         if (tp->t_rtttime == 0) {
16070                 tp->t_rtttime = ticks;
16071                 tp->t_rtseq = startseq;
16072                 KMOD_TCPSTAT_INC(tcps_segstimed);
16073         }
16074         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16075             (max_val > len) &&
16076             (tso == 0)) {
16077                 max_val -= len;
16078                 len = segsiz;
16079                 th = rack->r_ctl.fsb.th;
16080                 cnt_thru++;
16081                 goto again;
16082         }
16083         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16084         counter_u64_add(rack_fto_send, 1);
16085         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16086         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16087 #ifdef TCP_ACCOUNTING
16088         crtsc = get_cyclecount();
16089         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16090                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16091         }
16092         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16093         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16094                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16095         }
16096         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16097         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16098                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16099         }
16100         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16101         sched_unpin();
16102 #endif
16103         return (0);
16104 failed:
16105         if (m)
16106                 m_free(m);
16107         rack->r_fast_output = 0;
16108         return (-1);
16109 }
16110
16111 static int
16112 rack_output(struct tcpcb *tp)
16113 {
16114         struct socket *so;
16115         uint32_t recwin;
16116         uint32_t sb_offset, s_moff = 0;
16117         int32_t len, flags, error = 0;
16118         struct mbuf *m, *s_mb = NULL;
16119         struct mbuf *mb;
16120         uint32_t if_hw_tsomaxsegcount = 0;
16121         uint32_t if_hw_tsomaxsegsize;
16122         int32_t segsiz, minseg;
16123         long tot_len_this_send = 0;
16124 #ifdef INET
16125         struct ip *ip = NULL;
16126 #ifdef TCPDEBUG
16127         struct ipovly *ipov = NULL;
16128 #endif
16129 #endif
16130         struct udphdr *udp = NULL;
16131         struct tcp_rack *rack;
16132         struct tcphdr *th;
16133         uint8_t pass = 0;
16134         uint8_t mark = 0;
16135         uint8_t wanted_cookie = 0;
16136         u_char opt[TCP_MAXOLEN];
16137         unsigned ipoptlen, optlen, hdrlen;
16138 #if defined(INET) || defined(INET6)
16139         unsigned ulen=0;
16140 #endif
16141         uint32_t rack_seq;
16142
16143 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16144         unsigned ipsec_optlen = 0;
16145
16146 #endif
16147         int32_t idle, sendalot;
16148         int32_t sub_from_prr = 0;
16149         volatile int32_t sack_rxmit;
16150         struct rack_sendmap *rsm = NULL;
16151         int32_t tso, mtu;
16152         struct tcpopt to;
16153         int32_t slot = 0;
16154         int32_t sup_rack = 0;
16155         uint32_t cts, ms_cts, delayed, early;
16156         uint16_t add_flag = RACK_SENT_SP;
16157         /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16158         uint8_t hpts_calling,  doing_tlp = 0;
16159         uint32_t cwnd_to_use, pace_max_seg;
16160         int32_t do_a_prefetch = 0;
16161         int32_t prefetch_rsm = 0;
16162         int32_t orig_len = 0;
16163         struct timeval tv;
16164         int32_t prefetch_so_done = 0;
16165         struct tcp_log_buffer *lgb;
16166         struct inpcb *inp;
16167         struct sockbuf *sb;
16168         uint64_t ts_val = 0;
16169 #ifdef TCP_ACCOUNTING
16170         uint64_t crtsc;
16171 #endif
16172 #ifdef INET6
16173         struct ip6_hdr *ip6 = NULL;
16174         int32_t isipv6;
16175 #endif
16176         uint8_t filled_all = 0;
16177         bool hw_tls = false;
16178
16179         /* setup and take the cache hits here */
16180         rack = (struct tcp_rack *)tp->t_fb_ptr;
16181 #ifdef TCP_ACCOUNTING
16182         sched_pin();
16183         ts_val = get_cyclecount();
16184 #endif
16185         hpts_calling = rack->rc_inp->inp_hpts_calls;
16186         NET_EPOCH_ASSERT();
16187         INP_WLOCK_ASSERT(rack->rc_inp);
16188 #ifdef TCP_OFFLOAD
16189         if (tp->t_flags & TF_TOE) {
16190 #ifdef TCP_ACCOUNTING
16191                 sched_unpin();
16192 #endif
16193                 return (tcp_offload_output(tp));
16194         }
16195 #endif
16196         /*
16197          * For TFO connections in SYN_RECEIVED, only allow the initial
16198          * SYN|ACK and those sent by the retransmit timer.
16199          */
16200         if (IS_FASTOPEN(tp->t_flags) &&
16201             (tp->t_state == TCPS_SYN_RECEIVED) &&
16202             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16203             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16204 #ifdef TCP_ACCOUNTING
16205                 sched_unpin();
16206 #endif
16207                 return (0);
16208         }
16209 #ifdef INET6
16210         if (rack->r_state) {
16211                 /* Use the cache line loaded if possible */
16212                 isipv6 = rack->r_is_v6;
16213         } else {
16214                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16215         }
16216 #endif
16217         early = 0;
16218         cts = tcp_get_usecs(&tv);
16219         ms_cts = tcp_tv_to_mssectick(&tv);
16220         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16221             rack->rc_inp->inp_in_hpts) {
16222                 /*
16223                  * We are on the hpts for some timer but not hptsi output.
16224                  * Remove from the hpts unconditionally.
16225                  */
16226                 rack_timer_cancel(tp, rack, cts, __LINE__);
16227         }
16228         /* Are we pacing and late? */
16229         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16230             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16231                 /* We are delayed */
16232                 delayed = cts - rack->r_ctl.rc_last_output_to;
16233         } else {
16234                 delayed = 0;
16235         }
16236         /* Do the timers, which may override the pacer */
16237         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16238                 if (rack_process_timers(tp, rack, cts, hpts_calling, &doing_tlp)) {
16239                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16240 #ifdef TCP_ACCOUNTING
16241                         sched_unpin();
16242 #endif
16243                         return (0);
16244                 }
16245         }
16246         if (rack->rc_in_persist) {
16247                 if (rack->rc_inp->inp_in_hpts == 0) {
16248                         /* Timer is not running */
16249                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16250                 }
16251 #ifdef TCP_ACCOUNTING
16252                 sched_unpin();
16253 #endif
16254                 return (0);
16255         }
16256         if ((rack->r_timer_override) ||
16257             (rack->rc_ack_can_sendout_data) ||
16258             (delayed) ||
16259             (tp->t_state < TCPS_ESTABLISHED)) {
16260                 rack->rc_ack_can_sendout_data = 0;
16261                 if (rack->rc_inp->inp_in_hpts)
16262                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
16263         } else if (rack->rc_inp->inp_in_hpts) {
16264                 /*
16265                  * On the hpts you can't pass even if ACKNOW is on, we will
16266                  * when the hpts fires.
16267                  */
16268 #ifdef TCP_ACCOUNTING
16269                 crtsc = get_cyclecount();
16270                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16271                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16272                 }
16273                 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16274                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16275                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16276                 }
16277                 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16278                 sched_unpin();
16279 #endif
16280                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16281                 return (0);
16282         }
16283         rack->rc_inp->inp_hpts_calls = 0;
16284         /* Finish out both pacing early and late accounting */
16285         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16286             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16287                 early = rack->r_ctl.rc_last_output_to - cts;
16288         } else
16289                 early = 0;
16290         if (delayed) {
16291                 rack->r_ctl.rc_agg_delayed += delayed;
16292                 rack->r_late = 1;
16293         } else if (early) {
16294                 rack->r_ctl.rc_agg_early += early;
16295                 rack->r_early = 1;
16296         }
16297         /* Now that early/late accounting is done turn off the flag */
16298         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16299         rack->r_wanted_output = 0;
16300         rack->r_timer_override = 0;
16301         if ((tp->t_state != rack->r_state) &&
16302             TCPS_HAVEESTABLISHED(tp->t_state)) {
16303                 rack_set_state(tp, rack);
16304         }
16305         if ((rack->r_fast_output) &&
16306             (doing_tlp == 0) &&
16307             (tp->rcv_numsacks == 0)) {
16308                 int ret;
16309
16310                 error = 0;
16311                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16312                 if (ret >= 0)
16313                         return(ret);
16314                 else if (error) {
16315                         inp = rack->rc_inp;
16316                         so = inp->inp_socket;
16317                         sb = &so->so_snd;
16318                         goto nomore;
16319                 }
16320         }
16321         inp = rack->rc_inp;
16322         /*
16323          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16324          * only allow the initial SYN or SYN|ACK and those sent
16325          * by the retransmit timer.
16326          */
16327         if (IS_FASTOPEN(tp->t_flags) &&
16328             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16329              (tp->t_state == TCPS_SYN_SENT)) &&
16330             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16331             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16332                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16333                 so = inp->inp_socket;
16334                 sb = &so->so_snd;
16335                 goto just_return_nolock;
16336         }
16337         /*
16338          * Determine length of data that should be transmitted, and flags
16339          * that will be used. If there is some data or critical controls
16340          * (SYN, RST) to send, then transmit; otherwise, investigate
16341          * further.
16342          */
16343         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16344         if (tp->t_idle_reduce) {
16345                 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16346                         rack_cc_after_idle(rack, tp);
16347         }
16348         tp->t_flags &= ~TF_LASTIDLE;
16349         if (idle) {
16350                 if (tp->t_flags & TF_MORETOCOME) {
16351                         tp->t_flags |= TF_LASTIDLE;
16352                         idle = 0;
16353                 }
16354         }
16355         if ((tp->snd_una == tp->snd_max) &&
16356             rack->r_ctl.rc_went_idle_time &&
16357             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16358                 idle = cts - rack->r_ctl.rc_went_idle_time;
16359                 if (idle > rack_min_probertt_hold) {
16360                         /* Count as a probe rtt */
16361                         if (rack->in_probe_rtt == 0) {
16362                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16363                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16364                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16365                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16366                         } else {
16367                                 rack_exit_probertt(rack, cts);
16368                         }
16369                 }
16370                 idle = 0;
16371         }
16372         if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16373                 rack_init_fsb_block(tp, rack);
16374 again:
16375         /*
16376          * If we've recently taken a timeout, snd_max will be greater than
16377          * snd_nxt.  There may be SACK information that allows us to avoid
16378          * resending already delivered data.  Adjust snd_nxt accordingly.
16379          */
16380         sendalot = 0;
16381         cts = tcp_get_usecs(&tv);
16382         ms_cts = tcp_tv_to_mssectick(&tv);
16383         tso = 0;
16384         mtu = 0;
16385         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16386         minseg = segsiz;
16387         if (rack->r_ctl.rc_pace_max_segs == 0)
16388                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16389         else
16390                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16391         sb_offset = tp->snd_max - tp->snd_una;
16392         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16393         flags = tcp_outflags[tp->t_state];
16394         while (rack->rc_free_cnt < rack_free_cache) {
16395                 rsm = rack_alloc(rack);
16396                 if (rsm == NULL) {
16397                         if (inp->inp_hpts_calls)
16398                                 /* Retry in a ms */
16399                                 slot = (1 * HPTS_USEC_IN_MSEC);
16400                         so = inp->inp_socket;
16401                         sb = &so->so_snd;
16402                         goto just_return_nolock;
16403                 }
16404                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16405                 rack->rc_free_cnt++;
16406                 rsm = NULL;
16407         }
16408         if (inp->inp_hpts_calls)
16409                 inp->inp_hpts_calls = 0;
16410         sack_rxmit = 0;
16411         len = 0;
16412         rsm = NULL;
16413         if (flags & TH_RST) {
16414                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16415                 so = inp->inp_socket;
16416                 sb = &so->so_snd;
16417                 goto send;
16418         }
16419         if (rack->r_ctl.rc_resend) {
16420                 /* Retransmit timer */
16421                 rsm = rack->r_ctl.rc_resend;
16422                 rack->r_ctl.rc_resend = NULL;
16423                 rsm->r_flags &= ~RACK_TLP;
16424                 len = rsm->r_end - rsm->r_start;
16425                 sack_rxmit = 1;
16426                 sendalot = 0;
16427                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16428                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16429                          __func__, __LINE__,
16430                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16431                 sb_offset = rsm->r_start - tp->snd_una;
16432                 if (len >= segsiz)
16433                         len = segsiz;
16434         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16435                 /* We have a retransmit that takes precedence */
16436                 rsm->r_flags &= ~RACK_TLP;
16437                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16438                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16439                         /* Enter recovery if not induced by a time-out */
16440                         rack->r_ctl.rc_rsm_start = rsm->r_start;
16441                         rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
16442                         rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
16443                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
16444                 }
16445 #ifdef INVARIANTS
16446                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16447                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16448                               tp, rack, rsm, rsm->r_start, tp->snd_una);
16449                 }
16450 #endif
16451                 len = rsm->r_end - rsm->r_start;
16452                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16453                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16454                          __func__, __LINE__,
16455                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16456                 sb_offset = rsm->r_start - tp->snd_una;
16457                 sendalot = 0;
16458                 if (len >= segsiz)
16459                         len = segsiz;
16460                 if (len > 0) {
16461                         sack_rxmit = 1;
16462                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16463                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16464                             min(len, segsiz));
16465                         counter_u64_add(rack_rtm_prr_retran, 1);
16466                 }
16467         } else if (rack->r_ctl.rc_tlpsend) {
16468                 /* Tail loss probe */
16469                 long cwin;
16470                 long tlen;
16471
16472                 /*
16473                  * Check if we can do a TLP with a RACK'd packet
16474                  * this can happen if we are not doing the rack
16475                  * cheat and we skipped to a TLP and it
16476                  * went off.
16477                  */
16478                 rsm = rack->r_ctl.rc_tlpsend;
16479                 rsm->r_flags |= RACK_TLP;
16480
16481                 rack->r_ctl.rc_tlpsend = NULL;
16482                 sack_rxmit = 1;
16483                 tlen = rsm->r_end - rsm->r_start;
16484                 if (tlen > segsiz)
16485                         tlen = segsiz;
16486                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16487                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16488                          __func__, __LINE__,
16489                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16490                 sb_offset = rsm->r_start - tp->snd_una;
16491                 cwin = min(tp->snd_wnd, tlen);
16492                 len = cwin;
16493         }
16494         if (rack->r_must_retran &&
16495             (rsm == NULL)) {
16496                 /*
16497                  * Non-Sack and we had a RTO or MTU change, we
16498                  * need to retransmit until we reach
16499                  * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
16500                  */
16501                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
16502                         int sendwin, flight;
16503
16504                         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16505                         flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16506                         if (flight >= sendwin) {
16507                                 so = inp->inp_socket;
16508                                 sb = &so->so_snd;
16509                                 goto just_return_nolock;
16510                         }
16511                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16512                         KASSERT(rsm != NULL, ("rsm is NULL rack:%p r_must_retran set", rack));
16513                         if (rsm == NULL) {
16514                                 /* TSNH */
16515                                 rack->r_must_retran = 0;
16516                                 rack->r_ctl.rc_out_at_rto = 0;
16517                                 rack->r_must_retran = 0;
16518                                 so = inp->inp_socket;
16519                                 sb = &so->so_snd;
16520                                 goto just_return_nolock;
16521                         }
16522                         sack_rxmit = 1;
16523                         len = rsm->r_end - rsm->r_start;
16524                         sendalot = 0;
16525                         sb_offset = rsm->r_start - tp->snd_una;
16526                         if (len >= segsiz)
16527                                 len = segsiz;
16528                 } else {
16529                         /* We must be done if there is nothing outstanding */
16530                         rack->r_must_retran = 0;
16531                         rack->r_ctl.rc_out_at_rto = 0;
16532                 }
16533         }
16534         /*
16535          * Enforce a connection sendmap count limit if set
16536          * as long as we are not retransmiting.
16537          */
16538         if ((rsm == NULL) &&
16539             (rack->do_detection == 0) &&
16540             (V_tcp_map_entries_limit > 0) &&
16541             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16542                 counter_u64_add(rack_to_alloc_limited, 1);
16543                 if (!rack->alloc_limit_reported) {
16544                         rack->alloc_limit_reported = 1;
16545                         counter_u64_add(rack_alloc_limited_conns, 1);
16546                 }
16547                 so = inp->inp_socket;
16548                 sb = &so->so_snd;
16549                 goto just_return_nolock;
16550         }
16551         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16552                 /* we are retransmitting the fin */
16553                 len--;
16554                 if (len) {
16555                         /*
16556                          * When retransmitting data do *not* include the
16557                          * FIN. This could happen from a TLP probe.
16558                          */
16559                         flags &= ~TH_FIN;
16560                 }
16561         }
16562 #ifdef INVARIANTS
16563         /* For debugging */
16564         rack->r_ctl.rc_rsm_at_retran = rsm;
16565 #endif
16566         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16567             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16568                 int ret;
16569
16570                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
16571                 if (ret == 0)
16572                         return (0);
16573         }
16574         so = inp->inp_socket;
16575         sb = &so->so_snd;
16576         if (do_a_prefetch == 0) {
16577                 kern_prefetch(sb, &do_a_prefetch);
16578                 do_a_prefetch = 1;
16579         }
16580 #ifdef NETFLIX_SHARED_CWND
16581         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
16582             rack->rack_enable_scwnd) {
16583                 /* We are doing cwnd sharing */
16584                 if (rack->gp_ready &&
16585                     (rack->rack_attempted_scwnd == 0) &&
16586                     (rack->r_ctl.rc_scw == NULL) &&
16587                     tp->t_lib) {
16588                         /* The pcbid is in, lets make an attempt */
16589                         counter_u64_add(rack_try_scwnd, 1);
16590                         rack->rack_attempted_scwnd = 1;
16591                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
16592                                                                    &rack->r_ctl.rc_scw_index,
16593                                                                    segsiz);
16594                 }
16595                 if (rack->r_ctl.rc_scw &&
16596                     (rack->rack_scwnd_is_idle == 1) &&
16597                     sbavail(&so->so_snd)) {
16598                         /* we are no longer out of data */
16599                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16600                         rack->rack_scwnd_is_idle = 0;
16601                 }
16602                 if (rack->r_ctl.rc_scw) {
16603                         /* First lets update and get the cwnd */
16604                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
16605                                                                     rack->r_ctl.rc_scw_index,
16606                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
16607                 }
16608         }
16609 #endif
16610         /*
16611          * Get standard flags, and add SYN or FIN if requested by 'hidden'
16612          * state flags.
16613          */
16614         if (tp->t_flags & TF_NEEDFIN)
16615                 flags |= TH_FIN;
16616         if (tp->t_flags & TF_NEEDSYN)
16617                 flags |= TH_SYN;
16618         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
16619                 void *end_rsm;
16620                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
16621                 if (end_rsm)
16622                         kern_prefetch(end_rsm, &prefetch_rsm);
16623                 prefetch_rsm = 1;
16624         }
16625         SOCKBUF_LOCK(sb);
16626         /*
16627          * If snd_nxt == snd_max and we have transmitted a FIN, the
16628          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
16629          * negative length.  This can also occur when TCP opens up its
16630          * congestion window while receiving additional duplicate acks after
16631          * fast-retransmit because TCP will reset snd_nxt to snd_max after
16632          * the fast-retransmit.
16633          *
16634          * In the normal retransmit-FIN-only case, however, snd_nxt will be
16635          * set to snd_una, the sb_offset will be 0, and the length may wind
16636          * up 0.
16637          *
16638          * If sack_rxmit is true we are retransmitting from the scoreboard
16639          * in which case len is already set.
16640          */
16641         if ((sack_rxmit == 0) &&
16642             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
16643                 uint32_t avail;
16644
16645                 avail = sbavail(sb);
16646                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
16647                         sb_offset = tp->snd_nxt - tp->snd_una;
16648                 else
16649                         sb_offset = 0;
16650                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
16651                         if (rack->r_ctl.rc_tlp_new_data) {
16652                                 /* TLP is forcing out new data */
16653                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
16654                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
16655                                 }
16656                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
16657                                         if (tp->snd_wnd > sb_offset)
16658                                                 len = tp->snd_wnd - sb_offset;
16659                                         else
16660                                                 len = 0;
16661                                 } else {
16662                                         len = rack->r_ctl.rc_tlp_new_data;
16663                                 }
16664                         }  else {
16665                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
16666                         }
16667                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
16668                                 /*
16669                                  * For prr=off, we need to send only 1 MSS
16670                                  * at a time. We do this because another sack could
16671                                  * be arriving that causes us to send retransmits and
16672                                  * we don't want to be on a long pace due to a larger send
16673                                  * that keeps us from sending out the retransmit.
16674                                  */
16675                                 len = segsiz;
16676                         }
16677                 } else {
16678                         uint32_t outstanding;
16679                         /*
16680                          * We are inside of a Fast recovery episode, this
16681                          * is caused by a SACK or 3 dup acks. At this point
16682                          * we have sent all the retransmissions and we rely
16683                          * on PRR to dictate what we will send in the form of
16684                          * new data.
16685                          */
16686
16687                         outstanding = tp->snd_max - tp->snd_una;
16688                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
16689                                 if (tp->snd_wnd > outstanding) {
16690                                         len = tp->snd_wnd - outstanding;
16691                                         /* Check to see if we have the data */
16692                                         if ((sb_offset + len) > avail) {
16693                                                 /* It does not all fit */
16694                                                 if (avail > sb_offset)
16695                                                         len = avail - sb_offset;
16696                                                 else
16697                                                         len = 0;
16698                                         }
16699                                 } else {
16700                                         len = 0;
16701                                 }
16702                         } else if (avail > sb_offset) {
16703                                 len = avail - sb_offset;
16704                         } else {
16705                                 len = 0;
16706                         }
16707                         if (len > 0) {
16708                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
16709                                         len = rack->r_ctl.rc_prr_sndcnt;
16710                                 }
16711                                 if (len > 0) {
16712                                         sub_from_prr = 1;
16713                                         counter_u64_add(rack_rtm_prr_newdata, 1);
16714                                 }
16715                         }
16716                         if (len > segsiz) {
16717                                 /*
16718                                  * We should never send more than a MSS when
16719                                  * retransmitting or sending new data in prr
16720                                  * mode unless the override flag is on. Most
16721                                  * likely the PRR algorithm is not going to
16722                                  * let us send a lot as well :-)
16723                                  */
16724                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
16725                                         len = segsiz;
16726                                 }
16727                         } else if (len < segsiz) {
16728                                 /*
16729                                  * Do we send any? The idea here is if the
16730                                  * send empty's the socket buffer we want to
16731                                  * do it. However if not then lets just wait
16732                                  * for our prr_sndcnt to get bigger.
16733                                  */
16734                                 long leftinsb;
16735
16736                                 leftinsb = sbavail(sb) - sb_offset;
16737                                 if (leftinsb > len) {
16738                                         /* This send does not empty the sb */
16739                                         len = 0;
16740                                 }
16741                         }
16742                 }
16743         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
16744                 /*
16745                  * If you have not established
16746                  * and are not doing FAST OPEN
16747                  * no data please.
16748                  */
16749                 if ((sack_rxmit == 0) &&
16750                     (!IS_FASTOPEN(tp->t_flags))){
16751                         len = 0;
16752                         sb_offset = 0;
16753                 }
16754         }
16755         if (prefetch_so_done == 0) {
16756                 kern_prefetch(so, &prefetch_so_done);
16757                 prefetch_so_done = 1;
16758         }
16759         /*
16760          * Lop off SYN bit if it has already been sent.  However, if this is
16761          * SYN-SENT state and if segment contains data and if we don't know
16762          * that foreign host supports TAO, suppress sending segment.
16763          */
16764         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
16765             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
16766                 /*
16767                  * When sending additional segments following a TFO SYN|ACK,
16768                  * do not include the SYN bit.
16769                  */
16770                 if (IS_FASTOPEN(tp->t_flags) &&
16771                     (tp->t_state == TCPS_SYN_RECEIVED))
16772                         flags &= ~TH_SYN;
16773         }
16774         /*
16775          * Be careful not to send data and/or FIN on SYN segments. This
16776          * measure is needed to prevent interoperability problems with not
16777          * fully conformant TCP implementations.
16778          */
16779         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
16780                 len = 0;
16781                 flags &= ~TH_FIN;
16782         }
16783         /*
16784          * On TFO sockets, ensure no data is sent in the following cases:
16785          *
16786          *  - When retransmitting SYN|ACK on a passively-created socket
16787          *
16788          *  - When retransmitting SYN on an actively created socket
16789          *
16790          *  - When sending a zero-length cookie (cookie request) on an
16791          *    actively created socket
16792          *
16793          *  - When the socket is in the CLOSED state (RST is being sent)
16794          */
16795         if (IS_FASTOPEN(tp->t_flags) &&
16796             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
16797              ((tp->t_state == TCPS_SYN_SENT) &&
16798               (tp->t_tfo_client_cookie_len == 0)) ||
16799              (flags & TH_RST))) {
16800                 sack_rxmit = 0;
16801                 len = 0;
16802         }
16803         /* Without fast-open there should never be data sent on a SYN */
16804         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
16805                 tp->snd_nxt = tp->iss;
16806                 len = 0;
16807         }
16808         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
16809                 /* We only send 1 MSS if we have a DSACK block */
16810                 add_flag |= RACK_SENT_W_DSACK;
16811                 len = segsiz;
16812         }
16813         orig_len = len;
16814         if (len <= 0) {
16815                 /*
16816                  * If FIN has been sent but not acked, but we haven't been
16817                  * called to retransmit, len will be < 0.  Otherwise, window
16818                  * shrank after we sent into it.  If window shrank to 0,
16819                  * cancel pending retransmit, pull snd_nxt back to (closed)
16820                  * window, and set the persist timer if it isn't already
16821                  * going.  If the window didn't close completely, just wait
16822                  * for an ACK.
16823                  *
16824                  * We also do a general check here to ensure that we will
16825                  * set the persist timer when we have data to send, but a
16826                  * 0-byte window. This makes sure the persist timer is set
16827                  * even if the packet hits one of the "goto send" lines
16828                  * below.
16829                  */
16830                 len = 0;
16831                 if ((tp->snd_wnd == 0) &&
16832                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16833                     (tp->snd_una == tp->snd_max) &&
16834                     (sb_offset < (int)sbavail(sb))) {
16835                         rack_enter_persist(tp, rack, cts);
16836                 }
16837         } else if ((rsm == NULL) &&
16838                    (doing_tlp == 0) &&
16839                    (len < pace_max_seg)) {
16840                 /*
16841                  * We are not sending a maximum sized segment for
16842                  * some reason. Should we not send anything (think
16843                  * sws or persists)?
16844                  */
16845                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16846                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16847                     (len < minseg) &&
16848                     (len < (int)(sbavail(sb) - sb_offset))) {
16849                         /*
16850                          * Here the rwnd is less than
16851                          * the minimum pacing size, this is not a retransmit,
16852                          * we are established and
16853                          * the send is not the last in the socket buffer
16854                          * we send nothing, and we may enter persists
16855                          * if nothing is outstanding.
16856                          */
16857                         len = 0;
16858                         if (tp->snd_max == tp->snd_una) {
16859                                 /*
16860                                  * Nothing out we can
16861                                  * go into persists.
16862                                  */
16863                                 rack_enter_persist(tp, rack, cts);
16864                         }
16865                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
16866                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16867                            (len < (int)(sbavail(sb) - sb_offset)) &&
16868                            (len < minseg)) {
16869                         /*
16870                          * Here we are not retransmitting, and
16871                          * the cwnd is not so small that we could
16872                          * not send at least a min size (rxt timer
16873                          * not having gone off), We have 2 segments or
16874                          * more already in flight, its not the tail end
16875                          * of the socket buffer  and the cwnd is blocking
16876                          * us from sending out a minimum pacing segment size.
16877                          * Lets not send anything.
16878                          */
16879                         len = 0;
16880                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
16881                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16882                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16883                            (len < (int)(sbavail(sb) - sb_offset)) &&
16884                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
16885                         /*
16886                          * Here we have a send window but we have
16887                          * filled it up and we can't send another pacing segment.
16888                          * We also have in flight more than 2 segments
16889                          * and we are not completing the sb i.e. we allow
16890                          * the last bytes of the sb to go out even if
16891                          * its not a full pacing segment.
16892                          */
16893                         len = 0;
16894                 } else if ((rack->r_ctl.crte != NULL) &&
16895                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
16896                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
16897                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
16898                            (len < (int)(sbavail(sb) - sb_offset))) {
16899                         /*
16900                          * Here we are doing hardware pacing, this is not a TLP,
16901                          * we are not sending a pace max segment size, there is rwnd
16902                          * room to send at least N pace_max_seg, the cwnd is greater
16903                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
16904                          * more segments in flight and its not the tail of the socket buffer.
16905                          *
16906                          * We don't want to send instead we need to get more ack's in to
16907                          * allow us to send a full pacing segment. Normally, if we are pacing
16908                          * about the right speed, we should have finished our pacing
16909                          * send as most of the acks have come back if we are at the
16910                          * right rate. This is a bit fuzzy since return path delay
16911                          * can delay the acks, which is why we want to make sure we
16912                          * have cwnd space to have a bit more than a max pace segments in flight.
16913                          *
16914                          * If we have not gotten our acks back we are pacing at too high a
16915                          * rate delaying will not hurt and will bring our GP estimate down by
16916                          * injecting the delay. If we don't do this we will send
16917                          * 2 MSS out in response to the acks being clocked in which
16918                          * defeats the point of hw-pacing (i.e. to help us get
16919                          * larger TSO's out).
16920                          */
16921                         len = 0;
16922
16923                 }
16924
16925         }
16926         /* len will be >= 0 after this point. */
16927         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
16928         rack_sndbuf_autoscale(rack);
16929         /*
16930          * Decide if we can use TCP Segmentation Offloading (if supported by
16931          * hardware).
16932          *
16933          * TSO may only be used if we are in a pure bulk sending state.  The
16934          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
16935          * options prevent using TSO.  With TSO the TCP header is the same
16936          * (except for the sequence number) for all generated packets.  This
16937          * makes it impossible to transmit any options which vary per
16938          * generated segment or packet.
16939          *
16940          * IPv4 handling has a clear separation of ip options and ip header
16941          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
16942          * the right thing below to provide length of just ip options and thus
16943          * checking for ipoptlen is enough to decide if ip options are present.
16944          */
16945         ipoptlen = 0;
16946 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16947         /*
16948          * Pre-calculate here as we save another lookup into the darknesses
16949          * of IPsec that way and can actually decide if TSO is ok.
16950          */
16951 #ifdef INET6
16952         if (isipv6 && IPSEC_ENABLED(ipv6))
16953                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
16954 #ifdef INET
16955         else
16956 #endif
16957 #endif                          /* INET6 */
16958 #ifdef INET
16959                 if (IPSEC_ENABLED(ipv4))
16960                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
16961 #endif                          /* INET */
16962 #endif
16963
16964 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16965         ipoptlen += ipsec_optlen;
16966 #endif
16967         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
16968             (tp->t_port == 0) &&
16969             ((tp->t_flags & TF_SIGNATURE) == 0) &&
16970             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
16971             ipoptlen == 0)
16972                 tso = 1;
16973         {
16974                 uint32_t outstanding;
16975
16976                 outstanding = tp->snd_max - tp->snd_una;
16977                 if (tp->t_flags & TF_SENTFIN) {
16978                         /*
16979                          * If we sent a fin, snd_max is 1 higher than
16980                          * snd_una
16981                          */
16982                         outstanding--;
16983                 }
16984                 if (sack_rxmit) {
16985                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
16986                                 flags &= ~TH_FIN;
16987                 } else {
16988                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
16989                                    sbused(sb)))
16990                                 flags &= ~TH_FIN;
16991                 }
16992         }
16993         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
16994             (long)TCP_MAXWIN << tp->rcv_scale);
16995
16996         /*
16997          * Sender silly window avoidance.   We transmit under the following
16998          * conditions when len is non-zero:
16999          *
17000          * - We have a full segment (or more with TSO) - This is the last
17001          * buffer in a write()/send() and we are either idle or running
17002          * NODELAY - we've timed out (e.g. persist timer) - we have more
17003          * then 1/2 the maximum send window's worth of data (receiver may be
17004          * limited the window size) - we need to retransmit
17005          */
17006         if (len) {
17007                 if (len >= segsiz) {
17008                         goto send;
17009                 }
17010                 /*
17011                  * NOTE! on localhost connections an 'ack' from the remote
17012                  * end may occur synchronously with the output and cause us
17013                  * to flush a buffer queued with moretocome.  XXX
17014                  *
17015                  */
17016                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
17017                     (idle || (tp->t_flags & TF_NODELAY)) &&
17018                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17019                     (tp->t_flags & TF_NOPUSH) == 0) {
17020                         pass = 2;
17021                         goto send;
17022                 }
17023                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
17024                         pass = 22;
17025                         goto send;
17026                 }
17027                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17028                         pass = 4;
17029                         goto send;
17030                 }
17031                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
17032                         pass = 5;
17033                         goto send;
17034                 }
17035                 if (sack_rxmit) {
17036                         pass = 6;
17037                         goto send;
17038                 }
17039                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17040                     (ctf_outstanding(tp) < (segsiz * 2))) {
17041                         /*
17042                          * We have less than two MSS outstanding (delayed ack)
17043                          * and our rwnd will not let us send a full sized
17044                          * MSS. Lets go ahead and let this small segment
17045                          * out because we want to try to have at least two
17046                          * packets inflight to not be caught by delayed ack.
17047                          */
17048                         pass = 12;
17049                         goto send;
17050                 }
17051         }
17052         /*
17053          * Sending of standalone window updates.
17054          *
17055          * Window updates are important when we close our window due to a
17056          * full socket buffer and are opening it again after the application
17057          * reads data from it.  Once the window has opened again and the
17058          * remote end starts to send again the ACK clock takes over and
17059          * provides the most current window information.
17060          *
17061          * We must avoid the silly window syndrome whereas every read from
17062          * the receive buffer, no matter how small, causes a window update
17063          * to be sent.  We also should avoid sending a flurry of window
17064          * updates when the socket buffer had queued a lot of data and the
17065          * application is doing small reads.
17066          *
17067          * Prevent a flurry of pointless window updates by only sending an
17068          * update when we can increase the advertized window by more than
17069          * 1/4th of the socket buffer capacity.  When the buffer is getting
17070          * full or is very small be more aggressive and send an update
17071          * whenever we can increase by two mss sized segments. In all other
17072          * situations the ACK's to new incoming data will carry further
17073          * window increases.
17074          *
17075          * Don't send an independent window update if a delayed ACK is
17076          * pending (it will get piggy-backed on it) or the remote side
17077          * already has done a half-close and won't send more data.  Skip
17078          * this if the connection is in T/TCP half-open state.
17079          */
17080         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17081             !(tp->t_flags & TF_DELACK) &&
17082             !TCPS_HAVERCVDFIN(tp->t_state)) {
17083                 /*
17084                  * "adv" is the amount we could increase the window, taking
17085                  * into account that we are limited by TCP_MAXWIN <<
17086                  * tp->rcv_scale.
17087                  */
17088                 int32_t adv;
17089                 int oldwin;
17090
17091                 adv = recwin;
17092                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17093                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
17094                         if (adv > oldwin)
17095                             adv -= oldwin;
17096                         else {
17097                                 /* We can't increase the window */
17098                                 adv = 0;
17099                         }
17100                 } else
17101                         oldwin = 0;
17102
17103                 /*
17104                  * If the new window size ends up being the same as or less
17105                  * than the old size when it is scaled, then don't force
17106                  * a window update.
17107                  */
17108                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17109                         goto dontupdate;
17110
17111                 if (adv >= (int32_t)(2 * segsiz) &&
17112                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17113                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17114                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17115                         pass = 7;
17116                         goto send;
17117                 }
17118                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17119                         pass = 23;
17120                         goto send;
17121                 }
17122         }
17123 dontupdate:
17124
17125         /*
17126          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17127          * is also a catch-all for the retransmit timer timeout case.
17128          */
17129         if (tp->t_flags & TF_ACKNOW) {
17130                 pass = 8;
17131                 goto send;
17132         }
17133         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17134                 pass = 9;
17135                 goto send;
17136         }
17137         /*
17138          * If our state indicates that FIN should be sent and we have not
17139          * yet done so, then we need to send.
17140          */
17141         if ((flags & TH_FIN) &&
17142             (tp->snd_nxt == tp->snd_una)) {
17143                 pass = 11;
17144                 goto send;
17145         }
17146         /*
17147          * No reason to send a segment, just return.
17148          */
17149 just_return:
17150         SOCKBUF_UNLOCK(sb);
17151 just_return_nolock:
17152         {
17153                 int app_limited = CTF_JR_SENT_DATA;
17154
17155                 if (tot_len_this_send > 0) {
17156                         /* Make sure snd_nxt is up to max */
17157                         rack->r_ctl.fsb.recwin = recwin;
17158                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17159                         if ((error == 0) &&
17160                             rack_use_rfo &&
17161                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
17162                             (ipoptlen == 0) &&
17163                             (tp->snd_nxt == tp->snd_max) &&
17164                             (tp->rcv_numsacks == 0) &&
17165                             rack->r_fsb_inited &&
17166                             TCPS_HAVEESTABLISHED(tp->t_state) &&
17167                             (rack->r_must_retran == 0) &&
17168                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
17169                             (len > 0) && (orig_len > 0) &&
17170                             (orig_len > len) &&
17171                             ((orig_len - len) >= segsiz) &&
17172                             ((optlen == 0) ||
17173                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17174                                 /* We can send at least one more MSS using our fsb */
17175
17176                                 rack->r_fast_output = 1;
17177                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17178                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17179                                 rack->r_ctl.fsb.tcp_flags = flags;
17180                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
17181                                 if (hw_tls)
17182                                         rack->r_ctl.fsb.hw_tls = 1;
17183                                 else
17184                                         rack->r_ctl.fsb.hw_tls = 0;
17185                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17186                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
17187                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17188                                          (tp->snd_max - tp->snd_una)));
17189                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
17190                                         rack->r_fast_output = 0;
17191                                 else {
17192                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17193                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
17194                                         else
17195                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
17196                                 }
17197                         } else
17198                                 rack->r_fast_output = 0;
17199
17200
17201                         rack_log_fsb(rack, tp, so, flags,
17202                                      ipoptlen, orig_len, len, 0,
17203                                      1, optlen, __LINE__, 1);
17204                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17205                                 tp->snd_nxt = tp->snd_max;
17206                 } else {
17207                         int end_window = 0;
17208                         uint32_t seq = tp->gput_ack;
17209
17210                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17211                         if (rsm) {
17212                                 /*
17213                                  * Mark the last sent that we just-returned (hinting
17214                                  * that delayed ack may play a role in any rtt measurement).
17215                                  */
17216                                 rsm->r_just_ret = 1;
17217                         }
17218                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17219                         rack->r_ctl.rc_agg_delayed = 0;
17220                         rack->r_early = 0;
17221                         rack->r_late = 0;
17222                         rack->r_ctl.rc_agg_early = 0;
17223                         if ((ctf_outstanding(tp) +
17224                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17225                                  minseg)) >= tp->snd_wnd) {
17226                                 /* We are limited by the rwnd */
17227                                 app_limited = CTF_JR_RWND_LIMITED;
17228                                 if (IN_FASTRECOVERY(tp->t_flags))
17229                                     rack->r_ctl.rc_prr_sndcnt = 0;
17230                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17231                                 /* We are limited by whats available -- app limited */
17232                                 app_limited = CTF_JR_APP_LIMITED;
17233                                 if (IN_FASTRECOVERY(tp->t_flags))
17234                                     rack->r_ctl.rc_prr_sndcnt = 0;
17235                         } else if ((idle == 0) &&
17236                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17237                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17238                                    (len < segsiz)) {
17239                                 /*
17240                                  * No delay is not on and the
17241                                  * user is sending less than 1MSS. This
17242                                  * brings out SWS avoidance so we
17243                                  * don't send. Another app-limited case.
17244                                  */
17245                                 app_limited = CTF_JR_APP_LIMITED;
17246                         } else if (tp->t_flags & TF_NOPUSH) {
17247                                 /*
17248                                  * The user has requested no push of
17249                                  * the last segment and we are
17250                                  * at the last segment. Another app
17251                                  * limited case.
17252                                  */
17253                                 app_limited = CTF_JR_APP_LIMITED;
17254                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17255                                 /* Its the cwnd */
17256                                 app_limited = CTF_JR_CWND_LIMITED;
17257                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17258                                    (rack->rack_no_prr == 0) &&
17259                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17260                                 app_limited = CTF_JR_PRR;
17261                         } else {
17262                                 /* Now why here are we not sending? */
17263 #ifdef NOW
17264 #ifdef INVARIANTS
17265                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17266 #endif
17267 #endif
17268                                 app_limited = CTF_JR_ASSESSING;
17269                         }
17270                         /*
17271                          * App limited in some fashion, for our pacing GP
17272                          * measurements we don't want any gap (even cwnd).
17273                          * Close  down the measurement window.
17274                          */
17275                         if (rack_cwnd_block_ends_measure &&
17276                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17277                              (app_limited == CTF_JR_PRR))) {
17278                                 /*
17279                                  * The reason we are not sending is
17280                                  * the cwnd (or prr). We have been configured
17281                                  * to end the measurement window in
17282                                  * this case.
17283                                  */
17284                                 end_window = 1;
17285                         } else if (rack_rwnd_block_ends_measure &&
17286                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17287                                 /*
17288                                  * We are rwnd limited and have been
17289                                  * configured to end the measurement
17290                                  * window in this case.
17291                                  */
17292                                 end_window = 1;
17293                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17294                                 /*
17295                                  * A true application limited period, we have
17296                                  * ran out of data.
17297                                  */
17298                                 end_window = 1;
17299                         } else if (app_limited == CTF_JR_ASSESSING) {
17300                                 /*
17301                                  * In the assessing case we hit the end of
17302                                  * the if/else and had no known reason
17303                                  * This will panic us under invariants..
17304                                  *
17305                                  * If we get this out in logs we need to
17306                                  * investagate which reason we missed.
17307                                  */
17308                                 end_window = 1;
17309                         }
17310                         if (end_window) {
17311                                 uint8_t log = 0;
17312
17313                                 /* Adjust the Gput measurement */
17314                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17315                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17316                                         tp->gput_ack = tp->snd_max;
17317                                         if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17318                                                 /*
17319                                                  * There is not enough to measure.
17320                                                  */
17321                                                 tp->t_flags &= ~TF_GPUTINPROG;
17322                                                 rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17323                                                                            rack->r_ctl.rc_gp_srtt /*flex1*/,
17324                                                                            tp->gput_seq,
17325                                                                            0, 0, 18, __LINE__, NULL, 0);
17326                                         } else
17327                                                 log = 1;
17328                                 }
17329                                 /* Mark the last packet has app limited */
17330                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17331                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17332                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17333                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17334                                         else {
17335                                                 /*
17336                                                  * Go out to the end app limited and mark
17337                                                  * this new one as next and move the end_appl up
17338                                                  * to this guy.
17339                                                  */
17340                                                 if (rack->r_ctl.rc_end_appl)
17341                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17342                                                 rack->r_ctl.rc_end_appl = rsm;
17343                                         }
17344                                         rsm->r_flags |= RACK_APP_LIMITED;
17345                                         rack->r_ctl.rc_app_limited_cnt++;
17346                                 }
17347                                 if (log)
17348                                         rack_log_pacing_delay_calc(rack,
17349                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17350                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17351                         }
17352                 }
17353                 if (slot) {
17354                         /* set the rack tcb into the slot N */
17355                         counter_u64_add(rack_paced_segments, 1);
17356                 } else if (tot_len_this_send) {
17357                         counter_u64_add(rack_unpaced_segments, 1);
17358                 }
17359                 /* Check if we need to go into persists or not */
17360                 if ((tp->snd_max == tp->snd_una) &&
17361                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17362                     sbavail(sb) &&
17363                     (sbavail(sb) > tp->snd_wnd) &&
17364                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17365                         /* Yes lets make sure to move to persist before timer-start */
17366                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17367                 }
17368                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17369                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17370         }
17371 #ifdef NETFLIX_SHARED_CWND
17372         if ((sbavail(sb) == 0) &&
17373             rack->r_ctl.rc_scw) {
17374                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17375                 rack->rack_scwnd_is_idle = 1;
17376         }
17377 #endif
17378 #ifdef TCP_ACCOUNTING
17379         if (tot_len_this_send > 0) {
17380                 crtsc = get_cyclecount();
17381                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17382                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
17383                 }
17384                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17385                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17386                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17387                 }
17388                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17389                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17390                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17391                 }
17392                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17393         } else {
17394                 crtsc = get_cyclecount();
17395                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17396                         tp->tcp_cnt_counters[SND_LIMITED]++;
17397                 }
17398                 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17399                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17400                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17401                 }
17402                 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17403         }
17404         sched_unpin();
17405 #endif
17406         return (0);
17407
17408 send:
17409         if (rsm || sack_rxmit)
17410                 counter_u64_add(rack_nfto_resend, 1);
17411         else
17412                 counter_u64_add(rack_non_fto_send, 1);
17413         if ((flags & TH_FIN) &&
17414             sbavail(sb)) {
17415                 /*
17416                  * We do not transmit a FIN
17417                  * with data outstanding. We
17418                  * need to make it so all data
17419                  * is acked first.
17420                  */
17421                 flags &= ~TH_FIN;
17422         }
17423         /* Enforce stack imposed max seg size if we have one */
17424         if (rack->r_ctl.rc_pace_max_segs &&
17425             (len > rack->r_ctl.rc_pace_max_segs)) {
17426                 mark = 1;
17427                 len = rack->r_ctl.rc_pace_max_segs;
17428         }
17429         SOCKBUF_LOCK_ASSERT(sb);
17430         if (len > 0) {
17431                 if (len >= segsiz)
17432                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17433                 else
17434                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17435         }
17436         /*
17437          * Before ESTABLISHED, force sending of initial options unless TCP
17438          * set not to do any options. NOTE: we assume that the IP/TCP header
17439          * plus TCP options always fit in a single mbuf, leaving room for a
17440          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17441          * + optlen <= MCLBYTES
17442          */
17443         optlen = 0;
17444 #ifdef INET6
17445         if (isipv6)
17446                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17447         else
17448 #endif
17449                 hdrlen = sizeof(struct tcpiphdr);
17450
17451         /*
17452          * Compute options for segment. We only have to care about SYN and
17453          * established connection segments.  Options for SYN-ACK segments
17454          * are handled in TCP syncache.
17455          */
17456         to.to_flags = 0;
17457         if ((tp->t_flags & TF_NOOPT) == 0) {
17458                 /* Maximum segment size. */
17459                 if (flags & TH_SYN) {
17460                         tp->snd_nxt = tp->iss;
17461                         to.to_mss = tcp_mssopt(&inp->inp_inc);
17462                         if (tp->t_port)
17463                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
17464                         to.to_flags |= TOF_MSS;
17465
17466                         /*
17467                          * On SYN or SYN|ACK transmits on TFO connections,
17468                          * only include the TFO option if it is not a
17469                          * retransmit, as the presence of the TFO option may
17470                          * have caused the original SYN or SYN|ACK to have
17471                          * been dropped by a middlebox.
17472                          */
17473                         if (IS_FASTOPEN(tp->t_flags) &&
17474                             (tp->t_rxtshift == 0)) {
17475                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
17476                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17477                                         to.to_tfo_cookie =
17478                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
17479                                         to.to_flags |= TOF_FASTOPEN;
17480                                         wanted_cookie = 1;
17481                                 } else if (tp->t_state == TCPS_SYN_SENT) {
17482                                         to.to_tfo_len =
17483                                                 tp->t_tfo_client_cookie_len;
17484                                         to.to_tfo_cookie =
17485                                                 tp->t_tfo_cookie.client;
17486                                         to.to_flags |= TOF_FASTOPEN;
17487                                         wanted_cookie = 1;
17488                                         /*
17489                                          * If we wind up having more data to
17490                                          * send with the SYN than can fit in
17491                                          * one segment, don't send any more
17492                                          * until the SYN|ACK comes back from
17493                                          * the other end.
17494                                          */
17495                                         sendalot = 0;
17496                                 }
17497                         }
17498                 }
17499                 /* Window scaling. */
17500                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17501                         to.to_wscale = tp->request_r_scale;
17502                         to.to_flags |= TOF_SCALE;
17503                 }
17504                 /* Timestamps. */
17505                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
17506                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17507                         to.to_tsval = ms_cts + tp->ts_offset;
17508                         to.to_tsecr = tp->ts_recent;
17509                         to.to_flags |= TOF_TS;
17510                 }
17511                 /* Set receive buffer autosizing timestamp. */
17512                 if (tp->rfbuf_ts == 0 &&
17513                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
17514                         tp->rfbuf_ts = tcp_ts_getticks();
17515                 /* Selective ACK's. */
17516                 if (tp->t_flags & TF_SACK_PERMIT) {
17517                         if (flags & TH_SYN)
17518                                 to.to_flags |= TOF_SACKPERM;
17519                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17520                                  tp->rcv_numsacks > 0) {
17521                                 to.to_flags |= TOF_SACK;
17522                                 to.to_nsacks = tp->rcv_numsacks;
17523                                 to.to_sacks = (u_char *)tp->sackblks;
17524                         }
17525                 }
17526 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17527                 /* TCP-MD5 (RFC2385). */
17528                 if (tp->t_flags & TF_SIGNATURE)
17529                         to.to_flags |= TOF_SIGNATURE;
17530 #endif                          /* TCP_SIGNATURE */
17531
17532                 /* Processing the options. */
17533                 hdrlen += optlen = tcp_addoptions(&to, opt);
17534                 /*
17535                  * If we wanted a TFO option to be added, but it was unable
17536                  * to fit, ensure no data is sent.
17537                  */
17538                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17539                     !(to.to_flags & TOF_FASTOPEN))
17540                         len = 0;
17541         }
17542         if (tp->t_port) {
17543                 if (V_tcp_udp_tunneling_port == 0) {
17544                         /* The port was removed?? */
17545                         SOCKBUF_UNLOCK(&so->so_snd);
17546 #ifdef TCP_ACCOUNTING
17547                         crtsc = get_cyclecount();
17548                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17549                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17550                         }
17551                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
17552                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17553                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17554                         }
17555                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
17556                         sched_unpin();
17557 #endif
17558                         return (EHOSTUNREACH);
17559                 }
17560                 hdrlen += sizeof(struct udphdr);
17561         }
17562 #ifdef INET6
17563         if (isipv6)
17564                 ipoptlen = ip6_optlen(tp->t_inpcb);
17565         else
17566 #endif
17567                 if (tp->t_inpcb->inp_options)
17568                         ipoptlen = tp->t_inpcb->inp_options->m_len -
17569                                 offsetof(struct ipoption, ipopt_list);
17570                 else
17571                         ipoptlen = 0;
17572 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17573         ipoptlen += ipsec_optlen;
17574 #endif
17575
17576         /*
17577          * Adjust data length if insertion of options will bump the packet
17578          * length beyond the t_maxseg length. Clear the FIN bit because we
17579          * cut off the tail of the segment.
17580          */
17581         if (len + optlen + ipoptlen > tp->t_maxseg) {
17582                 if (tso) {
17583                         uint32_t if_hw_tsomax;
17584                         uint32_t moff;
17585                         int32_t max_len;
17586
17587                         /* extract TSO information */
17588                         if_hw_tsomax = tp->t_tsomax;
17589                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
17590                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
17591                         KASSERT(ipoptlen == 0,
17592                                 ("%s: TSO can't do IP options", __func__));
17593
17594                         /*
17595                          * Check if we should limit by maximum payload
17596                          * length:
17597                          */
17598                         if (if_hw_tsomax != 0) {
17599                                 /* compute maximum TSO length */
17600                                 max_len = (if_hw_tsomax - hdrlen -
17601                                            max_linkhdr);
17602                                 if (max_len <= 0) {
17603                                         len = 0;
17604                                 } else if (len > max_len) {
17605                                         sendalot = 1;
17606                                         len = max_len;
17607                                         mark = 2;
17608                                 }
17609                         }
17610                         /*
17611                          * Prevent the last segment from being fractional
17612                          * unless the send sockbuf can be emptied:
17613                          */
17614                         max_len = (tp->t_maxseg - optlen);
17615                         if ((sb_offset + len) < sbavail(sb)) {
17616                                 moff = len % (u_int)max_len;
17617                                 if (moff != 0) {
17618                                         mark = 3;
17619                                         len -= moff;
17620                                 }
17621                         }
17622                         /*
17623                          * In case there are too many small fragments don't
17624                          * use TSO:
17625                          */
17626                         if (len <= segsiz) {
17627                                 mark = 4;
17628                                 tso = 0;
17629                         }
17630                         /*
17631                          * Send the FIN in a separate segment after the bulk
17632                          * sending is done. We don't trust the TSO
17633                          * implementations to clear the FIN flag on all but
17634                          * the last segment.
17635                          */
17636                         if (tp->t_flags & TF_NEEDFIN) {
17637                                 sendalot = 4;
17638                         }
17639                 } else {
17640                         mark = 5;
17641                         if (optlen + ipoptlen >= tp->t_maxseg) {
17642                                 /*
17643                                  * Since we don't have enough space to put
17644                                  * the IP header chain and the TCP header in
17645                                  * one packet as required by RFC 7112, don't
17646                                  * send it. Also ensure that at least one
17647                                  * byte of the payload can be put into the
17648                                  * TCP segment.
17649                                  */
17650                                 SOCKBUF_UNLOCK(&so->so_snd);
17651                                 error = EMSGSIZE;
17652                                 sack_rxmit = 0;
17653                                 goto out;
17654                         }
17655                         len = tp->t_maxseg - optlen - ipoptlen;
17656                         sendalot = 5;
17657                 }
17658         } else {
17659                 tso = 0;
17660                 mark = 6;
17661         }
17662         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
17663                 ("%s: len > IP_MAXPACKET", __func__));
17664 #ifdef DIAGNOSTIC
17665 #ifdef INET6
17666         if (max_linkhdr + hdrlen > MCLBYTES)
17667 #else
17668                 if (max_linkhdr + hdrlen > MHLEN)
17669 #endif
17670                         panic("tcphdr too big");
17671 #endif
17672
17673         /*
17674          * This KASSERT is here to catch edge cases at a well defined place.
17675          * Before, those had triggered (random) panic conditions further
17676          * down.
17677          */
17678         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17679         if ((len == 0) &&
17680             (flags & TH_FIN) &&
17681             (sbused(sb))) {
17682                 /*
17683                  * We have outstanding data, don't send a fin by itself!.
17684                  */
17685                 goto just_return;
17686         }
17687         /*
17688          * Grab a header mbuf, attaching a copy of data to be transmitted,
17689          * and initialize the header from the template for sends on this
17690          * connection.
17691          */
17692         hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
17693         if (len) {
17694                 uint32_t max_val;
17695                 uint32_t moff;
17696
17697                 if (rack->r_ctl.rc_pace_max_segs)
17698                         max_val = rack->r_ctl.rc_pace_max_segs;
17699                 else if (rack->rc_user_set_max_segs)
17700                         max_val = rack->rc_user_set_max_segs * segsiz;
17701                 else
17702                         max_val = len;
17703                 /*
17704                  * We allow a limit on sending with hptsi.
17705                  */
17706                 if (len > max_val) {
17707                         mark = 7;
17708                         len = max_val;
17709                 }
17710 #ifdef INET6
17711                 if (MHLEN < hdrlen + max_linkhdr)
17712                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
17713                 else
17714 #endif
17715                         m = m_gethdr(M_NOWAIT, MT_DATA);
17716
17717                 if (m == NULL) {
17718                         SOCKBUF_UNLOCK(sb);
17719                         error = ENOBUFS;
17720                         sack_rxmit = 0;
17721                         goto out;
17722                 }
17723                 m->m_data += max_linkhdr;
17724                 m->m_len = hdrlen;
17725
17726                 /*
17727                  * Start the m_copy functions from the closest mbuf to the
17728                  * sb_offset in the socket buffer chain.
17729                  */
17730                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
17731                 s_mb = mb;
17732                 s_moff = moff;
17733                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
17734                         m_copydata(mb, moff, (int)len,
17735                                    mtod(m, caddr_t)+hdrlen);
17736                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17737                                 sbsndptr_adv(sb, mb, len);
17738                         m->m_len += len;
17739                 } else {
17740                         struct sockbuf *msb;
17741
17742                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17743                                 msb = NULL;
17744                         else
17745                                 msb = sb;
17746                         m->m_next = tcp_m_copym(
17747                                 mb, moff, &len,
17748                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
17749                                 ((rsm == NULL) ? hw_tls : 0)
17750 #ifdef NETFLIX_COPY_ARGS
17751                                 , &filled_all
17752 #endif
17753                                 );
17754                         if (len <= (tp->t_maxseg - optlen)) {
17755                                 /*
17756                                  * Must have ran out of mbufs for the copy
17757                                  * shorten it to no longer need tso. Lets
17758                                  * not put on sendalot since we are low on
17759                                  * mbufs.
17760                                  */
17761                                 tso = 0;
17762                         }
17763                         if (m->m_next == NULL) {
17764                                 SOCKBUF_UNLOCK(sb);
17765                                 (void)m_free(m);
17766                                 error = ENOBUFS;
17767                                 sack_rxmit = 0;
17768                                 goto out;
17769                         }
17770                 }
17771                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
17772                         if (rsm && (rsm->r_flags & RACK_TLP)) {
17773                                 /*
17774                                  * TLP should not count in retran count, but
17775                                  * in its own bin
17776                                  */
17777                                 counter_u64_add(rack_tlp_retran, 1);
17778                                 counter_u64_add(rack_tlp_retran_bytes, len);
17779                         } else {
17780                                 tp->t_sndrexmitpack++;
17781                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
17782                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
17783                         }
17784 #ifdef STATS
17785                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
17786                                                  len);
17787 #endif
17788                 } else {
17789                         KMOD_TCPSTAT_INC(tcps_sndpack);
17790                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
17791 #ifdef STATS
17792                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
17793                                                  len);
17794 #endif
17795                 }
17796                 /*
17797                  * If we're sending everything we've got, set PUSH. (This
17798                  * will keep happy those implementations which only give
17799                  * data to the user when a buffer fills or a PUSH comes in.)
17800                  */
17801                 if (sb_offset + len == sbused(sb) &&
17802                     sbused(sb) &&
17803                     !(flags & TH_SYN)) {
17804                         flags |= TH_PUSH;
17805                         add_flag |= RACK_HAD_PUSH;
17806                 }
17807
17808                 SOCKBUF_UNLOCK(sb);
17809         } else {
17810                 SOCKBUF_UNLOCK(sb);
17811                 if (tp->t_flags & TF_ACKNOW)
17812                         KMOD_TCPSTAT_INC(tcps_sndacks);
17813                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
17814                         KMOD_TCPSTAT_INC(tcps_sndctrl);
17815                 else
17816                         KMOD_TCPSTAT_INC(tcps_sndwinup);
17817
17818                 m = m_gethdr(M_NOWAIT, MT_DATA);
17819                 if (m == NULL) {
17820                         error = ENOBUFS;
17821                         sack_rxmit = 0;
17822                         goto out;
17823                 }
17824 #ifdef INET6
17825                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
17826                     MHLEN >= hdrlen) {
17827                         M_ALIGN(m, hdrlen);
17828                 } else
17829 #endif
17830                         m->m_data += max_linkhdr;
17831                 m->m_len = hdrlen;
17832         }
17833         SOCKBUF_UNLOCK_ASSERT(sb);
17834         m->m_pkthdr.rcvif = (struct ifnet *)0;
17835 #ifdef MAC
17836         mac_inpcb_create_mbuf(inp, m);
17837 #endif
17838         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
17839 #ifdef INET6
17840                 if (isipv6) {
17841                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
17842                 } else
17843 #endif                          /* INET6 */
17844                 {
17845 #ifdef INET
17846                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
17847 #endif
17848                 }
17849                 th = rack->r_ctl.fsb.th;
17850                 udp = rack->r_ctl.fsb.udp;
17851                 if (udp) {
17852 #ifdef INET6
17853                         if (isipv6) {
17854                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17855                         } else
17856 #endif                          /* INET6 */
17857                         {
17858 #ifdef INET
17859                                 ulen = hdrlen + len - sizeof(struct ip);
17860                                 udp->uh_ulen = htons(ulen);
17861 #endif
17862                         }
17863                 }
17864         } else {
17865 #ifdef INET6
17866                 if (isipv6) {
17867                         ip6 = mtod(m, struct ip6_hdr *);
17868                         if (tp->t_port) {
17869                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
17870                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17871                                 udp->uh_dport = tp->t_port;
17872                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17873                                 udp->uh_ulen = htons(ulen);
17874                                 th = (struct tcphdr *)(udp + 1);
17875                         } else
17876                                 th = (struct tcphdr *)(ip6 + 1);
17877                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
17878                 } else
17879 #endif                          /* INET6 */
17880                 {
17881 #ifdef INET
17882                         ip = mtod(m, struct ip *);
17883 #ifdef TCPDEBUG
17884                         ipov = (struct ipovly *)ip;
17885 #endif
17886                         if (tp->t_port) {
17887                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
17888                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17889                                 udp->uh_dport = tp->t_port;
17890                                 ulen = hdrlen + len - sizeof(struct ip);
17891                                 udp->uh_ulen = htons(ulen);
17892                                 th = (struct tcphdr *)(udp + 1);
17893                         } else
17894                                 th = (struct tcphdr *)(ip + 1);
17895                         tcpip_fillheaders(inp, tp->t_port, ip, th);
17896 #endif                          /* INET */
17897                 }
17898         }
17899         /*
17900          * Fill in fields, remembering maximum advertised window for use in
17901          * delaying messages about window sizes. If resending a FIN, be sure
17902          * not to use a new sequence number.
17903          */
17904         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
17905             tp->snd_nxt == tp->snd_max)
17906                 tp->snd_nxt--;
17907         /*
17908          * If we are starting a connection, send ECN setup SYN packet. If we
17909          * are on a retransmit, we may resend those bits a number of times
17910          * as per RFC 3168.
17911          */
17912         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
17913                 if (tp->t_rxtshift >= 1) {
17914                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
17915                                 flags |= TH_ECE | TH_CWR;
17916                 } else
17917                         flags |= TH_ECE | TH_CWR;
17918         }
17919         /* Handle parallel SYN for ECN */
17920         if ((tp->t_state == TCPS_SYN_RECEIVED) &&
17921             (tp->t_flags2 & TF2_ECN_SND_ECE)) {
17922                 flags |= TH_ECE;
17923                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
17924         }
17925         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17926             (tp->t_flags2 & TF2_ECN_PERMIT)) {
17927                 /*
17928                  * If the peer has ECN, mark data packets with ECN capable
17929                  * transmission (ECT). Ignore pure ack packets,
17930                  * retransmissions.
17931                  */
17932                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
17933                     (sack_rxmit == 0)) {
17934 #ifdef INET6
17935                         if (isipv6) {
17936                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
17937                         } else
17938 #endif
17939                         {
17940 #ifdef INET
17941                                 ip->ip_tos |= IPTOS_ECN_ECT0;
17942 #endif
17943                         }
17944                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
17945                         /*
17946                          * Reply with proper ECN notifications.
17947                          * Only set CWR on new data segments.
17948                          */
17949                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
17950                                 flags |= TH_CWR;
17951                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
17952                         }
17953                 }
17954                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
17955                         flags |= TH_ECE;
17956         }
17957         /*
17958          * If we are doing retransmissions, then snd_nxt will not reflect
17959          * the first unsent octet.  For ACK only packets, we do not want the
17960          * sequence number of the retransmitted packet, we want the sequence
17961          * number of the next unsent octet.  So, if there is no data (and no
17962          * SYN or FIN), use snd_max instead of snd_nxt when filling in
17963          * ti_seq.  But if we are in persist state, snd_max might reflect
17964          * one byte beyond the right edge of the window, so use snd_nxt in
17965          * that case, since we know we aren't doing a retransmission.
17966          * (retransmit and persist are mutually exclusive...)
17967          */
17968         if (sack_rxmit == 0) {
17969                 if (len || (flags & (TH_SYN | TH_FIN))) {
17970                         th->th_seq = htonl(tp->snd_nxt);
17971                         rack_seq = tp->snd_nxt;
17972                 } else {
17973                         th->th_seq = htonl(tp->snd_max);
17974                         rack_seq = tp->snd_max;
17975                 }
17976         } else {
17977                 th->th_seq = htonl(rsm->r_start);
17978                 rack_seq = rsm->r_start;
17979         }
17980         th->th_ack = htonl(tp->rcv_nxt);
17981         th->th_flags = flags;
17982         /*
17983          * Calculate receive window.  Don't shrink window, but avoid silly
17984          * window syndrome.
17985          * If a RST segment is sent, advertise a window of zero.
17986          */
17987         if (flags & TH_RST) {
17988                 recwin = 0;
17989         } else {
17990                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
17991                     recwin < (long)segsiz) {
17992                         recwin = 0;
17993                 }
17994                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
17995                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
17996                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
17997         }
17998
17999         /*
18000          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18001          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18002          * handled in syncache.
18003          */
18004         if (flags & TH_SYN)
18005                 th->th_win = htons((u_short)
18006                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18007         else {
18008                 /* Avoid shrinking window with window scaling. */
18009                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
18010                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18011         }
18012         /*
18013          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18014          * window.  This may cause the remote transmitter to stall.  This
18015          * flag tells soreceive() to disable delayed acknowledgements when
18016          * draining the buffer.  This can occur if the receiver is
18017          * attempting to read more data than can be buffered prior to
18018          * transmitting on the connection.
18019          */
18020         if (th->th_win == 0) {
18021                 tp->t_sndzerowin++;
18022                 tp->t_flags |= TF_RXWIN0SENT;
18023         } else
18024                 tp->t_flags &= ~TF_RXWIN0SENT;
18025         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
18026         /* Now are we using fsb?, if so copy the template data to the mbuf */
18027         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18028                 uint8_t *cpto;
18029
18030                 cpto = mtod(m, uint8_t *);
18031                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18032                 /*
18033                  * We have just copied in:
18034                  * IP/IP6
18035                  * <optional udphdr>
18036                  * tcphdr (no options)
18037                  *
18038                  * We need to grab the correct pointers into the mbuf
18039                  * for both the tcp header, and possibly the udp header (if tunneling).
18040                  * We do this by using the offset in the copy buffer and adding it
18041                  * to the mbuf base pointer (cpto).
18042                  */
18043 #ifdef INET6
18044                 if (isipv6)
18045                         ip6 = mtod(m, struct ip6_hdr *);
18046                 else
18047 #endif                          /* INET6 */
18048 #ifdef INET
18049                         ip = mtod(m, struct ip *);
18050 #endif                          /* INET */
18051                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18052                 /* If we have a udp header lets set it into the mbuf as well */
18053                 if (udp)
18054                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18055         }
18056 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18057         if (to.to_flags & TOF_SIGNATURE) {
18058                 /*
18059                  * Calculate MD5 signature and put it into the place
18060                  * determined before.
18061                  * NOTE: since TCP options buffer doesn't point into
18062                  * mbuf's data, calculate offset and use it.
18063                  */
18064                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18065                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18066                         /*
18067                          * Do not send segment if the calculation of MD5
18068                          * digest has failed.
18069                          */
18070                         goto out;
18071                 }
18072         }
18073 #endif
18074         if (optlen) {
18075                 bcopy(opt, th + 1, optlen);
18076                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18077         }
18078         /*
18079          * Put TCP length in extended header, and then checksum extended
18080          * header and data.
18081          */
18082         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
18083 #ifdef INET6
18084         if (isipv6) {
18085                 /*
18086                  * ip6_plen is not need to be filled now, and will be filled
18087                  * in ip6_output.
18088                  */
18089                 if (tp->t_port) {
18090                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18091                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18092                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18093                         th->th_sum = htons(0);
18094                         UDPSTAT_INC(udps_opackets);
18095                 } else {
18096                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18097                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18098                         th->th_sum = in6_cksum_pseudo(ip6,
18099                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18100                                                       0);
18101                 }
18102         }
18103 #endif
18104 #if defined(INET6) && defined(INET)
18105         else
18106 #endif
18107 #ifdef INET
18108         {
18109                 if (tp->t_port) {
18110                         m->m_pkthdr.csum_flags = CSUM_UDP;
18111                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18112                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18113                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18114                         th->th_sum = htons(0);
18115                         UDPSTAT_INC(udps_opackets);
18116                 } else {
18117                         m->m_pkthdr.csum_flags = CSUM_TCP;
18118                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18119                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
18120                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18121                                                                         IPPROTO_TCP + len + optlen));
18122                 }
18123                 /* IP version must be set here for ipv4/ipv6 checking later */
18124                 KASSERT(ip->ip_v == IPVERSION,
18125                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
18126         }
18127 #endif
18128         /*
18129          * Enable TSO and specify the size of the segments. The TCP pseudo
18130          * header checksum is always provided. XXX: Fixme: This is currently
18131          * not the case for IPv6.
18132          */
18133         if (tso) {
18134                 KASSERT(len > tp->t_maxseg - optlen,
18135                         ("%s: len <= tso_segsz", __func__));
18136                 m->m_pkthdr.csum_flags |= CSUM_TSO;
18137                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18138         }
18139         KASSERT(len + hdrlen == m_length(m, NULL),
18140                 ("%s: mbuf chain different than expected: %d + %u != %u",
18141                  __func__, len, hdrlen, m_length(m, NULL)));
18142
18143 #ifdef TCP_HHOOK
18144         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18145         hhook_run_tcp_est_out(tp, th, &to, len, tso);
18146 #endif
18147         /* We're getting ready to send; log now. */
18148         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18149                 union tcp_log_stackspecific log;
18150
18151                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18152                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
18153                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
18154                 if (rack->rack_no_prr)
18155                         log.u_bbr.flex1 = 0;
18156                 else
18157                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18158                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18159                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18160                 log.u_bbr.flex4 = orig_len;
18161                 if (filled_all)
18162                         log.u_bbr.flex5 = 0x80000000;
18163                 else
18164                         log.u_bbr.flex5 = 0;
18165                 /* Save off the early/late values */
18166                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18167                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18168                 log.u_bbr.bw_inuse = rack_get_bw(rack);
18169                 if (rsm || sack_rxmit) {
18170                         if (doing_tlp)
18171                                 log.u_bbr.flex8 = 2;
18172                         else
18173                                 log.u_bbr.flex8 = 1;
18174                 } else {
18175                         log.u_bbr.flex8 = 0;
18176                 }
18177                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18178                 log.u_bbr.flex7 = mark;
18179                 log.u_bbr.flex7 <<= 8;
18180                 log.u_bbr.flex7 |= pass;
18181                 log.u_bbr.pkts_out = tp->t_maxseg;
18182                 log.u_bbr.timeStamp = cts;
18183                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18184                 log.u_bbr.lt_epoch = cwnd_to_use;
18185                 log.u_bbr.delivered = sendalot;
18186                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18187                                      len, &log, false, NULL, NULL, 0, &tv);
18188         } else
18189                 lgb = NULL;
18190
18191         /*
18192          * Fill in IP length and desired time to live and send to IP level.
18193          * There should be a better way to handle ttl and tos; we could keep
18194          * them in the template, but need a way to checksum without them.
18195          */
18196         /*
18197          * m->m_pkthdr.len should have been set before cksum calcuration,
18198          * because in6_cksum() need it.
18199          */
18200 #ifdef INET6
18201         if (isipv6) {
18202                 /*
18203                  * we separately set hoplimit for every segment, since the
18204                  * user might want to change the value via setsockopt. Also,
18205                  * desired default hop limit might be changed via Neighbor
18206                  * Discovery.
18207                  */
18208                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18209
18210                 /*
18211                  * Set the packet size here for the benefit of DTrace
18212                  * probes. ip6_output() will set it properly; it's supposed
18213                  * to include the option header lengths as well.
18214                  */
18215                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18216
18217                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18218                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18219                 else
18220                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18221
18222                 if (tp->t_state == TCPS_SYN_SENT)
18223                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18224
18225                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18226                 /* TODO: IPv6 IP6TOS_ECT bit on */
18227                 error = ip6_output(m,
18228 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18229                                    inp->in6p_outputopts,
18230 #else
18231                                    NULL,
18232 #endif
18233                                    &inp->inp_route6,
18234                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18235                                    NULL, NULL, inp);
18236
18237                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18238                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18239         }
18240 #endif                          /* INET6 */
18241 #if defined(INET) && defined(INET6)
18242         else
18243 #endif
18244 #ifdef INET
18245         {
18246                 ip->ip_len = htons(m->m_pkthdr.len);
18247 #ifdef INET6
18248                 if (inp->inp_vflag & INP_IPV6PROTO)
18249                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18250 #endif                          /* INET6 */
18251                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18252                 /*
18253                  * If we do path MTU discovery, then we set DF on every
18254                  * packet. This might not be the best thing to do according
18255                  * to RFC3390 Section 2. However the tcp hostcache migitates
18256                  * the problem so it affects only the first tcp connection
18257                  * with a host.
18258                  *
18259                  * NB: Don't set DF on small MTU/MSS to have a safe
18260                  * fallback.
18261                  */
18262                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18263                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18264                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18265                                 ip->ip_off |= htons(IP_DF);
18266                         }
18267                 } else {
18268                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18269                 }
18270
18271                 if (tp->t_state == TCPS_SYN_SENT)
18272                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18273
18274                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18275
18276                 error = ip_output(m,
18277 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18278                                   inp->inp_options,
18279 #else
18280                                   NULL,
18281 #endif
18282                                   &inp->inp_route,
18283                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18284                                   inp);
18285                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18286                         mtu = inp->inp_route.ro_nh->nh_mtu;
18287         }
18288 #endif                          /* INET */
18289
18290 out:
18291         if (lgb) {
18292                 lgb->tlb_errno = error;
18293                 lgb = NULL;
18294         }
18295         /*
18296          * In transmit state, time the transmission and arrange for the
18297          * retransmit.  In persist state, just set snd_max.
18298          */
18299         if (error == 0) {
18300                 tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18301                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18302                 if (rsm && (doing_tlp == 0)) {
18303                         /* Set we retransmitted */
18304                         rack->rc_gp_saw_rec = 1;
18305                 } else {
18306                         if (cwnd_to_use > tp->snd_ssthresh) {
18307                                 /* Set we sent in CA */
18308                                 rack->rc_gp_saw_ca = 1;
18309                         } else {
18310                                 /* Set we sent in SS */
18311                                 rack->rc_gp_saw_ss = 1;
18312                         }
18313                 }
18314                 if (doing_tlp && (rsm == NULL)) {
18315                         /* Make sure new data TLP cnt is clear */
18316                         rack->r_ctl.rc_tlp_new_data = 0;
18317                 }
18318                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18319                     (tp->t_flags & TF_SACK_PERMIT) &&
18320                     tp->rcv_numsacks > 0)
18321                         tcp_clean_dsack_blocks(tp);
18322                 tot_len_this_send += len;
18323                 if (len == 0)
18324                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18325                 else if (len == 1) {
18326                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18327                 } else if (len > 1) {
18328                         int idx;
18329
18330                         idx = (len / segsiz) + 3;
18331                         if (idx >= TCP_MSS_ACCT_ATIMER)
18332                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18333                         else
18334                                 counter_u64_add(rack_out_size[idx], 1);
18335                 }
18336         }
18337         if ((rack->rack_no_prr == 0) &&
18338             sub_from_prr &&
18339             (error == 0)) {
18340                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18341                         rack->r_ctl.rc_prr_sndcnt -= len;
18342                 else
18343                         rack->r_ctl.rc_prr_sndcnt = 0;
18344         }
18345         sub_from_prr = 0;
18346         if (doing_tlp && (rsm == NULL)) {
18347                 /* New send doing a TLP */
18348                 add_flag |= RACK_TLP;
18349         }
18350         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18351                         rack_to_usec_ts(&tv),
18352                         rsm, add_flag, s_mb, s_moff, hw_tls);
18353
18354
18355         if ((error == 0) &&
18356             (len > 0) &&
18357             (tp->snd_una == tp->snd_max))
18358                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18359         {
18360                 tcp_seq startseq = tp->snd_nxt;
18361
18362                 /* Track our lost count */
18363                 if (rsm && (doing_tlp == 0))
18364                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18365                 /*
18366                  * Advance snd_nxt over sequence space of this segment.
18367                  */
18368                 if (error)
18369                         /* We don't log or do anything with errors */
18370                         goto nomore;
18371                 if (doing_tlp == 0) {
18372                         if (rsm == NULL) {
18373                                 /*
18374                                  * Not a retransmission of some
18375                                  * sort, new data is going out so
18376                                  * clear our TLP count and flag.
18377                                  */
18378                                 rack->rc_tlp_in_progress = 0;
18379                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18380                         }
18381                 } else {
18382                         /*
18383                          * We have just sent a TLP, mark that it is true
18384                          * and make sure our in progress is set so we
18385                          * continue to check the count.
18386                          */
18387                         rack->rc_tlp_in_progress = 1;
18388                         rack->r_ctl.rc_tlp_cnt_out++;
18389                 }
18390                 if (flags & (TH_SYN | TH_FIN)) {
18391                         if (flags & TH_SYN)
18392                                 tp->snd_nxt++;
18393                         if (flags & TH_FIN) {
18394                                 tp->snd_nxt++;
18395                                 tp->t_flags |= TF_SENTFIN;
18396                         }
18397                 }
18398                 /* In the ENOBUFS case we do *not* update snd_max */
18399                 if (sack_rxmit)
18400                         goto nomore;
18401
18402                 tp->snd_nxt += len;
18403                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18404                         if (tp->snd_una == tp->snd_max) {
18405                                 /*
18406                                  * Update the time we just added data since
18407                                  * none was outstanding.
18408                                  */
18409                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18410                                 tp->t_acktime = ticks;
18411                         }
18412                         tp->snd_max = tp->snd_nxt;
18413                         /*
18414                          * Time this transmission if not a retransmission and
18415                          * not currently timing anything.
18416                          * This is only relevant in case of switching back to
18417                          * the base stack.
18418                          */
18419                         if (tp->t_rtttime == 0) {
18420                                 tp->t_rtttime = ticks;
18421                                 tp->t_rtseq = startseq;
18422                                 KMOD_TCPSTAT_INC(tcps_segstimed);
18423                         }
18424                         if (len &&
18425                             ((tp->t_flags & TF_GPUTINPROG) == 0))
18426                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18427                 }
18428                 /*
18429                  * If we are doing FO we need to update the mbuf position and subtract
18430                  * this happens when the peer sends us duplicate information and
18431                  * we thus want to send a DSACK.
18432                  *
18433                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18434                  * turned off? If not then we are going to echo multiple DSACK blocks
18435                  * out (with the TSO), which we should not be doing.
18436                  */
18437                 if (rack->r_fast_output && len) {
18438                         if (rack->r_ctl.fsb.left_to_send > len)
18439                                 rack->r_ctl.fsb.left_to_send -= len;
18440                         else
18441                                 rack->r_ctl.fsb.left_to_send = 0;
18442                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18443                                 rack->r_fast_output = 0;
18444                         if (rack->r_fast_output) {
18445                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18446                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18447                         }
18448                 }
18449         }
18450 nomore:
18451         if (error) {
18452                 rack->r_ctl.rc_agg_delayed = 0;
18453                 rack->r_early = 0;
18454                 rack->r_late = 0;
18455                 rack->r_ctl.rc_agg_early = 0;
18456                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
18457                 /*
18458                  * Failures do not advance the seq counter above. For the
18459                  * case of ENOBUFS we will fall out and retry in 1ms with
18460                  * the hpts. Everything else will just have to retransmit
18461                  * with the timer.
18462                  *
18463                  * In any case, we do not want to loop around for another
18464                  * send without a good reason.
18465                  */
18466                 sendalot = 0;
18467                 switch (error) {
18468                 case EPERM:
18469                         tp->t_softerror = error;
18470 #ifdef TCP_ACCOUNTING
18471                         crtsc = get_cyclecount();
18472                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18473                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18474                         }
18475                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18476                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18477                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18478                         }
18479                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18480                         sched_unpin();
18481 #endif
18482                         return (error);
18483                 case ENOBUFS:
18484                         /*
18485                          * Pace us right away to retry in a some
18486                          * time
18487                          */
18488                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18489                         if (rack->rc_enobuf < 0x7f)
18490                                 rack->rc_enobuf++;
18491                         if (slot < (10 * HPTS_USEC_IN_MSEC))
18492                                 slot = 10 * HPTS_USEC_IN_MSEC;
18493                         if (rack->r_ctl.crte != NULL) {
18494                                 counter_u64_add(rack_saw_enobuf_hw, 1);
18495                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
18496                         }
18497                         counter_u64_add(rack_saw_enobuf, 1);
18498                         goto enobufs;
18499                 case EMSGSIZE:
18500                         /*
18501                          * For some reason the interface we used initially
18502                          * to send segments changed to another or lowered
18503                          * its MTU. If TSO was active we either got an
18504                          * interface without TSO capabilits or TSO was
18505                          * turned off. If we obtained mtu from ip_output()
18506                          * then update it and try again.
18507                          */
18508                         if (tso)
18509                                 tp->t_flags &= ~TF_TSO;
18510                         if (mtu != 0) {
18511                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
18512                                 goto again;
18513                         }
18514                         slot = 10 * HPTS_USEC_IN_MSEC;
18515                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18516 #ifdef TCP_ACCOUNTING
18517                         crtsc = get_cyclecount();
18518                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18519                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18520                         }
18521                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18522                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18523                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18524                         }
18525                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18526                         sched_unpin();
18527 #endif
18528                         return (error);
18529                 case ENETUNREACH:
18530                         counter_u64_add(rack_saw_enetunreach, 1);
18531                 case EHOSTDOWN:
18532                 case EHOSTUNREACH:
18533                 case ENETDOWN:
18534                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
18535                                 tp->t_softerror = error;
18536                         }
18537                         /* FALLTHROUGH */
18538                 default:
18539                         slot = 10 * HPTS_USEC_IN_MSEC;
18540                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18541 #ifdef TCP_ACCOUNTING
18542                         crtsc = get_cyclecount();
18543                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18544                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18545                         }
18546                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18547                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18548                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18549                         }
18550                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18551                         sched_unpin();
18552 #endif
18553                         return (error);
18554                 }
18555         } else {
18556                 rack->rc_enobuf = 0;
18557                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18558                         rack->r_ctl.retran_during_recovery += len;
18559         }
18560         KMOD_TCPSTAT_INC(tcps_sndtotal);
18561
18562         /*
18563          * Data sent (as far as we can tell). If this advertises a larger
18564          * window than any other segment, then remember the size of the
18565          * advertised window. Any pending ACK has now been sent.
18566          */
18567         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18568                 tp->rcv_adv = tp->rcv_nxt + recwin;
18569
18570         tp->last_ack_sent = tp->rcv_nxt;
18571         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18572 enobufs:
18573         if (sendalot) {
18574                 /* Do we need to turn off sendalot? */
18575                 if (rack->r_ctl.rc_pace_max_segs &&
18576                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18577                         /* We hit our max. */
18578                         sendalot = 0;
18579                 } else if ((rack->rc_user_set_max_segs) &&
18580                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18581                         /* We hit the user defined max */
18582                         sendalot = 0;
18583                 }
18584         }
18585         if ((error == 0) && (flags & TH_FIN))
18586                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18587         if (flags & TH_RST) {
18588                 /*
18589                  * We don't send again after sending a RST.
18590                  */
18591                 slot = 0;
18592                 sendalot = 0;
18593                 if (error == 0)
18594                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18595         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18596                 /*
18597                  * Get our pacing rate, if an error
18598                  * occurred in sending (ENOBUF) we would
18599                  * hit the else if with slot preset. Other
18600                  * errors return.
18601                  */
18602                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18603         }
18604         if (rsm &&
18605             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18606             rack->use_rack_rr) {
18607                 /* Its a retransmit and we use the rack cheat? */
18608                 if ((slot == 0) ||
18609                     (rack->rc_always_pace == 0) ||
18610                     (rack->r_rr_config == 1)) {
18611                         /*
18612                          * We have no pacing set or we
18613                          * are using old-style rack or
18614                          * we are overriden to use the old 1ms pacing.
18615                          */
18616                         slot = rack->r_ctl.rc_min_to;
18617                 }
18618         }
18619         /* We have sent clear the flag */
18620         rack->r_ent_rec_ns = 0;
18621         if (rack->r_must_retran) {
18622                 if (rsm) {
18623                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
18624                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
18625                                 /*
18626                                  * We have retransmitted all.
18627                                  */
18628                                 rack->r_must_retran = 0;
18629                                 rack->r_ctl.rc_out_at_rto = 0;
18630                         }
18631                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18632                         /*
18633                          * Sending new data will also kill
18634                          * the loop.
18635                          */
18636                         rack->r_must_retran = 0;
18637                         rack->r_ctl.rc_out_at_rto = 0;
18638                 }
18639         }
18640         rack->r_ctl.fsb.recwin = recwin;
18641         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
18642             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18643                 /*
18644                  * We hit an RTO and now have past snd_max at the RTO
18645                  * clear all the WAS flags.
18646                  */
18647                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
18648         }
18649         if (slot) {
18650                 /* set the rack tcb into the slot N */
18651                 counter_u64_add(rack_paced_segments, 1);
18652                 if ((error == 0) &&
18653                     rack_use_rfo &&
18654                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18655                     (rsm == NULL) &&
18656                     (tp->snd_nxt == tp->snd_max) &&
18657                     (ipoptlen == 0) &&
18658                     (tp->rcv_numsacks == 0) &&
18659                     rack->r_fsb_inited &&
18660                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18661                     (rack->r_must_retran == 0) &&
18662                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18663                     (len > 0) && (orig_len > 0) &&
18664                     (orig_len > len) &&
18665                     ((orig_len - len) >= segsiz) &&
18666                     ((optlen == 0) ||
18667                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18668                         /* We can send at least one more MSS using our fsb */
18669
18670                         rack->r_fast_output = 1;
18671                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18672                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18673                         rack->r_ctl.fsb.tcp_flags = flags;
18674                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18675                         if (hw_tls)
18676                                 rack->r_ctl.fsb.hw_tls = 1;
18677                         else
18678                                 rack->r_ctl.fsb.hw_tls = 0;
18679                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18680                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18681                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18682                                  (tp->snd_max - tp->snd_una)));
18683                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18684                                 rack->r_fast_output = 0;
18685                         else {
18686                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18687                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18688                                 else
18689                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18690                         }
18691                 } else
18692                         rack->r_fast_output = 0;
18693                 rack_log_fsb(rack, tp, so, flags,
18694                              ipoptlen, orig_len, len, error,
18695                              (rsm == NULL), optlen, __LINE__, 2);
18696         } else if (sendalot) {
18697                 int ret;
18698
18699                 if (len)
18700                         counter_u64_add(rack_unpaced_segments, 1);
18701                 sack_rxmit = 0;
18702                 if ((error == 0) &&
18703                     rack_use_rfo &&
18704                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18705                     (rsm == NULL) &&
18706                     (ipoptlen == 0) &&
18707                     (tp->rcv_numsacks == 0) &&
18708                     (tp->snd_nxt == tp->snd_max) &&
18709                     (rack->r_must_retran == 0) &&
18710                     rack->r_fsb_inited &&
18711                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18712                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18713                     (len > 0) && (orig_len > 0) &&
18714                     (orig_len > len) &&
18715                     ((orig_len - len) >= segsiz) &&
18716                     ((optlen == 0) ||
18717                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18718                         /* we can use fast_output for more */
18719
18720                         rack->r_fast_output = 1;
18721                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18722                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18723                         rack->r_ctl.fsb.tcp_flags = flags;
18724                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18725                         if (hw_tls)
18726                                 rack->r_ctl.fsb.hw_tls = 1;
18727                         else
18728                                 rack->r_ctl.fsb.hw_tls = 0;
18729                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18730                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18731                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18732                                  (tp->snd_max - tp->snd_una)));
18733                         if (rack->r_ctl.fsb.left_to_send < segsiz) {
18734                                 rack->r_fast_output = 0;
18735                         }
18736                         if (rack->r_fast_output) {
18737                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18738                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18739                                 else
18740                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18741                                 rack_log_fsb(rack, tp, so, flags,
18742                                              ipoptlen, orig_len, len, error,
18743                                              (rsm == NULL), optlen, __LINE__, 3);
18744                                 error = 0;
18745                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
18746                                 if (ret >= 0)
18747                                         return (ret);
18748                                 else if (error)
18749                                         goto nomore;
18750
18751                         }
18752                 }
18753                 goto again;
18754         } else if (len) {
18755                 counter_u64_add(rack_unpaced_segments, 1);
18756         }
18757         /* Assure when we leave that snd_nxt will point to top */
18758         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
18759                 tp->snd_nxt = tp->snd_max;
18760         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
18761 #ifdef TCP_ACCOUNTING
18762         crtsc = get_cyclecount() - ts_val;
18763         if (tot_len_this_send) {
18764                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18765                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
18766                 }
18767                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
18768                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18769                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
18770                 }
18771                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
18772                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18773                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
18774                 }
18775                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
18776         } else {
18777                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18778                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
18779                 }
18780                 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
18781                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18782                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
18783                 }
18784                 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
18785         }
18786         sched_unpin();
18787 #endif
18788         if (error == ENOBUFS)
18789                 error = 0;
18790         return (error);
18791 }
18792
18793 static void
18794 rack_update_seg(struct tcp_rack *rack)
18795 {
18796         uint32_t orig_val;
18797
18798         orig_val = rack->r_ctl.rc_pace_max_segs;
18799         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
18800         if (orig_val != rack->r_ctl.rc_pace_max_segs)
18801                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
18802 }
18803
18804 static void
18805 rack_mtu_change(struct tcpcb *tp)
18806 {
18807         /*
18808          * The MSS may have changed
18809          */
18810         struct tcp_rack *rack;
18811
18812         rack = (struct tcp_rack *)tp->t_fb_ptr;
18813         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
18814                 /*
18815                  * The MTU has changed we need to resend everything
18816                  * since all we have sent is lost. We first fix
18817                  * up the mtu though.
18818                  */
18819                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
18820                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
18821                 rack_remxt_tmr(tp);
18822                 rack->r_fast_output = 0;
18823                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
18824                                                 rack->r_ctl.rc_sacked);
18825                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
18826                 rack->r_must_retran = 1;
18827
18828         }
18829         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
18830         /* We don't use snd_nxt to retransmit */
18831         tp->snd_nxt = tp->snd_max;
18832 }
18833
18834 static int
18835 rack_set_profile(struct tcp_rack *rack, int prof)
18836 {
18837         int err = EINVAL;
18838         if (prof == 1) {
18839                 /* pace_always=1 */
18840                 if (rack->rc_always_pace == 0) {
18841                         if (tcp_can_enable_pacing() == 0)
18842                                 return (EBUSY);
18843                 }
18844                 rack->rc_always_pace = 1;
18845                 if (rack->use_fixed_rate || rack->gp_ready)
18846                         rack_set_cc_pacing(rack);
18847                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18848                 rack->rack_attempt_hdwr_pace = 0;
18849                 /* cmpack=1 */
18850                 if (rack_use_cmp_acks)
18851                         rack->r_use_cmp_ack = 1;
18852                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18853                     rack->r_use_cmp_ack)
18854                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18855                 /* scwnd=1 */
18856                 rack->rack_enable_scwnd = 1;
18857                 /* dynamic=100 */
18858                 rack->rc_gp_dyn_mul = 1;
18859                 /* gp_inc_ca */
18860                 rack->r_ctl.rack_per_of_gp_ca = 100;
18861                 /* rrr_conf=3 */
18862                 rack->r_rr_config = 3;
18863                 /* npush=2 */
18864                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18865                 /* fillcw=1 */
18866                 rack->rc_pace_to_cwnd = 1;
18867                 rack->rc_pace_fill_if_rttin_range = 0;
18868                 rack->rtt_limit_mul = 0;
18869                 /* noprr=1 */
18870                 rack->rack_no_prr = 1;
18871                 /* lscwnd=1 */
18872                 rack->r_limit_scw = 1;
18873                 /* gp_inc_rec */
18874                 rack->r_ctl.rack_per_of_gp_rec = 90;
18875                 err = 0;
18876
18877         } else if (prof == 3) {
18878                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
18879                 /* pace_always=1 */
18880                 if (rack->rc_always_pace == 0) {
18881                         if (tcp_can_enable_pacing() == 0)
18882                                 return (EBUSY);
18883                 }
18884                 rack->rc_always_pace = 1;
18885                 if (rack->use_fixed_rate || rack->gp_ready)
18886                         rack_set_cc_pacing(rack);
18887                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18888                 rack->rack_attempt_hdwr_pace = 0;
18889                 /* cmpack=1 */
18890                 if (rack_use_cmp_acks)
18891                         rack->r_use_cmp_ack = 1;
18892                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18893                     rack->r_use_cmp_ack)
18894                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18895                 /* scwnd=1 */
18896                 rack->rack_enable_scwnd = 1;
18897                 /* dynamic=100 */
18898                 rack->rc_gp_dyn_mul = 1;
18899                 /* gp_inc_ca */
18900                 rack->r_ctl.rack_per_of_gp_ca = 100;
18901                 /* rrr_conf=3 */
18902                 rack->r_rr_config = 3;
18903                 /* npush=2 */
18904                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18905                 /* fillcw=2 */
18906                 rack->rc_pace_to_cwnd = 1;
18907                 rack->r_fill_less_agg = 1;
18908                 rack->rc_pace_fill_if_rttin_range = 0;
18909                 rack->rtt_limit_mul = 0;
18910                 /* noprr=1 */
18911                 rack->rack_no_prr = 1;
18912                 /* lscwnd=1 */
18913                 rack->r_limit_scw = 1;
18914                 /* gp_inc_rec */
18915                 rack->r_ctl.rack_per_of_gp_rec = 90;
18916                 err = 0;
18917
18918
18919         } else if (prof == 2) {
18920                 /* cmpack=1 */
18921                 if (rack->rc_always_pace == 0) {
18922                         if (tcp_can_enable_pacing() == 0)
18923                                 return (EBUSY);
18924                 }
18925                 rack->rc_always_pace = 1;
18926                 if (rack->use_fixed_rate || rack->gp_ready)
18927                         rack_set_cc_pacing(rack);
18928                 rack->r_use_cmp_ack = 1;
18929                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18930                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18931                 /* pace_always=1 */
18932                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18933                 /* scwnd=1 */
18934                 rack->rack_enable_scwnd = 1;
18935                 /* dynamic=100 */
18936                 rack->rc_gp_dyn_mul = 1;
18937                 rack->r_ctl.rack_per_of_gp_ca = 100;
18938                 /* rrr_conf=3 */
18939                 rack->r_rr_config = 3;
18940                 /* npush=2 */
18941                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18942                 /* fillcw=1 */
18943                 rack->rc_pace_to_cwnd = 1;
18944                 rack->rc_pace_fill_if_rttin_range = 0;
18945                 rack->rtt_limit_mul = 0;
18946                 /* noprr=1 */
18947                 rack->rack_no_prr = 1;
18948                 /* lscwnd=0 */
18949                 rack->r_limit_scw = 0;
18950                 err = 0;
18951         } else if (prof == 0) {
18952                 /* This changes things back to the default settings */
18953                 err = 0;
18954                 if (rack->rc_always_pace) {
18955                         tcp_decrement_paced_conn();
18956                         rack_undo_cc_pacing(rack);
18957                         rack->rc_always_pace = 0;
18958                 }
18959                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
18960                         rack->rc_always_pace = 1;
18961                         if (rack->use_fixed_rate || rack->gp_ready)
18962                                 rack_set_cc_pacing(rack);
18963                 } else
18964                         rack->rc_always_pace = 0;
18965                 if (rack_use_cmp_acks)
18966                         rack->r_use_cmp_ack = 1;
18967                 else
18968                         rack->r_use_cmp_ack = 0;
18969                 if (rack_disable_prr)
18970                         rack->rack_no_prr = 1;
18971                 else
18972                         rack->rack_no_prr = 0;
18973                 if (rack_gp_no_rec_chg)
18974                         rack->rc_gp_no_rec_chg = 1;
18975                 else
18976                         rack->rc_gp_no_rec_chg = 0;
18977                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
18978                         rack->r_mbuf_queue = 1;
18979                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18980                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18981                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18982                 } else {
18983                         rack->r_mbuf_queue = 0;
18984                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
18985                 }
18986                 if (rack_enable_shared_cwnd)
18987                         rack->rack_enable_scwnd = 1;
18988                 else
18989                         rack->rack_enable_scwnd = 0;
18990                 if (rack_do_dyn_mul) {
18991                         /* When dynamic adjustment is on CA needs to start at 100% */
18992                         rack->rc_gp_dyn_mul = 1;
18993                         if (rack_do_dyn_mul >= 100)
18994                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
18995                 } else {
18996                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
18997                         rack->rc_gp_dyn_mul = 0;
18998                 }
18999                 rack->r_rr_config = 0;
19000                 rack->r_ctl.rc_no_push_at_mrtt = 0;
19001                 rack->rc_pace_to_cwnd = 0;
19002                 rack->rc_pace_fill_if_rttin_range = 0;
19003                 rack->rtt_limit_mul = 0;
19004
19005                 if (rack_enable_hw_pacing)
19006                         rack->rack_hdw_pace_ena = 1;
19007                 else
19008                         rack->rack_hdw_pace_ena = 0;
19009                 if (rack_disable_prr)
19010                         rack->rack_no_prr = 1;
19011                 else
19012                         rack->rack_no_prr = 0;
19013                 if (rack_limits_scwnd)
19014                         rack->r_limit_scw  = 1;
19015                 else
19016                         rack->r_limit_scw  = 0;
19017                 err = 0;
19018         }
19019         return (err);
19020 }
19021
19022 static int
19023 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19024 {
19025         struct deferred_opt_list *dol;
19026
19027         dol = malloc(sizeof(struct deferred_opt_list),
19028                      M_TCPFSB, M_NOWAIT|M_ZERO);
19029         if (dol == NULL) {
19030                 /*
19031                  * No space yikes -- fail out..
19032                  */
19033                 return (0);
19034         }
19035         dol->optname = sopt_name;
19036         dol->optval = loptval;
19037         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19038         return (1);
19039 }
19040
19041 static int
19042 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19043                     uint32_t optval, uint64_t loptval)
19044 {
19045         struct epoch_tracker et;
19046         struct sockopt sopt;
19047         struct cc_newreno_opts opt;
19048         uint64_t val;
19049         int error = 0;
19050         uint16_t ca, ss;
19051
19052         switch (sopt_name) {
19053
19054         case TCP_RACK_PACING_BETA:
19055                 RACK_OPTS_INC(tcp_rack_beta);
19056                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19057                         /* This only works for newreno. */
19058                         error = EINVAL;
19059                         break;
19060                 }
19061                 if (rack->rc_pacing_cc_set) {
19062                         /*
19063                          * Set them into the real CC module
19064                          * whats in the rack pcb is the old values
19065                          * to be used on restoral/
19066                          */
19067                         sopt.sopt_dir = SOPT_SET;
19068                         opt.name = CC_NEWRENO_BETA;
19069                         opt.val = optval;
19070                         if (CC_ALGO(tp)->ctl_output != NULL)
19071                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19072                         else {
19073                                 error = ENOENT;
19074                                 break;
19075                         }
19076                 } else {
19077                         /*
19078                          * Not pacing yet so set it into our local
19079                          * rack pcb storage.
19080                          */
19081                         rack->r_ctl.rc_saved_beta.beta = optval;
19082                 }
19083                 break;
19084         case TCP_RACK_TIMER_SLOP:
19085                 RACK_OPTS_INC(tcp_rack_timer_slop);
19086                 rack->r_ctl.timer_slop = optval;
19087                 if (rack->rc_tp->t_srtt) {
19088                         /*
19089                          * If we have an SRTT lets update t_rxtcur
19090                          * to have the new slop.
19091                          */
19092                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19093                                            rack_rto_min, rack_rto_max,
19094                                            rack->r_ctl.timer_slop);
19095                 }
19096                 break;
19097         case TCP_RACK_PACING_BETA_ECN:
19098                 RACK_OPTS_INC(tcp_rack_beta_ecn);
19099                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19100                         /* This only works for newreno. */
19101                         error = EINVAL;
19102                         break;
19103                 }
19104                 if (rack->rc_pacing_cc_set) {
19105                         /*
19106                          * Set them into the real CC module
19107                          * whats in the rack pcb is the old values
19108                          * to be used on restoral/
19109                          */
19110                         sopt.sopt_dir = SOPT_SET;
19111                         opt.name = CC_NEWRENO_BETA_ECN;
19112                         opt.val = optval;
19113                         if (CC_ALGO(tp)->ctl_output != NULL)
19114                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19115                         else
19116                                 error = ENOENT;
19117                 } else {
19118                         /*
19119                          * Not pacing yet so set it into our local
19120                          * rack pcb storage.
19121                          */
19122                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19123                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
19124                 }
19125                 break;
19126         case TCP_DEFER_OPTIONS:
19127                 RACK_OPTS_INC(tcp_defer_opt);
19128                 if (optval) {
19129                         if (rack->gp_ready) {
19130                                 /* Too late */
19131                                 error = EINVAL;
19132                                 break;
19133                         }
19134                         rack->defer_options = 1;
19135                 } else
19136                         rack->defer_options = 0;
19137                 break;
19138         case TCP_RACK_MEASURE_CNT:
19139                 RACK_OPTS_INC(tcp_rack_measure_cnt);
19140                 if (optval && (optval <= 0xff)) {
19141                         rack->r_ctl.req_measurements = optval;
19142                 } else
19143                         error = EINVAL;
19144                 break;
19145         case TCP_REC_ABC_VAL:
19146                 RACK_OPTS_INC(tcp_rec_abc_val);
19147                 if (optval > 0)
19148                         rack->r_use_labc_for_rec = 1;
19149                 else
19150                         rack->r_use_labc_for_rec = 0;
19151                 break;
19152         case TCP_RACK_ABC_VAL:
19153                 RACK_OPTS_INC(tcp_rack_abc_val);
19154                 if ((optval > 0) && (optval < 255))
19155                         rack->rc_labc = optval;
19156                 else
19157                         error = EINVAL;
19158                 break;
19159         case TCP_HDWR_UP_ONLY:
19160                 RACK_OPTS_INC(tcp_pacing_up_only);
19161                 if (optval)
19162                         rack->r_up_only = 1;
19163                 else
19164                         rack->r_up_only = 0;
19165                 break;
19166         case TCP_PACING_RATE_CAP:
19167                 RACK_OPTS_INC(tcp_pacing_rate_cap);
19168                 rack->r_ctl.bw_rate_cap = loptval;
19169                 break;
19170         case TCP_RACK_PROFILE:
19171                 RACK_OPTS_INC(tcp_profile);
19172                 error = rack_set_profile(rack, optval);
19173                 break;
19174         case TCP_USE_CMP_ACKS:
19175                 RACK_OPTS_INC(tcp_use_cmp_acks);
19176                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19177                         /* You can't turn it off once its on! */
19178                         error = EINVAL;
19179                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19180                         rack->r_use_cmp_ack = 1;
19181                         rack->r_mbuf_queue = 1;
19182                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19183                 }
19184                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19185                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19186                 break;
19187         case TCP_SHARED_CWND_TIME_LIMIT:
19188                 RACK_OPTS_INC(tcp_lscwnd);
19189                 if (optval)
19190                         rack->r_limit_scw = 1;
19191                 else
19192                         rack->r_limit_scw = 0;
19193                 break;
19194         case TCP_RACK_PACE_TO_FILL:
19195                 RACK_OPTS_INC(tcp_fillcw);
19196                 if (optval == 0)
19197                         rack->rc_pace_to_cwnd = 0;
19198                 else {
19199                         rack->rc_pace_to_cwnd = 1;
19200                         if (optval > 1)
19201                                 rack->r_fill_less_agg = 1;
19202                 }
19203                 if ((optval >= rack_gp_rtt_maxmul) &&
19204                     rack_gp_rtt_maxmul &&
19205                     (optval < 0xf)) {
19206                         rack->rc_pace_fill_if_rttin_range = 1;
19207                         rack->rtt_limit_mul = optval;
19208                 } else {
19209                         rack->rc_pace_fill_if_rttin_range = 0;
19210                         rack->rtt_limit_mul = 0;
19211                 }
19212                 break;
19213         case TCP_RACK_NO_PUSH_AT_MAX:
19214                 RACK_OPTS_INC(tcp_npush);
19215                 if (optval == 0)
19216                         rack->r_ctl.rc_no_push_at_mrtt = 0;
19217                 else if (optval < 0xff)
19218                         rack->r_ctl.rc_no_push_at_mrtt = optval;
19219                 else
19220                         error = EINVAL;
19221                 break;
19222         case TCP_SHARED_CWND_ENABLE:
19223                 RACK_OPTS_INC(tcp_rack_scwnd);
19224                 if (optval == 0)
19225                         rack->rack_enable_scwnd = 0;
19226                 else
19227                         rack->rack_enable_scwnd = 1;
19228                 break;
19229         case TCP_RACK_MBUF_QUEUE:
19230                 /* Now do we use the LRO mbuf-queue feature */
19231                 RACK_OPTS_INC(tcp_rack_mbufq);
19232                 if (optval || rack->r_use_cmp_ack)
19233                         rack->r_mbuf_queue = 1;
19234                 else
19235                         rack->r_mbuf_queue = 0;
19236                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19237                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19238                 else
19239                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19240                 break;
19241         case TCP_RACK_NONRXT_CFG_RATE:
19242                 RACK_OPTS_INC(tcp_rack_cfg_rate);
19243                 if (optval == 0)
19244                         rack->rack_rec_nonrxt_use_cr = 0;
19245                 else
19246                         rack->rack_rec_nonrxt_use_cr = 1;
19247                 break;
19248         case TCP_NO_PRR:
19249                 RACK_OPTS_INC(tcp_rack_noprr);
19250                 if (optval == 0)
19251                         rack->rack_no_prr = 0;
19252                 else if (optval == 1)
19253                         rack->rack_no_prr = 1;
19254                 else if (optval == 2)
19255                         rack->no_prr_addback = 1;
19256                 else
19257                         error = EINVAL;
19258                 break;
19259         case TCP_TIMELY_DYN_ADJ:
19260                 RACK_OPTS_INC(tcp_timely_dyn);
19261                 if (optval == 0)
19262                         rack->rc_gp_dyn_mul = 0;
19263                 else {
19264                         rack->rc_gp_dyn_mul = 1;
19265                         if (optval >= 100) {
19266                                 /*
19267                                  * If the user sets something 100 or more
19268                                  * its the gp_ca value.
19269                                  */
19270                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19271                         }
19272                 }
19273                 break;
19274         case TCP_RACK_DO_DETECTION:
19275                 RACK_OPTS_INC(tcp_rack_do_detection);
19276                 if (optval == 0)
19277                         rack->do_detection = 0;
19278                 else
19279                         rack->do_detection = 1;
19280                 break;
19281         case TCP_RACK_TLP_USE:
19282                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19283                         error = EINVAL;
19284                         break;
19285                 }
19286                 RACK_OPTS_INC(tcp_tlp_use);
19287                 rack->rack_tlp_threshold_use = optval;
19288                 break;
19289         case TCP_RACK_TLP_REDUCE:
19290                 /* RACK TLP cwnd reduction (bool) */
19291                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19292                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19293                 break;
19294         /*  Pacing related ones */
19295         case TCP_RACK_PACE_ALWAYS:
19296                 /*
19297                  * zero is old rack method, 1 is new
19298                  * method using a pacing rate.
19299                  */
19300                 RACK_OPTS_INC(tcp_rack_pace_always);
19301                 if (optval > 0) {
19302                         if (rack->rc_always_pace) {
19303                                 error = EALREADY;
19304                                 break;
19305                         } else if (tcp_can_enable_pacing()) {
19306                                 rack->rc_always_pace = 1;
19307                                 if (rack->use_fixed_rate || rack->gp_ready)
19308                                         rack_set_cc_pacing(rack);
19309                         }
19310                         else {
19311                                 error = ENOSPC;
19312                                 break;
19313                         }
19314                 } else {
19315                         if (rack->rc_always_pace) {
19316                                 tcp_decrement_paced_conn();
19317                                 rack->rc_always_pace = 0;
19318                                 rack_undo_cc_pacing(rack);
19319                         }
19320                 }
19321                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19322                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19323                 else
19324                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19325                 /* A rate may be set irate or other, if so set seg size */
19326                 rack_update_seg(rack);
19327                 break;
19328         case TCP_BBR_RACK_INIT_RATE:
19329                 RACK_OPTS_INC(tcp_initial_rate);
19330                 val = optval;
19331                 /* Change from kbits per second to bytes per second */
19332                 val *= 1000;
19333                 val /= 8;
19334                 rack->r_ctl.init_rate = val;
19335                 if (rack->rc_init_win != rack_default_init_window) {
19336                         uint32_t win, snt;
19337
19338                         /*
19339                          * Options don't always get applied
19340                          * in the order you think. So in order
19341                          * to assure we update a cwnd we need
19342                          * to check and see if we are still
19343                          * where we should raise the cwnd.
19344                          */
19345                         win = rc_init_window(rack);
19346                         if (SEQ_GT(tp->snd_max, tp->iss))
19347                                 snt = tp->snd_max - tp->iss;
19348                         else
19349                                 snt = 0;
19350                         if ((snt < win) &&
19351                             (tp->snd_cwnd < win))
19352                                 tp->snd_cwnd = win;
19353                 }
19354                 if (rack->rc_always_pace)
19355                         rack_update_seg(rack);
19356                 break;
19357         case TCP_BBR_IWINTSO:
19358                 RACK_OPTS_INC(tcp_initial_win);
19359                 if (optval && (optval <= 0xff)) {
19360                         uint32_t win, snt;
19361
19362                         rack->rc_init_win = optval;
19363                         win = rc_init_window(rack);
19364                         if (SEQ_GT(tp->snd_max, tp->iss))
19365                                 snt = tp->snd_max - tp->iss;
19366                         else
19367                                 snt = 0;
19368                         if ((snt < win) &&
19369                             (tp->t_srtt |
19370 #ifdef NETFLIX_PEAKRATE
19371                              tp->t_maxpeakrate |
19372 #endif
19373                              rack->r_ctl.init_rate)) {
19374                                 /*
19375                                  * We are not past the initial window
19376                                  * and we have some bases for pacing,
19377                                  * so we need to possibly adjust up
19378                                  * the cwnd. Note even if we don't set
19379                                  * the cwnd, its still ok to raise the rc_init_win
19380                                  * which can be used coming out of idle when we
19381                                  * would have a rate.
19382                                  */
19383                                 if (tp->snd_cwnd < win)
19384                                         tp->snd_cwnd = win;
19385                         }
19386                         if (rack->rc_always_pace)
19387                                 rack_update_seg(rack);
19388                 } else
19389                         error = EINVAL;
19390                 break;
19391         case TCP_RACK_FORCE_MSEG:
19392                 RACK_OPTS_INC(tcp_rack_force_max_seg);
19393                 if (optval)
19394                         rack->rc_force_max_seg = 1;
19395                 else
19396                         rack->rc_force_max_seg = 0;
19397                 break;
19398         case TCP_RACK_PACE_MAX_SEG:
19399                 /* Max segments size in a pace in bytes */
19400                 RACK_OPTS_INC(tcp_rack_max_seg);
19401                 rack->rc_user_set_max_segs = optval;
19402                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19403                 break;
19404         case TCP_RACK_PACE_RATE_REC:
19405                 /* Set the fixed pacing rate in Bytes per second ca */
19406                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19407                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19408                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19409                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19410                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19411                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19412                 rack->use_fixed_rate = 1;
19413                 if (rack->rc_always_pace)
19414                         rack_set_cc_pacing(rack);
19415                 rack_log_pacing_delay_calc(rack,
19416                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19417                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19418                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19419                                            __LINE__, NULL,0);
19420                 break;
19421
19422         case TCP_RACK_PACE_RATE_SS:
19423                 /* Set the fixed pacing rate in Bytes per second ca */
19424                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19425                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19426                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19427                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19428                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19429                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19430                 rack->use_fixed_rate = 1;
19431                 if (rack->rc_always_pace)
19432                         rack_set_cc_pacing(rack);
19433                 rack_log_pacing_delay_calc(rack,
19434                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19435                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19436                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19437                                            __LINE__, NULL, 0);
19438                 break;
19439
19440         case TCP_RACK_PACE_RATE_CA:
19441                 /* Set the fixed pacing rate in Bytes per second ca */
19442                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19443                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19444                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19445                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19446                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19447                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19448                 rack->use_fixed_rate = 1;
19449                 if (rack->rc_always_pace)
19450                         rack_set_cc_pacing(rack);
19451                 rack_log_pacing_delay_calc(rack,
19452                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19453                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19454                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19455                                            __LINE__, NULL, 0);
19456                 break;
19457         case TCP_RACK_GP_INCREASE_REC:
19458                 RACK_OPTS_INC(tcp_gp_inc_rec);
19459                 rack->r_ctl.rack_per_of_gp_rec = optval;
19460                 rack_log_pacing_delay_calc(rack,
19461                                            rack->r_ctl.rack_per_of_gp_ss,
19462                                            rack->r_ctl.rack_per_of_gp_ca,
19463                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19464                                            __LINE__, NULL, 0);
19465                 break;
19466         case TCP_RACK_GP_INCREASE_CA:
19467                 RACK_OPTS_INC(tcp_gp_inc_ca);
19468                 ca = optval;
19469                 if (ca < 100) {
19470                         /*
19471                          * We don't allow any reduction
19472                          * over the GP b/w.
19473                          */
19474                         error = EINVAL;
19475                         break;
19476                 }
19477                 rack->r_ctl.rack_per_of_gp_ca = ca;
19478                 rack_log_pacing_delay_calc(rack,
19479                                            rack->r_ctl.rack_per_of_gp_ss,
19480                                            rack->r_ctl.rack_per_of_gp_ca,
19481                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19482                                            __LINE__, NULL, 0);
19483                 break;
19484         case TCP_RACK_GP_INCREASE_SS:
19485                 RACK_OPTS_INC(tcp_gp_inc_ss);
19486                 ss = optval;
19487                 if (ss < 100) {
19488                         /*
19489                          * We don't allow any reduction
19490                          * over the GP b/w.
19491                          */
19492                         error = EINVAL;
19493                         break;
19494                 }
19495                 rack->r_ctl.rack_per_of_gp_ss = ss;
19496                 rack_log_pacing_delay_calc(rack,
19497                                            rack->r_ctl.rack_per_of_gp_ss,
19498                                            rack->r_ctl.rack_per_of_gp_ca,
19499                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19500                                            __LINE__, NULL, 0);
19501                 break;
19502         case TCP_RACK_RR_CONF:
19503                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19504                 if (optval && optval <= 3)
19505                         rack->r_rr_config = optval;
19506                 else
19507                         rack->r_rr_config = 0;
19508                 break;
19509         case TCP_HDWR_RATE_CAP:
19510                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
19511                 if (optval) {
19512                         if (rack->r_rack_hw_rate_caps == 0)
19513                                 rack->r_rack_hw_rate_caps = 1;
19514                         else
19515                                 error = EALREADY;
19516                 } else {
19517                         rack->r_rack_hw_rate_caps = 0;
19518                 }
19519                 break;
19520         case TCP_BBR_HDWR_PACE:
19521                 RACK_OPTS_INC(tcp_hdwr_pacing);
19522                 if (optval){
19523                         if (rack->rack_hdrw_pacing == 0) {
19524                                 rack->rack_hdw_pace_ena = 1;
19525                                 rack->rack_attempt_hdwr_pace = 0;
19526                         } else
19527                                 error = EALREADY;
19528                 } else {
19529                         rack->rack_hdw_pace_ena = 0;
19530 #ifdef RATELIMIT
19531                         if (rack->r_ctl.crte != NULL) {
19532                                 rack->rack_hdrw_pacing = 0;
19533                                 rack->rack_attempt_hdwr_pace = 0;
19534                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19535                                 rack->r_ctl.crte = NULL;
19536                         }
19537 #endif
19538                 }
19539                 break;
19540         /*  End Pacing related ones */
19541         case TCP_RACK_PRR_SENDALOT:
19542                 /* Allow PRR to send more than one seg */
19543                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
19544                 rack->r_ctl.rc_prr_sendalot = optval;
19545                 break;
19546         case TCP_RACK_MIN_TO:
19547                 /* Minimum time between rack t-o's in ms */
19548                 RACK_OPTS_INC(tcp_rack_min_to);
19549                 rack->r_ctl.rc_min_to = optval;
19550                 break;
19551         case TCP_RACK_EARLY_SEG:
19552                 /* If early recovery max segments */
19553                 RACK_OPTS_INC(tcp_rack_early_seg);
19554                 rack->r_ctl.rc_early_recovery_segs = optval;
19555                 break;
19556         case TCP_RACK_REORD_THRESH:
19557                 /* RACK reorder threshold (shift amount) */
19558                 RACK_OPTS_INC(tcp_rack_reord_thresh);
19559                 if ((optval > 0) && (optval < 31))
19560                         rack->r_ctl.rc_reorder_shift = optval;
19561                 else
19562                         error = EINVAL;
19563                 break;
19564         case TCP_RACK_REORD_FADE:
19565                 /* Does reordering fade after ms time */
19566                 RACK_OPTS_INC(tcp_rack_reord_fade);
19567                 rack->r_ctl.rc_reorder_fade = optval;
19568                 break;
19569         case TCP_RACK_TLP_THRESH:
19570                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19571                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
19572                 if (optval)
19573                         rack->r_ctl.rc_tlp_threshold = optval;
19574                 else
19575                         error = EINVAL;
19576                 break;
19577         case TCP_BBR_USE_RACK_RR:
19578                 RACK_OPTS_INC(tcp_rack_rr);
19579                 if (optval)
19580                         rack->use_rack_rr = 1;
19581                 else
19582                         rack->use_rack_rr = 0;
19583                 break;
19584         case TCP_FAST_RSM_HACK:
19585                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
19586                 if (optval)
19587                         rack->fast_rsm_hack = 1;
19588                 else
19589                         rack->fast_rsm_hack = 0;
19590                 break;
19591         case TCP_RACK_PKT_DELAY:
19592                 /* RACK added ms i.e. rack-rtt + reord + N */
19593                 RACK_OPTS_INC(tcp_rack_pkt_delay);
19594                 rack->r_ctl.rc_pkt_delay = optval;
19595                 break;
19596         case TCP_DELACK:
19597                 RACK_OPTS_INC(tcp_rack_delayed_ack);
19598                 if (optval == 0)
19599                         tp->t_delayed_ack = 0;
19600                 else
19601                         tp->t_delayed_ack = 1;
19602                 if (tp->t_flags & TF_DELACK) {
19603                         tp->t_flags &= ~TF_DELACK;
19604                         tp->t_flags |= TF_ACKNOW;
19605                         NET_EPOCH_ENTER(et);
19606                         rack_output(tp);
19607                         NET_EPOCH_EXIT(et);
19608                 }
19609                 break;
19610
19611         case TCP_BBR_RACK_RTT_USE:
19612                 RACK_OPTS_INC(tcp_rack_rtt_use);
19613                 if ((optval != USE_RTT_HIGH) &&
19614                     (optval != USE_RTT_LOW) &&
19615                     (optval != USE_RTT_AVG))
19616                         error = EINVAL;
19617                 else
19618                         rack->r_ctl.rc_rate_sample_method = optval;
19619                 break;
19620         case TCP_DATA_AFTER_CLOSE:
19621                 RACK_OPTS_INC(tcp_data_after_close);
19622                 if (optval)
19623                         rack->rc_allow_data_af_clo = 1;
19624                 else
19625                         rack->rc_allow_data_af_clo = 0;
19626                 break;
19627         default:
19628                 break;
19629         }
19630 #ifdef NETFLIX_STATS
19631         tcp_log_socket_option(tp, sopt_name, optval, error);
19632 #endif
19633         return (error);
19634 }
19635
19636
19637 static void
19638 rack_apply_deferred_options(struct tcp_rack *rack)
19639 {
19640         struct deferred_opt_list *dol, *sdol;
19641         uint32_t s_optval;
19642
19643         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
19644                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
19645                 /* Disadvantage of deferal is you loose the error return */
19646                 s_optval = (uint32_t)dol->optval;
19647                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
19648                 free(dol, M_TCPDO);
19649         }
19650 }
19651
19652 static void
19653 rack_hw_tls_change(struct tcpcb *tp, int chg)
19654 {
19655         /*
19656          * HW tls state has changed.. fix all
19657          * rsm's in flight.
19658          */
19659         struct tcp_rack *rack;
19660         struct rack_sendmap *rsm;
19661
19662         rack = (struct tcp_rack *)tp->t_fb_ptr;
19663         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
19664                 if (chg)
19665                         rsm->r_hw_tls = 1;
19666                 else
19667                         rsm->r_hw_tls = 0;
19668         }
19669         if (chg)
19670                 rack->r_ctl.fsb.hw_tls = 1;
19671         else
19672                 rack->r_ctl.fsb.hw_tls = 0;
19673 }
19674
19675 static int
19676 rack_pru_options(struct tcpcb *tp, int flags)
19677 {
19678         if (flags & PRUS_OOB)
19679                 return (EOPNOTSUPP);
19680         return (0);
19681 }
19682
19683 static struct tcp_function_block __tcp_rack = {
19684         .tfb_tcp_block_name = __XSTRING(STACKNAME),
19685         .tfb_tcp_output = rack_output,
19686         .tfb_do_queued_segments = ctf_do_queued_segments,
19687         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
19688         .tfb_tcp_do_segment = rack_do_segment,
19689         .tfb_tcp_ctloutput = rack_ctloutput,
19690         .tfb_tcp_fb_init = rack_init,
19691         .tfb_tcp_fb_fini = rack_fini,
19692         .tfb_tcp_timer_stop_all = rack_stopall,
19693         .tfb_tcp_timer_activate = rack_timer_activate,
19694         .tfb_tcp_timer_active = rack_timer_active,
19695         .tfb_tcp_timer_stop = rack_timer_stop,
19696         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
19697         .tfb_tcp_handoff_ok = rack_handoff_ok,
19698         .tfb_tcp_mtu_chg = rack_mtu_change,
19699         .tfb_pru_options = rack_pru_options,
19700         .tfb_hwtls_change = rack_hw_tls_change,
19701 };
19702
19703 /*
19704  * rack_ctloutput() must drop the inpcb lock before performing copyin on
19705  * socket option arguments.  When it re-acquires the lock after the copy, it
19706  * has to revalidate that the connection is still valid for the socket
19707  * option.
19708  */
19709 static int
19710 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
19711     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19712 {
19713         uint64_t loptval;
19714         int32_t error = 0, optval;
19715
19716         switch (sopt->sopt_name) {
19717         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
19718         /*  Pacing related ones */
19719         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
19720         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
19721         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
19722         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
19723         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
19724         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
19725         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
19726         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
19727         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
19728         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
19729         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
19730         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
19731         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
19732         case TCP_HDWR_RATE_CAP:                 /*  URL:hdwrcap boolean */
19733         case TCP_PACING_RATE_CAP:               /*  URL:cap  -- used by side-channel */
19734         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
19735        /* End pacing related */
19736         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
19737         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
19738         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
19739         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
19740         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
19741         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
19742         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
19743         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
19744         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
19745         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
19746         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
19747         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
19748         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
19749         case TCP_NO_PRR:                        /*  URL:noprr */
19750         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
19751         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
19752         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
19753         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
19754         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
19755         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
19756         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
19757         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
19758         case TCP_RACK_PROFILE:                  /*  URL:profile */
19759         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
19760         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
19761         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
19762         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
19763         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
19764         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
19765         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
19766         case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
19767                 break;
19768         default:
19769                 /* Filter off all unknown options to the base stack */
19770                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19771                 break;
19772         }
19773         INP_WUNLOCK(inp);
19774         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
19775                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
19776                 /*
19777                  * We truncate it down to 32 bits for the socket-option trace this
19778                  * means rates > 34Gbps won't show right, but thats probably ok.
19779                  */
19780                 optval = (uint32_t)loptval;
19781         } else {
19782                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
19783                 /* Save it in 64 bit form too */
19784                 loptval = optval;
19785         }
19786         if (error)
19787                 return (error);
19788         INP_WLOCK(inp);
19789         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
19790                 INP_WUNLOCK(inp);
19791                 return (ECONNRESET);
19792         }
19793         if (tp->t_fb != &__tcp_rack) {
19794                 INP_WUNLOCK(inp);
19795                 return (ENOPROTOOPT);
19796         }
19797         if (rack->defer_options && (rack->gp_ready == 0) &&
19798             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
19799             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
19800             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
19801             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
19802                 /* Options are beind deferred */
19803                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
19804                         INP_WUNLOCK(inp);
19805                         return (0);
19806                 } else {
19807                         /* No memory to defer, fail */
19808                         INP_WUNLOCK(inp);
19809                         return (ENOMEM);
19810                 }
19811         }
19812         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
19813         INP_WUNLOCK(inp);
19814         return (error);
19815 }
19816
19817 static void
19818 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
19819 {
19820
19821         INP_WLOCK_ASSERT(tp->t_inpcb);
19822         bzero(ti, sizeof(*ti));
19823
19824         ti->tcpi_state = tp->t_state;
19825         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
19826                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
19827         if (tp->t_flags & TF_SACK_PERMIT)
19828                 ti->tcpi_options |= TCPI_OPT_SACK;
19829         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
19830                 ti->tcpi_options |= TCPI_OPT_WSCALE;
19831                 ti->tcpi_snd_wscale = tp->snd_scale;
19832                 ti->tcpi_rcv_wscale = tp->rcv_scale;
19833         }
19834         if (tp->t_flags2 & TF2_ECN_PERMIT)
19835                 ti->tcpi_options |= TCPI_OPT_ECN;
19836         if (tp->t_flags & TF_FASTOPEN)
19837                 ti->tcpi_options |= TCPI_OPT_TFO;
19838         /* still kept in ticks is t_rcvtime */
19839         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
19840         /* Since we hold everything in precise useconds this is easy */
19841         ti->tcpi_rtt = tp->t_srtt;
19842         ti->tcpi_rttvar = tp->t_rttvar;
19843         ti->tcpi_rto = tp->t_rxtcur;
19844         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
19845         ti->tcpi_snd_cwnd = tp->snd_cwnd;
19846         /*
19847          * FreeBSD-specific extension fields for tcp_info.
19848          */
19849         ti->tcpi_rcv_space = tp->rcv_wnd;
19850         ti->tcpi_rcv_nxt = tp->rcv_nxt;
19851         ti->tcpi_snd_wnd = tp->snd_wnd;
19852         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
19853         ti->tcpi_snd_nxt = tp->snd_nxt;
19854         ti->tcpi_snd_mss = tp->t_maxseg;
19855         ti->tcpi_rcv_mss = tp->t_maxseg;
19856         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
19857         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
19858         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
19859 #ifdef NETFLIX_STATS
19860         ti->tcpi_total_tlp = tp->t_sndtlppack;
19861         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
19862         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
19863 #endif
19864 #ifdef TCP_OFFLOAD
19865         if (tp->t_flags & TF_TOE) {
19866                 ti->tcpi_options |= TCPI_OPT_TOE;
19867                 tcp_offload_tcp_info(tp, ti);
19868         }
19869 #endif
19870 }
19871
19872 static int
19873 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
19874     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19875 {
19876         int32_t error, optval;
19877         uint64_t val, loptval;
19878         struct  tcp_info ti;
19879         /*
19880          * Because all our options are either boolean or an int, we can just
19881          * pull everything into optval and then unlock and copy. If we ever
19882          * add a option that is not a int, then this will have quite an
19883          * impact to this routine.
19884          */
19885         error = 0;
19886         switch (sopt->sopt_name) {
19887         case TCP_INFO:
19888                 /* First get the info filled */
19889                 rack_fill_info(tp, &ti);
19890                 /* Fix up the rtt related fields if needed */
19891                 INP_WUNLOCK(inp);
19892                 error = sooptcopyout(sopt, &ti, sizeof ti);
19893                 return (error);
19894         /*
19895          * Beta is the congestion control value for NewReno that influences how
19896          * much of a backoff happens when loss is detected. It is normally set
19897          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
19898          * when you exit recovery.
19899          */
19900         case TCP_RACK_PACING_BETA:
19901                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19902                         error = EINVAL;
19903                 else if (rack->rc_pacing_cc_set == 0)
19904                         optval = rack->r_ctl.rc_saved_beta.beta;
19905                 else {
19906                         /*
19907                          * Reach out into the CC data and report back what
19908                          * I have previously set. Yeah it looks hackish but
19909                          * we don't want to report the saved values.
19910                          */
19911                         if (tp->ccv->cc_data)
19912                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta;
19913                         else
19914                                 error = EINVAL;
19915                 }
19916                 break;
19917                 /*
19918                  * Beta_ecn is the congestion control value for NewReno that influences how
19919                  * much of a backoff happens when a ECN mark is detected. It is normally set
19920                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
19921                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
19922                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
19923                  */
19924
19925         case TCP_RACK_PACING_BETA_ECN:
19926                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19927                         error = EINVAL;
19928                 else if (rack->rc_pacing_cc_set == 0)
19929                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
19930                 else {
19931                         /*
19932                          * Reach out into the CC data and report back what
19933                          * I have previously set. Yeah it looks hackish but
19934                          * we don't want to report the saved values.
19935                          */
19936                         if (tp->ccv->cc_data)
19937                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
19938                         else
19939                                 error = EINVAL;
19940                 }
19941                 break;
19942         case TCP_FAST_RSM_HACK:
19943                 optval = rack->fast_rsm_hack;
19944                 break;
19945         case TCP_DEFER_OPTIONS:
19946                 optval = rack->defer_options;
19947                 break;
19948         case TCP_RACK_MEASURE_CNT:
19949                 optval = rack->r_ctl.req_measurements;
19950                 break;
19951         case TCP_REC_ABC_VAL:
19952                 optval = rack->r_use_labc_for_rec;
19953                 break;
19954         case TCP_RACK_ABC_VAL:
19955                 optval = rack->rc_labc;
19956                 break;
19957         case TCP_HDWR_UP_ONLY:
19958                 optval= rack->r_up_only;
19959                 break;
19960         case TCP_PACING_RATE_CAP:
19961                 loptval = rack->r_ctl.bw_rate_cap;
19962                 break;
19963         case TCP_RACK_PROFILE:
19964                 /* You cannot retrieve a profile, its write only */
19965                 error = EINVAL;
19966                 break;
19967         case TCP_USE_CMP_ACKS:
19968                 optval = rack->r_use_cmp_ack;
19969                 break;
19970         case TCP_RACK_PACE_TO_FILL:
19971                 optval = rack->rc_pace_to_cwnd;
19972                 if (optval && rack->r_fill_less_agg)
19973                         optval++;
19974                 break;
19975         case TCP_RACK_NO_PUSH_AT_MAX:
19976                 optval = rack->r_ctl.rc_no_push_at_mrtt;
19977                 break;
19978         case TCP_SHARED_CWND_ENABLE:
19979                 optval = rack->rack_enable_scwnd;
19980                 break;
19981         case TCP_RACK_NONRXT_CFG_RATE:
19982                 optval = rack->rack_rec_nonrxt_use_cr;
19983                 break;
19984         case TCP_NO_PRR:
19985                 if (rack->rack_no_prr  == 1)
19986                         optval = 1;
19987                 else if (rack->no_prr_addback == 1)
19988                         optval = 2;
19989                 else
19990                         optval = 0;
19991                 break;
19992         case TCP_RACK_DO_DETECTION:
19993                 optval = rack->do_detection;
19994                 break;
19995         case TCP_RACK_MBUF_QUEUE:
19996                 /* Now do we use the LRO mbuf-queue feature */
19997                 optval = rack->r_mbuf_queue;
19998                 break;
19999         case TCP_TIMELY_DYN_ADJ:
20000                 optval = rack->rc_gp_dyn_mul;
20001                 break;
20002         case TCP_BBR_IWINTSO:
20003                 optval = rack->rc_init_win;
20004                 break;
20005         case TCP_RACK_TLP_REDUCE:
20006                 /* RACK TLP cwnd reduction (bool) */
20007                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20008                 break;
20009         case TCP_BBR_RACK_INIT_RATE:
20010                 val = rack->r_ctl.init_rate;
20011                 /* convert to kbits per sec */
20012                 val *= 8;
20013                 val /= 1000;
20014                 optval = (uint32_t)val;
20015                 break;
20016         case TCP_RACK_FORCE_MSEG:
20017                 optval = rack->rc_force_max_seg;
20018                 break;
20019         case TCP_RACK_PACE_MAX_SEG:
20020                 /* Max segments in a pace */
20021                 optval = rack->rc_user_set_max_segs;
20022                 break;
20023         case TCP_RACK_PACE_ALWAYS:
20024                 /* Use the always pace method */
20025                 optval = rack->rc_always_pace;
20026                 break;
20027         case TCP_RACK_PRR_SENDALOT:
20028                 /* Allow PRR to send more than one seg */
20029                 optval = rack->r_ctl.rc_prr_sendalot;
20030                 break;
20031         case TCP_RACK_MIN_TO:
20032                 /* Minimum time between rack t-o's in ms */
20033                 optval = rack->r_ctl.rc_min_to;
20034                 break;
20035         case TCP_RACK_EARLY_SEG:
20036                 /* If early recovery max segments */
20037                 optval = rack->r_ctl.rc_early_recovery_segs;
20038                 break;
20039         case TCP_RACK_REORD_THRESH:
20040                 /* RACK reorder threshold (shift amount) */
20041                 optval = rack->r_ctl.rc_reorder_shift;
20042                 break;
20043         case TCP_RACK_REORD_FADE:
20044                 /* Does reordering fade after ms time */
20045                 optval = rack->r_ctl.rc_reorder_fade;
20046                 break;
20047         case TCP_BBR_USE_RACK_RR:
20048                 /* Do we use the rack cheat for rxt */
20049                 optval = rack->use_rack_rr;
20050                 break;
20051         case TCP_RACK_RR_CONF:
20052                 optval = rack->r_rr_config;
20053                 break;
20054         case TCP_HDWR_RATE_CAP:
20055                 optval = rack->r_rack_hw_rate_caps;
20056                 break;
20057         case TCP_BBR_HDWR_PACE:
20058                 optval = rack->rack_hdw_pace_ena;
20059                 break;
20060         case TCP_RACK_TLP_THRESH:
20061                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
20062                 optval = rack->r_ctl.rc_tlp_threshold;
20063                 break;
20064         case TCP_RACK_PKT_DELAY:
20065                 /* RACK added ms i.e. rack-rtt + reord + N */
20066                 optval = rack->r_ctl.rc_pkt_delay;
20067                 break;
20068         case TCP_RACK_TLP_USE:
20069                 optval = rack->rack_tlp_threshold_use;
20070                 break;
20071         case TCP_RACK_PACE_RATE_CA:
20072                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20073                 break;
20074         case TCP_RACK_PACE_RATE_SS:
20075                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20076                 break;
20077         case TCP_RACK_PACE_RATE_REC:
20078                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20079                 break;
20080         case TCP_RACK_GP_INCREASE_SS:
20081                 optval = rack->r_ctl.rack_per_of_gp_ca;
20082                 break;
20083         case TCP_RACK_GP_INCREASE_CA:
20084                 optval = rack->r_ctl.rack_per_of_gp_ss;
20085                 break;
20086         case TCP_BBR_RACK_RTT_USE:
20087                 optval = rack->r_ctl.rc_rate_sample_method;
20088                 break;
20089         case TCP_DELACK:
20090                 optval = tp->t_delayed_ack;
20091                 break;
20092         case TCP_DATA_AFTER_CLOSE:
20093                 optval = rack->rc_allow_data_af_clo;
20094                 break;
20095         case TCP_SHARED_CWND_TIME_LIMIT:
20096                 optval = rack->r_limit_scw;
20097                 break;
20098         case TCP_RACK_TIMER_SLOP:
20099                 optval = rack->r_ctl.timer_slop;
20100                 break;
20101         default:
20102                 return (tcp_default_ctloutput(so, sopt, inp, tp));
20103                 break;
20104         }
20105         INP_WUNLOCK(inp);
20106         if (error == 0) {
20107                 if (TCP_PACING_RATE_CAP)
20108                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
20109                 else
20110                         error = sooptcopyout(sopt, &optval, sizeof optval);
20111         }
20112         return (error);
20113 }
20114
20115 static int
20116 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
20117 {
20118         int32_t error = EINVAL;
20119         struct tcp_rack *rack;
20120
20121         rack = (struct tcp_rack *)tp->t_fb_ptr;
20122         if (rack == NULL) {
20123                 /* Huh? */
20124                 goto out;
20125         }
20126         if (sopt->sopt_dir == SOPT_SET) {
20127                 return (rack_set_sockopt(so, sopt, inp, tp, rack));
20128         } else if (sopt->sopt_dir == SOPT_GET) {
20129                 return (rack_get_sockopt(so, sopt, inp, tp, rack));
20130         }
20131 out:
20132         INP_WUNLOCK(inp);
20133         return (error);
20134 }
20135
20136 static const char *rack_stack_names[] = {
20137         __XSTRING(STACKNAME),
20138 #ifdef STACKALIAS
20139         __XSTRING(STACKALIAS),
20140 #endif
20141 };
20142
20143 static int
20144 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20145 {
20146         memset(mem, 0, size);
20147         return (0);
20148 }
20149
20150 static void
20151 rack_dtor(void *mem, int32_t size, void *arg)
20152 {
20153
20154 }
20155
20156 static bool rack_mod_inited = false;
20157
20158 static int
20159 tcp_addrack(module_t mod, int32_t type, void *data)
20160 {
20161         int32_t err = 0;
20162         int num_stacks;
20163
20164         switch (type) {
20165         case MOD_LOAD:
20166                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20167                     sizeof(struct rack_sendmap),
20168                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20169
20170                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20171                     sizeof(struct tcp_rack),
20172                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20173
20174                 sysctl_ctx_init(&rack_sysctl_ctx);
20175                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20176                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20177                     OID_AUTO,
20178 #ifdef STACKALIAS
20179                     __XSTRING(STACKALIAS),
20180 #else
20181                     __XSTRING(STACKNAME),
20182 #endif
20183                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20184                     "");
20185                 if (rack_sysctl_root == NULL) {
20186                         printf("Failed to add sysctl node\n");
20187                         err = EFAULT;
20188                         goto free_uma;
20189                 }
20190                 rack_init_sysctls();
20191                 num_stacks = nitems(rack_stack_names);
20192                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20193                     rack_stack_names, &num_stacks);
20194                 if (err) {
20195                         printf("Failed to register %s stack name for "
20196                             "%s module\n", rack_stack_names[num_stacks],
20197                             __XSTRING(MODNAME));
20198                         sysctl_ctx_free(&rack_sysctl_ctx);
20199 free_uma:
20200                         uma_zdestroy(rack_zone);
20201                         uma_zdestroy(rack_pcb_zone);
20202                         rack_counter_destroy();
20203                         printf("Failed to register rack module -- err:%d\n", err);
20204                         return (err);
20205                 }
20206                 tcp_lro_reg_mbufq();
20207                 rack_mod_inited = true;
20208                 break;
20209         case MOD_QUIESCE:
20210                 err = deregister_tcp_functions(&__tcp_rack, true, false);
20211                 break;
20212         case MOD_UNLOAD:
20213                 err = deregister_tcp_functions(&__tcp_rack, false, true);
20214                 if (err == EBUSY)
20215                         break;
20216                 if (rack_mod_inited) {
20217                         uma_zdestroy(rack_zone);
20218                         uma_zdestroy(rack_pcb_zone);
20219                         sysctl_ctx_free(&rack_sysctl_ctx);
20220                         rack_counter_destroy();
20221                         rack_mod_inited = false;
20222                 }
20223                 tcp_lro_dereg_mbufq();
20224                 err = 0;
20225                 break;
20226         default:
20227                 return (EOPNOTSUPP);
20228         }
20229         return (err);
20230 }
20231
20232 static moduledata_t tcp_rack = {
20233         .name = __XSTRING(MODNAME),
20234         .evhand = tcp_addrack,
20235         .priv = 0
20236 };
20237
20238 MODULE_VERSION(MODNAME, 1);
20239 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20240 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);