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 <sys/param.h>
  36 #include <sys/arb.h>
  37 #include <sys/module.h>
  38 #include <sys/kernel.h>
  39 #ifdef TCP_HHOOK
  40 #include <sys/hhook.h>
  41 #endif
  42 #include <sys/lock.h>
  43 #include <sys/malloc.h>
  44 #include <sys/lock.h>
  45 #include <sys/mutex.h>
  46 #include <sys/mbuf.h>
  47 #include <sys/proc.h>           /* for proc0 declaration */
  48 #include <sys/socket.h>
  49 #include <sys/socketvar.h>
  50 #include <sys/sysctl.h>
  51 #include <sys/systm.h>
  52 #ifdef STATS
  53 #include <sys/qmath.h>
  54 #include <sys/tree.h>
  55 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
  56 #else
  57 #include <sys/tree.h>
  58 #endif
  59 #include <sys/refcount.h>
  60 #include <sys/queue.h>
  61 #include <sys/tim_filter.h>
  62 #include <sys/smp.h>
  63 #include <sys/kthread.h>
  64 #include <sys/kern_prefetch.h>
  65 #include <sys/protosw.h>
  66 #ifdef TCP_ACCOUNTING
  67 #include <sys/sched.h>
  68 #include <machine/cpu.h>
  69 #endif
  70 #include <vm/uma.h>
  71
  72 #include <net/route.h>
  73 #include <net/route/nhop.h>
  74 #include <net/vnet.h>
  75
  76 #define TCPSTATES               /* for logging */
  77
  78 #include <netinet/in.h>
  79 #include <netinet/in_kdtrace.h>
  80 #include <netinet/in_pcb.h>
  81 #include <netinet/ip.h>
  82 #include <netinet/ip_icmp.h>    /* required for icmp_var.h */
  83 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
  84 #include <netinet/ip_var.h>
  85 #include <netinet/ip6.h>
  86 #include <netinet6/in6_pcb.h>
  87 #include <netinet6/ip6_var.h>
  88 #include <netinet/tcp.h>
  89 #define TCPOUTFLAGS
  90 #include <netinet/tcp_fsm.h>
  91 #include <netinet/tcp_log_buf.h>
  92 #include <netinet/tcp_seq.h>
  93 #include <netinet/tcp_timer.h>
  94 #include <netinet/tcp_var.h>
  95 #include <netinet/tcp_hpts.h>
  96 #include <netinet/tcp_ratelimit.h>
  97 #include <netinet/tcp_accounting.h>
  98 #include <netinet/tcpip.h>
  99 #include <netinet/cc/cc.h>
 100 #include <netinet/cc/cc_newreno.h>
 101 #include <netinet/tcp_fastopen.h>
 102 #include <netinet/tcp_lro.h>
 103 #ifdef NETFLIX_SHARED_CWND
 104 #include <netinet/tcp_shared_cwnd.h>
 105 #endif
 106 #ifdef TCPDEBUG
 107 #include <netinet/tcp_debug.h>
 108 #endif                          /* TCPDEBUG */
 109 #ifdef TCP_OFFLOAD
 110 #include <netinet/tcp_offload.h>
 111 #endif
 112 #ifdef INET6
 113 #include <netinet6/tcp6_var.h>
 114 #endif
 115
 116 #include <netipsec/ipsec_support.h>
 117
 118 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 119 #include <netipsec/ipsec.h>
 120 #include <netipsec/ipsec6.h>
 121 #endif                          /* IPSEC */
 122
 123 #include <netinet/udp.h>
 124 #include <netinet/udp_var.h>
 125 #include <machine/in_cksum.h>
 126
 127 #ifdef MAC
 128 #include <security/mac/mac_framework.h>
 129 #endif
 130 #include "sack_filter.h"
 131 #include "tcp_rack.h"
 132 #include "rack_bbr_common.h"
 133
 134 uma_zone_t rack_zone;
 135 uma_zone_t rack_pcb_zone;
 136
 137 #ifndef TICKS2SBT
 138 #define TICKS2SBT(__t)  (tick_sbt * ((sbintime_t)(__t)))
 139 #endif
 140
 141 VNET_DECLARE(uint32_t, newreno_beta);
 142 VNET_DECLARE(uint32_t, newreno_beta_ecn);
 143 #define V_newreno_beta VNET(newreno_beta)
 144 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
 145
 146
 147 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
 148 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
 149
 150 struct sysctl_ctx_list rack_sysctl_ctx;
 151 struct sysctl_oid *rack_sysctl_root;
 152
 153 #define CUM_ACKED 1
 154 #define SACKED 2
 155
 156 /*
 157  * The RACK module incorporates a number of
 158  * TCP ideas that have been put out into the IETF
 159  * over the last few years:
 160  * - Matt Mathis's Rate Halving which slowly drops
 161  *    the congestion window so that the ack clock can
 162  *    be maintained during a recovery.
 163  * - Yuchung Cheng's RACK TCP (for which its named) that
 164  *    will stop us using the number of dup acks and instead
 165  *    use time as the gage of when we retransmit.
 166  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
 167  *    of Dukkipati et.al.
 168  * RACK depends on SACK, so if an endpoint arrives that
 169  * cannot do SACK the state machine below will shuttle the
 170  * connection back to using the "default" TCP stack that is
 171  * in FreeBSD.
 172  *
 173  * To implement RACK the original TCP stack was first decomposed
 174  * into a functional state machine with individual states
 175  * for each of the possible TCP connection states. The do_segement
 176  * functions role in life is to mandate the connection supports SACK
 177  * initially and then assure that the RACK state matches the conenction
 178  * state before calling the states do_segment function. Each
 179  * state is simplified due to the fact that the original do_segment
 180  * has been decomposed and we *know* what state we are in (no
 181  * switches on the state) and all tests for SACK are gone. This
 182  * greatly simplifies what each state does.
 183  *
 184  * TCP output is also over-written with a new version since it
 185  * must maintain the new rack scoreboard.
 186  *
 187  */
 188 static int32_t rack_tlp_thresh = 1;
 189 static int32_t rack_tlp_limit = 2;      /* No more than 2 TLPs w-out new data */
 190 static int32_t rack_tlp_use_greater = 1;
 191 static int32_t rack_reorder_thresh = 2;
 192 static int32_t rack_reorder_fade = 60000000;    /* 0 - never fade, def 60,000,000
 193                                                  * - 60 seconds */
 194 static uint8_t rack_req_measurements = 1;
 195 /* Attack threshold detections */
 196 static uint32_t rack_highest_sack_thresh_seen = 0;
 197 static uint32_t rack_highest_move_thresh_seen = 0;
 198 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
 199 static int32_t rack_hw_pace_extra_slots = 2;    /* 2 extra MSS time betweens */
 200 static int32_t rack_hw_rate_caps = 1; /* 1; */
 201 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
 202 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
 203 static int32_t rack_hw_up_only = 1;
 204 static int32_t rack_stats_gets_ms_rtt = 1;
 205 static int32_t rack_prr_addbackmax = 2;
 206
 207 static int32_t rack_pkt_delay = 1000;
 208 static int32_t rack_send_a_lot_in_prr = 1;
 209 static int32_t rack_min_to = 1000;      /* Number of microsecond  min timeout */
 210 static int32_t rack_verbose_logging = 0;
 211 static int32_t rack_ignore_data_after_close = 1;
 212 static int32_t rack_enable_shared_cwnd = 1;
 213 static int32_t rack_use_cmp_acks = 1;
 214 static int32_t rack_use_fsb = 1;
 215 static int32_t rack_use_rfo = 1;
 216 static int32_t rack_use_rsm_rfo = 1;
 217 static int32_t rack_max_abc_post_recovery = 2;
 218 static int32_t rack_client_low_buf = 0;
 219 #ifdef TCP_ACCOUNTING
 220 static int32_t rack_tcp_accounting = 0;
 221 #endif
 222 static int32_t rack_limits_scwnd = 1;
 223 static int32_t rack_enable_mqueue_for_nonpaced = 0;
 224 static int32_t rack_disable_prr = 0;
 225 static int32_t use_rack_rr = 1;
 226 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
 227 static int32_t rack_persist_min = 250000;       /* 250usec */
 228 static int32_t rack_persist_max = 2000000;      /* 2 Second in usec's */
 229 static int32_t rack_sack_not_required = 1;      /* set to one to allow non-sack to use rack */
 230 static int32_t rack_default_init_window = 0;    /* Use system default */
 231 static int32_t rack_limit_time_with_srtt = 0;
 232 static int32_t rack_autosndbuf_inc = 20;        /* In percentage form */
 233 static int32_t rack_enobuf_hw_boost_mult = 2;   /* How many times the hw rate we boost slot using time_between */
 234 static int32_t rack_enobuf_hw_max = 12000;      /* 12 ms in usecs */
 235 static int32_t rack_enobuf_hw_min = 10000;      /* 10 ms in usecs */
 236 static int32_t rack_hw_rwnd_factor = 2;         /* How many max_segs the rwnd must be before we hold off sending */
 237 /*
 238  * Currently regular tcp has a rto_min of 30ms
 239  * the backoff goes 12 times so that ends up
 240  * being a total of 122.850 seconds before a
 241  * connection is killed.
 242  */
 243 static uint32_t rack_def_data_window = 20;
 244 static uint32_t rack_goal_bdp = 2;
 245 static uint32_t rack_min_srtts = 1;
 246 static uint32_t rack_min_measure_usec = 0;
 247 static int32_t rack_tlp_min = 10000;    /* 10ms */
 248 static int32_t rack_rto_min = 30000;    /* 30,000 usec same as main freebsd */
 249 static int32_t rack_rto_max = 4000000;  /* 4 seconds in usec's */
 250 static const int32_t rack_free_cache = 2;
 251 static int32_t rack_hptsi_segments = 40;
 252 static int32_t rack_rate_sample_method = USE_RTT_LOW;
 253 static int32_t rack_pace_every_seg = 0;
 254 static int32_t rack_delayed_ack_time = 40000;   /* 40ms in usecs */
 255 static int32_t rack_slot_reduction = 4;
 256 static int32_t rack_wma_divisor = 8;            /* For WMA calculation */
 257 static int32_t rack_cwnd_block_ends_measure = 0;
 258 static int32_t rack_rwnd_block_ends_measure = 0;
 259 static int32_t rack_def_profile = 0;
 260
 261 static int32_t rack_lower_cwnd_at_tlp = 0;
 262 static int32_t rack_limited_retran = 0;
 263 static int32_t rack_always_send_oldest = 0;
 264 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
 265
 266 static uint16_t rack_per_of_gp_ss = 250;        /* 250 % slow-start */
 267 static uint16_t rack_per_of_gp_ca = 200;        /* 200 % congestion-avoidance */
 268 static uint16_t rack_per_of_gp_rec = 200;       /* 200 % of bw */
 269
 270 /* Probertt */
 271 static uint16_t rack_per_of_gp_probertt = 60;   /* 60% of bw */
 272 static uint16_t rack_per_of_gp_lowthresh = 40;  /* 40% is bottom */
 273 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
 274 static uint16_t rack_atexit_prtt_hbp = 130;     /* Clamp to 130% on exit prtt if highly buffered path */
 275 static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */
 276
 277 static uint32_t rack_max_drain_wait = 2;        /* How man gp srtt's before we give up draining */
 278 static uint32_t rack_must_drain = 1;            /* How many GP srtt's we *must* wait */
 279 static uint32_t rack_probertt_use_min_rtt_entry = 1;    /* Use the min to calculate the goal else gp_srtt */
 280 static uint32_t rack_probertt_use_min_rtt_exit = 0;
 281 static uint32_t rack_probe_rtt_sets_cwnd = 0;
 282 static uint32_t rack_probe_rtt_safety_val = 2000000;    /* No more than 2 sec in probe-rtt */
 283 static uint32_t rack_time_between_probertt = 9600000;   /* 9.6 sec in usecs */
 284 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;       /* How many srtt periods does probe-rtt last top fraction */
 285 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;       /* How many srtt periods does probe-rtt last bottom fraction */
 286 static uint32_t rack_min_probertt_hold = 40000;         /* Equal to delayed ack time */
 287 static uint32_t rack_probertt_filter_life = 10000000;
 288 static uint32_t rack_probertt_lower_within = 10;
 289 static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds)  to count as a lowering */
 290 static int32_t rack_pace_one_seg = 0;           /* Shall we pace for less than 1.4Meg 1MSS at a time */
 291 static int32_t rack_probertt_clear_is = 1;
 292 static int32_t rack_max_drain_hbp = 1;          /* Extra drain times gpsrtt for highly buffered paths */
 293 static int32_t rack_hbp_thresh = 3;             /* what is the divisor max_rtt/min_rtt to decided a hbp */
 294
 295 /* Part of pacing */
 296 static int32_t rack_max_per_above = 30;         /* When we go to increment stop if above 100+this% */
 297
 298 /* Timely information */
 299 /* Combine these two gives the range of 'no change' to bw */
 300 /* ie the up/down provide the upper and lower bound */
 301 static int32_t rack_gp_per_bw_mul_up = 2;       /* 2% */
 302 static int32_t rack_gp_per_bw_mul_down = 4;     /* 4% */
 303 static int32_t rack_gp_rtt_maxmul = 3;          /* 3 x maxmin */
 304 static int32_t rack_gp_rtt_minmul = 1;          /* minrtt + (minrtt/mindiv) is lower rtt */
 305 static int32_t rack_gp_rtt_mindiv = 4;          /* minrtt + (minrtt * minmul/mindiv) is lower rtt */
 306 static int32_t rack_gp_decrease_per = 20;       /* 20% decrease in multipler */
 307 static int32_t rack_gp_increase_per = 2;        /* 2% increase in multipler */
 308 static int32_t rack_per_lower_bound = 50;       /* Don't allow to drop below this multiplier */
 309 static int32_t rack_per_upper_bound_ss = 0;     /* Don't allow SS to grow above this */
 310 static int32_t rack_per_upper_bound_ca = 0;     /* Don't allow CA to grow above this */
 311 static int32_t rack_do_dyn_mul = 0;             /* Are the rack gp multipliers dynamic */
 312 static int32_t rack_gp_no_rec_chg = 1;          /* Prohibit recovery from reducing it's multiplier */
 313 static int32_t rack_timely_dec_clear = 6;       /* Do we clear decrement count at a value (6)? */
 314 static int32_t rack_timely_max_push_rise = 3;   /* One round of pushing */
 315 static int32_t rack_timely_max_push_drop = 3;   /* Three round of pushing */
 316 static int32_t rack_timely_min_segs = 4;        /* 4 segment minimum */
 317 static int32_t rack_use_max_for_nobackoff = 0;
 318 static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */
 319 static int32_t rack_timely_no_stopping = 0;
 320 static int32_t rack_down_raise_thresh = 100;
 321 static int32_t rack_req_segs = 1;
 322 static uint64_t rack_bw_rate_cap = 0;
 323
 324 /* Weird delayed ack mode */
 325 static int32_t rack_use_imac_dack = 0;
 326 /* Rack specific counters */
 327 counter_u64_t rack_badfr;
 328 counter_u64_t rack_badfr_bytes;
 329 counter_u64_t rack_rtm_prr_retran;
 330 counter_u64_t rack_rtm_prr_newdata;
 331 counter_u64_t rack_timestamp_mismatch;
 332 counter_u64_t rack_reorder_seen;
 333 counter_u64_t rack_paced_segments;
 334 counter_u64_t rack_unpaced_segments;
 335 counter_u64_t rack_calc_zero;
 336 counter_u64_t rack_calc_nonzero;
 337 counter_u64_t rack_saw_enobuf;
 338 counter_u64_t rack_saw_enobuf_hw;
 339 counter_u64_t rack_saw_enetunreach;
 340 counter_u64_t rack_per_timer_hole;
 341 counter_u64_t rack_large_ackcmp;
 342 counter_u64_t rack_small_ackcmp;
 343 #ifdef INVARIANTS
 344 counter_u64_t rack_adjust_map_bw;
 345 #endif
 346 /* Tail loss probe counters */
 347 counter_u64_t rack_tlp_tot;
 348 counter_u64_t rack_tlp_newdata;
 349 counter_u64_t rack_tlp_retran;
 350 counter_u64_t rack_tlp_retran_bytes;
 351 counter_u64_t rack_tlp_retran_fail;
 352 counter_u64_t rack_to_tot;
 353 counter_u64_t rack_to_arm_rack;
 354 counter_u64_t rack_to_arm_tlp;
 355 counter_u64_t rack_hot_alloc;
 356 counter_u64_t rack_to_alloc;
 357 counter_u64_t rack_to_alloc_hard;
 358 counter_u64_t rack_to_alloc_emerg;
 359 counter_u64_t rack_to_alloc_limited;
 360 counter_u64_t rack_alloc_limited_conns;
 361 counter_u64_t rack_split_limited;
 362
 363 #define MAX_NUM_OF_CNTS 13
 364 counter_u64_t rack_proc_comp_ack[MAX_NUM_OF_CNTS];
 365 counter_u64_t rack_multi_single_eq;
 366 counter_u64_t rack_proc_non_comp_ack;
 367
 368 counter_u64_t rack_fto_send;
 369 counter_u64_t rack_fto_rsm_send;
 370 counter_u64_t rack_nfto_resend;
 371 counter_u64_t rack_non_fto_send;
 372 counter_u64_t rack_extended_rfo;
 373
 374 counter_u64_t rack_sack_proc_all;
 375 counter_u64_t rack_sack_proc_short;
 376 counter_u64_t rack_sack_proc_restart;
 377 counter_u64_t rack_sack_attacks_detected;
 378 counter_u64_t rack_sack_attacks_reversed;
 379 counter_u64_t rack_sack_used_next_merge;
 380 counter_u64_t rack_sack_splits;
 381 counter_u64_t rack_sack_used_prev_merge;
 382 counter_u64_t rack_sack_skipped_acked;
 383 counter_u64_t rack_ack_total;
 384 counter_u64_t rack_express_sack;
 385 counter_u64_t rack_sack_total;
 386 counter_u64_t rack_move_none;
 387 counter_u64_t rack_move_some;
 388
 389 counter_u64_t rack_used_tlpmethod;
 390 counter_u64_t rack_used_tlpmethod2;
 391 counter_u64_t rack_enter_tlp_calc;
 392 counter_u64_t rack_input_idle_reduces;
 393 counter_u64_t rack_collapsed_win;
 394 counter_u64_t rack_tlp_does_nada;
 395 counter_u64_t rack_try_scwnd;
 396 counter_u64_t rack_hw_pace_init_fail;
 397 counter_u64_t rack_hw_pace_lost;
 398 counter_u64_t rack_sbsndptr_right;
 399 counter_u64_t rack_sbsndptr_wrong;
 400
 401 /* Temp CPU counters */
 402 counter_u64_t rack_find_high;
 403
 404 counter_u64_t rack_progress_drops;
 405 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
 406 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
 407
 408
 409 #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
 410
 411 #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {  \
 412         (tv) = (value) + slop;   \
 413         if ((u_long)(tv) < (u_long)(tvmin)) \
 414                 (tv) = (tvmin); \
 415         if ((u_long)(tv) > (u_long)(tvmax)) \
 416                 (tv) = (tvmax); \
 417 } while (0)
 418
 419 static void
 420 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
 421
 422 static int
 423 rack_process_ack(struct mbuf *m, struct tcphdr *th,
 424     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
 425     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
 426 static int
 427 rack_process_data(struct mbuf *m, struct tcphdr *th,
 428     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 429     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
 430 static void
 431 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
 432    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
 433 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
 434 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
 435     uint8_t limit_type);
 436 static struct rack_sendmap *
 437 rack_check_recovery_mode(struct tcpcb *tp,
 438     uint32_t tsused);
 439 static void
 440 rack_cong_signal(struct tcpcb *tp,
 441                  uint32_t type, uint32_t ack);
 442 static void rack_counter_destroy(void);
 443 static int
 444 rack_ctloutput(struct socket *so, struct sockopt *sopt,
 445     struct inpcb *inp, struct tcpcb *tp);
 446 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
 447 static void
 448 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
 449 static void
 450 rack_do_segment(struct mbuf *m, struct tcphdr *th,
 451     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
 452     uint8_t iptos);
 453 static void rack_dtor(void *mem, int32_t size, void *arg);
 454 static void
 455 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
 456     uint32_t flex1, uint32_t flex2,
 457     uint32_t flex3, uint32_t flex4,
 458     uint32_t flex5, uint32_t flex6,
 459     uint16_t flex7, uint8_t mod);
 460 static void
 461 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
 462    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm);
 463 static struct rack_sendmap *
 464 rack_find_high_nonack(struct tcp_rack *rack,
 465     struct rack_sendmap *rsm);
 466 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
 467 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
 468 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
 469 static int
 470 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
 471     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
 472 static void
 473 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
 474                             tcp_seq th_ack, int line);
 475 static uint32_t
 476 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
 477 static int32_t rack_handoff_ok(struct tcpcb *tp);
 478 static int32_t rack_init(struct tcpcb *tp);
 479 static void rack_init_sysctls(void);
 480 static void
 481 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
 482     struct tcphdr *th, int entered_rec, int dup_ack_struck);
 483 static void
 484 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
 485     uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t ts,
 486     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff);
 487
 488 static void
 489 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
 490     struct rack_sendmap *rsm);
 491 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
 492 static int32_t rack_output(struct tcpcb *tp);
 493
 494 static uint32_t
 495 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
 496     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
 497     uint32_t cts, int *moved_two);
 498 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
 499 static void rack_remxt_tmr(struct tcpcb *tp);
 500 static int
 501 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
 502     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
 503 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
 504 static int32_t rack_stopall(struct tcpcb *tp);
 505 static void
 506 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
 507     uint32_t delta);
 508 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
 509 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
 510 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
 511 static uint32_t
 512 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
 513     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
 514 static void
 515 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
 516     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
 517 static int
 518 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
 519     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
 520 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
 521 static int
 522 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
 523     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 524     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 525 static int
 526 rack_do_closing(struct mbuf *m, struct tcphdr *th,
 527     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 528     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 529 static int
 530 rack_do_established(struct mbuf *m, struct tcphdr *th,
 531     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 532     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 533 static int
 534 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
 535     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 536     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
 537 static int
 538 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
 539     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 540     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 541 static int
 542 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
 543     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 544     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 545 static int
 546 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
 547     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 548     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 549 static int
 550 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
 551     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 552     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 553 static int
 554 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
 555     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
 556     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
 557 struct rack_sendmap *
 558 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
 559     uint32_t tsused);
 560 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
 561     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
 562 static void
 563      tcp_rack_partialack(struct tcpcb *tp);
 564 static int
 565 rack_set_profile(struct tcp_rack *rack, int prof);
 566 static void
 567 rack_apply_deferred_options(struct tcp_rack *rack);
 568
 569 int32_t rack_clear_counter=0;
 570
 571 static void
 572 rack_set_cc_pacing(struct tcp_rack *rack)
 573 {
 574         struct sockopt sopt;
 575         struct cc_newreno_opts opt;
 576         struct newreno old, *ptr;
 577         struct tcpcb *tp;
 578         int error;
 579
 580         if (rack->rc_pacing_cc_set)
 581                 return;
 582
 583         tp = rack->rc_tp;
 584         if (tp->cc_algo == NULL) {
 585                 /* Tcb is leaving */
 586                 printf("No cc algorithm?\n");
 587                 return;
 588         }
 589         rack->rc_pacing_cc_set = 1;
 590         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 591                 /* Not new-reno we can't play games with beta! */
 592                 printf("cc_algo:%s is not NEWRENO:%s\n",
 593                        tp->cc_algo->name, CCALGONAME_NEWRENO);
 594                 goto out;
 595         }
 596         ptr = ((struct newreno *)tp->ccv->cc_data);
 597         if (CC_ALGO(tp)->ctl_output == NULL)  {
 598                 /* Huh, why does new_reno no longer have a set function? */
 599                 printf("no ctl_output for algo:%s\n", tp->cc_algo->name);
 600                 goto out;
 601         }
 602         if (ptr == NULL) {
 603                 /* Just the default values */
 604                 old.beta = V_newreno_beta_ecn;
 605                 old.beta_ecn = V_newreno_beta_ecn;
 606                 old.newreno_flags = 0;
 607         } else {
 608                 old.beta = ptr->beta;
 609                 old.beta_ecn = ptr->beta_ecn;
 610                 old.newreno_flags = ptr->newreno_flags;
 611         }
 612         sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
 613         sopt.sopt_dir = SOPT_SET;
 614         opt.name = CC_NEWRENO_BETA;
 615         opt.val = rack->r_ctl.rc_saved_beta.beta;
 616         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 617         if (error)  {
 618                 printf("Error returned by ctl_output %d\n", error);
 619                 goto out;
 620         }
 621         /*
 622          * Hack alert we need to set in our newreno_flags
 623          * so that Abe behavior is also applied.
 624          */
 625         ((struct newreno *)tp->ccv->cc_data)->newreno_flags = CC_NEWRENO_BETA_ECN;
 626         opt.name = CC_NEWRENO_BETA_ECN;
 627         opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
 628         error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
 629         if (error) {
 630                 printf("Error returned by ctl_output %d\n", error);
 631                 goto out;
 632         }
 633         /* Save off the original values for restoral */
 634         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 635 out:
 636         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 637                 union tcp_log_stackspecific log;
 638                 struct timeval tv;
 639
 640                 ptr = ((struct newreno *)tp->ccv->cc_data);
 641                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 642                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 643                 if (ptr) {
 644                         log.u_bbr.flex1 = ptr->beta;
 645                         log.u_bbr.flex2 = ptr->beta_ecn;
 646                         log.u_bbr.flex3 = ptr->newreno_flags;
 647                 }
 648                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 649                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 650                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 651                 log.u_bbr.flex7 = rack->gp_ready;
 652                 log.u_bbr.flex7 <<= 1;
 653                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 654                 log.u_bbr.flex7 <<= 1;
 655                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 656                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 657                 log.u_bbr.flex8 = 3;
 658                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
 659                                0, &log, false, NULL, NULL, 0, &tv);
 660         }
 661 }
 662
 663 static void
 664 rack_undo_cc_pacing(struct tcp_rack *rack)
 665 {
 666         struct newreno old, *ptr;
 667         struct tcpcb *tp;
 668
 669         if (rack->rc_pacing_cc_set == 0)
 670                 return;
 671         tp = rack->rc_tp;
 672         rack->rc_pacing_cc_set = 0;
 673         if (tp->cc_algo == NULL)
 674                 /* Tcb is leaving */
 675                 return;
 676         if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
 677                 /* Not new-reno nothing to do! */
 678                 return;
 679         }
 680         ptr = ((struct newreno *)tp->ccv->cc_data);
 681         if (ptr == NULL) {
 682                 /*
 683                  * This happens at rack_fini() if the
 684                  * cc module gets freed on us. In that
 685                  * case we loose our "new" settings but
 686                  * thats ok, since the tcb is going away anyway.
 687                  */
 688                 return;
 689         }
 690         /* Grab out our set values */
 691         memcpy(&old, ptr, sizeof(struct newreno));
 692         /* Copy back in the original values */
 693         memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
 694         /* Now save back the values we had set in (for when pacing is restored) */
 695         memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
 696         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
 697                 union tcp_log_stackspecific log;
 698                 struct timeval tv;
 699
 700                 ptr = ((struct newreno *)tp->ccv->cc_data);
 701                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
 702                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
 703                 log.u_bbr.flex1 = ptr->beta;
 704                 log.u_bbr.flex2 = ptr->beta_ecn;
 705                 log.u_bbr.flex3 = ptr->newreno_flags;
 706                 log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
 707                 log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
 708                 log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
 709                 log.u_bbr.flex7 = rack->gp_ready;
 710                 log.u_bbr.flex7 <<= 1;
 711                 log.u_bbr.flex7 |= rack->use_fixed_rate;
 712                 log.u_bbr.flex7 <<= 1;
 713                 log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
 714                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
 715                 log.u_bbr.flex8 = 4;
 716                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
 717                                0, &log, false, NULL, NULL, 0, &tv);
 718         }
 719 }
 720
 721 #ifdef NETFLIX_PEAKRATE
 722 static inline void
 723 rack_update_peakrate_thr(struct tcpcb *tp)
 724 {
 725         /* Keep in mind that t_maxpeakrate is in B/s. */
 726         uint64_t peak;
 727         peak = uqmax((tp->t_maxseg * 2),
 728                      (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
 729         tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
 730 }
 731 #endif
 732
 733 static int
 734 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
 735 {
 736         uint32_t stat;
 737         int32_t error;
 738         int i;
 739
 740         error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
 741         if (error || req->newptr == NULL)
 742                 return error;
 743
 744         error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
 745         if (error)
 746                 return (error);
 747         if (stat == 1) {
 748 #ifdef INVARIANTS
 749                 printf("Clearing RACK counters\n");
 750 #endif
 751                 counter_u64_zero(rack_badfr);
 752                 counter_u64_zero(rack_badfr_bytes);
 753                 counter_u64_zero(rack_rtm_prr_retran);
 754                 counter_u64_zero(rack_rtm_prr_newdata);
 755                 counter_u64_zero(rack_timestamp_mismatch);
 756                 counter_u64_zero(rack_reorder_seen);
 757                 counter_u64_zero(rack_tlp_tot);
 758                 counter_u64_zero(rack_tlp_newdata);
 759                 counter_u64_zero(rack_tlp_retran);
 760                 counter_u64_zero(rack_tlp_retran_bytes);
 761                 counter_u64_zero(rack_tlp_retran_fail);
 762                 counter_u64_zero(rack_to_tot);
 763                 counter_u64_zero(rack_to_arm_rack);
 764                 counter_u64_zero(rack_to_arm_tlp);
 765                 counter_u64_zero(rack_paced_segments);
 766                 counter_u64_zero(rack_calc_zero);
 767                 counter_u64_zero(rack_calc_nonzero);
 768                 counter_u64_zero(rack_unpaced_segments);
 769                 counter_u64_zero(rack_saw_enobuf);
 770                 counter_u64_zero(rack_saw_enobuf_hw);
 771                 counter_u64_zero(rack_saw_enetunreach);
 772                 counter_u64_zero(rack_per_timer_hole);
 773                 counter_u64_zero(rack_large_ackcmp);
 774                 counter_u64_zero(rack_small_ackcmp);
 775 #ifdef INVARIANTS
 776                 counter_u64_zero(rack_adjust_map_bw);
 777 #endif
 778                 counter_u64_zero(rack_to_alloc_hard);
 779                 counter_u64_zero(rack_to_alloc_emerg);
 780                 counter_u64_zero(rack_sack_proc_all);
 781                 counter_u64_zero(rack_fto_send);
 782                 counter_u64_zero(rack_fto_rsm_send);
 783                 counter_u64_zero(rack_extended_rfo);
 784                 counter_u64_zero(rack_hw_pace_init_fail);
 785                 counter_u64_zero(rack_hw_pace_lost);
 786                 counter_u64_zero(rack_sbsndptr_wrong);
 787                 counter_u64_zero(rack_sbsndptr_right);
 788                 counter_u64_zero(rack_non_fto_send);
 789                 counter_u64_zero(rack_nfto_resend);
 790                 counter_u64_zero(rack_sack_proc_short);
 791                 counter_u64_zero(rack_sack_proc_restart);
 792                 counter_u64_zero(rack_to_alloc);
 793                 counter_u64_zero(rack_to_alloc_limited);
 794                 counter_u64_zero(rack_alloc_limited_conns);
 795                 counter_u64_zero(rack_split_limited);
 796                 for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
 797                         counter_u64_zero(rack_proc_comp_ack[i]);
 798                 }
 799                 counter_u64_zero(rack_multi_single_eq);
 800                 counter_u64_zero(rack_proc_non_comp_ack);
 801                 counter_u64_zero(rack_find_high);
 802                 counter_u64_zero(rack_sack_attacks_detected);
 803                 counter_u64_zero(rack_sack_attacks_reversed);
 804                 counter_u64_zero(rack_sack_used_next_merge);
 805                 counter_u64_zero(rack_sack_used_prev_merge);
 806                 counter_u64_zero(rack_sack_splits);
 807                 counter_u64_zero(rack_sack_skipped_acked);
 808                 counter_u64_zero(rack_ack_total);
 809                 counter_u64_zero(rack_express_sack);
 810                 counter_u64_zero(rack_sack_total);
 811                 counter_u64_zero(rack_move_none);
 812                 counter_u64_zero(rack_move_some);
 813                 counter_u64_zero(rack_used_tlpmethod);
 814                 counter_u64_zero(rack_used_tlpmethod2);
 815                 counter_u64_zero(rack_enter_tlp_calc);
 816                 counter_u64_zero(rack_progress_drops);
 817                 counter_u64_zero(rack_tlp_does_nada);
 818                 counter_u64_zero(rack_try_scwnd);
 819                 counter_u64_zero(rack_collapsed_win);
 820         }
 821         rack_clear_counter = 0;
 822         return (0);
 823 }
 824
 825 static void
 826 rack_init_sysctls(void)
 827 {
 828         int i;
 829         struct sysctl_oid *rack_counters;
 830         struct sysctl_oid *rack_attack;
 831         struct sysctl_oid *rack_pacing;
 832         struct sysctl_oid *rack_timely;
 833         struct sysctl_oid *rack_timers;
 834         struct sysctl_oid *rack_tlp;
 835         struct sysctl_oid *rack_misc;
 836         struct sysctl_oid *rack_measure;
 837         struct sysctl_oid *rack_probertt;
 838         struct sysctl_oid *rack_hw_pacing;
 839
 840         rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 841             SYSCTL_CHILDREN(rack_sysctl_root),
 842             OID_AUTO,
 843             "sack_attack",
 844             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 845             "Rack Sack Attack Counters and Controls");
 846         rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 847             SYSCTL_CHILDREN(rack_sysctl_root),
 848             OID_AUTO,
 849             "stats",
 850             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 851             "Rack Counters");
 852         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 853             SYSCTL_CHILDREN(rack_sysctl_root),
 854             OID_AUTO, "rate_sample_method", CTLFLAG_RW,
 855             &rack_rate_sample_method , USE_RTT_LOW,
 856             "What method should we use for rate sampling 0=high, 1=low ");
 857         /* Probe rtt related controls */
 858         rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 859             SYSCTL_CHILDREN(rack_sysctl_root),
 860             OID_AUTO,
 861             "probertt",
 862             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 863             "ProbeRTT related Controls");
 864         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 865             SYSCTL_CHILDREN(rack_probertt),
 866             OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
 867             &rack_atexit_prtt_hbp, 130,
 868             "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
 869         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 870             SYSCTL_CHILDREN(rack_probertt),
 871             OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
 872             &rack_atexit_prtt, 130,
 873             "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
 874         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 875             SYSCTL_CHILDREN(rack_probertt),
 876             OID_AUTO, "gp_per_mul", CTLFLAG_RW,
 877             &rack_per_of_gp_probertt, 60,
 878             "What percentage of goodput do we pace at in probertt");
 879         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 880             SYSCTL_CHILDREN(rack_probertt),
 881             OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
 882             &rack_per_of_gp_probertt_reduce, 10,
 883             "What percentage of goodput do we reduce every gp_srtt");
 884         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 885             SYSCTL_CHILDREN(rack_probertt),
 886             OID_AUTO, "gp_per_low", CTLFLAG_RW,
 887             &rack_per_of_gp_lowthresh, 40,
 888             "What percentage of goodput do we allow the multiplier to fall to");
 889         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 890             SYSCTL_CHILDREN(rack_probertt),
 891             OID_AUTO, "time_between", CTLFLAG_RW,
 892             & rack_time_between_probertt, 96000000,
 893             "How many useconds between the lowest rtt falling must past before we enter probertt");
 894         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 895             SYSCTL_CHILDREN(rack_probertt),
 896             OID_AUTO, "safety", CTLFLAG_RW,
 897             &rack_probe_rtt_safety_val, 2000000,
 898             "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
 899         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 900             SYSCTL_CHILDREN(rack_probertt),
 901             OID_AUTO, "sets_cwnd", CTLFLAG_RW,
 902             &rack_probe_rtt_sets_cwnd, 0,
 903             "Do we set the cwnd too (if always_lower is on)");
 904         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 905             SYSCTL_CHILDREN(rack_probertt),
 906             OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
 907             &rack_max_drain_wait, 2,
 908             "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
 909         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 910             SYSCTL_CHILDREN(rack_probertt),
 911             OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
 912             &rack_must_drain, 1,
 913             "We must drain this many gp_srtt's waiting for flight to reach goal");
 914         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 915             SYSCTL_CHILDREN(rack_probertt),
 916             OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
 917             &rack_probertt_use_min_rtt_entry, 1,
 918             "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
 919         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 920             SYSCTL_CHILDREN(rack_probertt),
 921             OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
 922             &rack_probertt_use_min_rtt_exit, 0,
 923             "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
 924         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 925             SYSCTL_CHILDREN(rack_probertt),
 926             OID_AUTO, "length_div", CTLFLAG_RW,
 927             &rack_probertt_gpsrtt_cnt_div, 0,
 928             "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
 929         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 930             SYSCTL_CHILDREN(rack_probertt),
 931             OID_AUTO, "length_mul", CTLFLAG_RW,
 932             &rack_probertt_gpsrtt_cnt_mul, 0,
 933             "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
 934         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 935             SYSCTL_CHILDREN(rack_probertt),
 936             OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
 937             &rack_min_probertt_hold, 200000,
 938             "What is the minimum time we hold probertt at target");
 939         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 940             SYSCTL_CHILDREN(rack_probertt),
 941             OID_AUTO, "filter_life", CTLFLAG_RW,
 942             &rack_probertt_filter_life, 10000000,
 943             "What is the time for the filters life in useconds");
 944         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 945             SYSCTL_CHILDREN(rack_probertt),
 946             OID_AUTO, "lower_within", CTLFLAG_RW,
 947             &rack_probertt_lower_within, 10,
 948             "If the rtt goes lower within this percentage of the time, go into probe-rtt");
 949         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 950             SYSCTL_CHILDREN(rack_probertt),
 951             OID_AUTO, "must_move", CTLFLAG_RW,
 952             &rack_min_rtt_movement, 250,
 953             "How much is the minimum movement in rtt to count as a drop for probertt purposes");
 954         SYSCTL_ADD_U32(&rack_sysctl_ctx,
 955             SYSCTL_CHILDREN(rack_probertt),
 956             OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
 957             &rack_probertt_clear_is, 1,
 958             "Do we clear I/S counts on exiting probe-rtt");
 959         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 960             SYSCTL_CHILDREN(rack_probertt),
 961             OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
 962             &rack_max_drain_hbp, 1,
 963             "How many extra drain gpsrtt's do we get in highly buffered paths");
 964         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 965             SYSCTL_CHILDREN(rack_probertt),
 966             OID_AUTO, "hbp_threshold", CTLFLAG_RW,
 967             &rack_hbp_thresh, 3,
 968             "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
 969         /* Pacing related sysctls */
 970         rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
 971             SYSCTL_CHILDREN(rack_sysctl_root),
 972             OID_AUTO,
 973             "pacing",
 974             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 975             "Pacing related Controls");
 976         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 977             SYSCTL_CHILDREN(rack_pacing),
 978             OID_AUTO, "max_pace_over", CTLFLAG_RW,
 979             &rack_max_per_above, 30,
 980             "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
 981         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 982             SYSCTL_CHILDREN(rack_pacing),
 983             OID_AUTO, "pace_to_one", CTLFLAG_RW,
 984             &rack_pace_one_seg, 0,
 985             "Do we allow low b/w pacing of 1MSS instead of two");
 986         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 987             SYSCTL_CHILDREN(rack_pacing),
 988             OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
 989             &rack_limit_time_with_srtt, 0,
 990             "Do we limit pacing time based on srtt");
 991         SYSCTL_ADD_S32(&rack_sysctl_ctx,
 992             SYSCTL_CHILDREN(rack_pacing),
 993             OID_AUTO, "init_win", CTLFLAG_RW,
 994             &rack_default_init_window, 0,
 995             "Do we have a rack initial window 0 = system default");
 996         SYSCTL_ADD_U16(&rack_sysctl_ctx,
 997             SYSCTL_CHILDREN(rack_pacing),
 998             OID_AUTO, "gp_per_ss", CTLFLAG_RW,
 999             &rack_per_of_gp_ss, 250,
1000             "If non zero, what percentage of goodput to pace at in slow start");
1001         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1002             SYSCTL_CHILDREN(rack_pacing),
1003             OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1004             &rack_per_of_gp_ca, 150,
1005             "If non zero, what percentage of goodput to pace at in congestion avoidance");
1006         SYSCTL_ADD_U16(&rack_sysctl_ctx,
1007             SYSCTL_CHILDREN(rack_pacing),
1008             OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1009             &rack_per_of_gp_rec, 200,
1010             "If non zero, what percentage of goodput to pace at in recovery");
1011         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1012             SYSCTL_CHILDREN(rack_pacing),
1013             OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1014             &rack_hptsi_segments, 40,
1015             "What size is the max for TSO segments in pacing and burst mitigation");
1016         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1017             SYSCTL_CHILDREN(rack_pacing),
1018             OID_AUTO, "burst_reduces", CTLFLAG_RW,
1019             &rack_slot_reduction, 4,
1020             "When doing only burst mitigation what is the reduce divisor");
1021         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1022             SYSCTL_CHILDREN(rack_sysctl_root),
1023             OID_AUTO, "use_pacing", CTLFLAG_RW,
1024             &rack_pace_every_seg, 0,
1025             "If set we use pacing, if clear we use only the original burst mitigation");
1026         SYSCTL_ADD_U64(&rack_sysctl_ctx,
1027             SYSCTL_CHILDREN(rack_pacing),
1028             OID_AUTO, "rate_cap", CTLFLAG_RW,
1029             &rack_bw_rate_cap, 0,
1030             "If set we apply this value to the absolute rate cap used by pacing");
1031         SYSCTL_ADD_U8(&rack_sysctl_ctx,
1032             SYSCTL_CHILDREN(rack_sysctl_root),
1033             OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1034             &rack_req_measurements, 1,
1035             "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1036         /* Hardware pacing */
1037         rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1038             SYSCTL_CHILDREN(rack_sysctl_root),
1039             OID_AUTO,
1040             "hdwr_pacing",
1041             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1042             "Pacing related Controls");
1043         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1044             SYSCTL_CHILDREN(rack_hw_pacing),
1045             OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1046             &rack_hw_rwnd_factor, 2,
1047             "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1048         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1049             SYSCTL_CHILDREN(rack_hw_pacing),
1050             OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1051             &rack_enobuf_hw_boost_mult, 2,
1052             "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1053         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1054             SYSCTL_CHILDREN(rack_hw_pacing),
1055             OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1056             &rack_enobuf_hw_max, 2,
1057             "What is the max boost the pacing time if we see a ENOBUFS?");
1058         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1059             SYSCTL_CHILDREN(rack_hw_pacing),
1060             OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1061             &rack_enobuf_hw_min, 2,
1062             "What is the min boost the pacing time if we see a ENOBUFS?");
1063         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1064             SYSCTL_CHILDREN(rack_hw_pacing),
1065             OID_AUTO, "enable", CTLFLAG_RW,
1066             &rack_enable_hw_pacing, 0,
1067             "Should RACK attempt to use hw pacing?");
1068         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1069             SYSCTL_CHILDREN(rack_hw_pacing),
1070             OID_AUTO, "rate_cap", CTLFLAG_RW,
1071             &rack_hw_rate_caps, 1,
1072             "Does the highest hardware pacing rate cap the rate we will send at??");
1073         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1074             SYSCTL_CHILDREN(rack_hw_pacing),
1075             OID_AUTO, "rate_min", CTLFLAG_RW,
1076             &rack_hw_rate_min, 0,
1077             "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1078         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1079             SYSCTL_CHILDREN(rack_hw_pacing),
1080             OID_AUTO, "rate_to_low", CTLFLAG_RW,
1081             &rack_hw_rate_to_low, 0,
1082             "If we fall below this rate, dis-engage hw pacing?");
1083         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1084             SYSCTL_CHILDREN(rack_hw_pacing),
1085             OID_AUTO, "up_only", CTLFLAG_RW,
1086             &rack_hw_up_only, 1,
1087             "Do we allow hw pacing to lower the rate selected?");
1088         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1089             SYSCTL_CHILDREN(rack_hw_pacing),
1090             OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1091             &rack_hw_pace_extra_slots, 2,
1092             "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1093         rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1094             SYSCTL_CHILDREN(rack_sysctl_root),
1095             OID_AUTO,
1096             "timely",
1097             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1098             "Rack Timely RTT Controls");
1099         /* Timely based GP dynmics */
1100         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1101             SYSCTL_CHILDREN(rack_timely),
1102             OID_AUTO, "upper", CTLFLAG_RW,
1103             &rack_gp_per_bw_mul_up, 2,
1104             "Rack timely upper range for equal b/w (in percentage)");
1105         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1106             SYSCTL_CHILDREN(rack_timely),
1107             OID_AUTO, "lower", CTLFLAG_RW,
1108             &rack_gp_per_bw_mul_down, 4,
1109             "Rack timely lower range for equal b/w (in percentage)");
1110         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1111             SYSCTL_CHILDREN(rack_timely),
1112             OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1113             &rack_gp_rtt_maxmul, 3,
1114             "Rack timely multipler of lowest rtt for rtt_max");
1115         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1116             SYSCTL_CHILDREN(rack_timely),
1117             OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1118             &rack_gp_rtt_mindiv, 4,
1119             "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1120         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1121             SYSCTL_CHILDREN(rack_timely),
1122             OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1123             &rack_gp_rtt_minmul, 1,
1124             "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1125         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1126             SYSCTL_CHILDREN(rack_timely),
1127             OID_AUTO, "decrease", CTLFLAG_RW,
1128             &rack_gp_decrease_per, 20,
1129             "Rack timely decrease percentage of our GP multiplication factor");
1130         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1131             SYSCTL_CHILDREN(rack_timely),
1132             OID_AUTO, "increase", CTLFLAG_RW,
1133             &rack_gp_increase_per, 2,
1134             "Rack timely increase perentage of our GP multiplication factor");
1135         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1136             SYSCTL_CHILDREN(rack_timely),
1137             OID_AUTO, "lowerbound", CTLFLAG_RW,
1138             &rack_per_lower_bound, 50,
1139             "Rack timely lowest percentage we allow GP multiplier to fall to");
1140         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1141             SYSCTL_CHILDREN(rack_timely),
1142             OID_AUTO, "upperboundss", CTLFLAG_RW,
1143             &rack_per_upper_bound_ss, 0,
1144             "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1145         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1146             SYSCTL_CHILDREN(rack_timely),
1147             OID_AUTO, "upperboundca", CTLFLAG_RW,
1148             &rack_per_upper_bound_ca, 0,
1149             "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1150         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1151             SYSCTL_CHILDREN(rack_timely),
1152             OID_AUTO, "dynamicgp", CTLFLAG_RW,
1153             &rack_do_dyn_mul, 0,
1154             "Rack timely do we enable dynmaic timely goodput by default");
1155         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1156             SYSCTL_CHILDREN(rack_timely),
1157             OID_AUTO, "no_rec_red", CTLFLAG_RW,
1158             &rack_gp_no_rec_chg, 1,
1159             "Rack timely do we prohibit the recovery multiplier from being lowered");
1160         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1161             SYSCTL_CHILDREN(rack_timely),
1162             OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1163             &rack_timely_dec_clear, 6,
1164             "Rack timely what threshold do we count to before another boost during b/w decent");
1165         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1166             SYSCTL_CHILDREN(rack_timely),
1167             OID_AUTO, "max_push_rise", CTLFLAG_RW,
1168             &rack_timely_max_push_rise, 3,
1169             "Rack timely how many times do we push up with b/w increase");
1170         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1171             SYSCTL_CHILDREN(rack_timely),
1172             OID_AUTO, "max_push_drop", CTLFLAG_RW,
1173             &rack_timely_max_push_drop, 3,
1174             "Rack timely how many times do we push back on b/w decent");
1175         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1176             SYSCTL_CHILDREN(rack_timely),
1177             OID_AUTO, "min_segs", CTLFLAG_RW,
1178             &rack_timely_min_segs, 4,
1179             "Rack timely when setting the cwnd what is the min num segments");
1180         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1181             SYSCTL_CHILDREN(rack_timely),
1182             OID_AUTO, "noback_max", CTLFLAG_RW,
1183             &rack_use_max_for_nobackoff, 0,
1184             "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1185         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1186             SYSCTL_CHILDREN(rack_timely),
1187             OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1188             &rack_timely_int_timely_only, 0,
1189             "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1190         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1191             SYSCTL_CHILDREN(rack_timely),
1192             OID_AUTO, "nonstop", CTLFLAG_RW,
1193             &rack_timely_no_stopping, 0,
1194             "Rack timely don't stop increase");
1195         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1196             SYSCTL_CHILDREN(rack_timely),
1197             OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1198             &rack_down_raise_thresh, 100,
1199             "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1200         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1201             SYSCTL_CHILDREN(rack_timely),
1202             OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1203             &rack_req_segs, 1,
1204             "Bottom dragging if not these many segments outstanding and room");
1205
1206         /* TLP and Rack related parameters */
1207         rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1208             SYSCTL_CHILDREN(rack_sysctl_root),
1209             OID_AUTO,
1210             "tlp",
1211             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1212             "TLP and Rack related Controls");
1213         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1214             SYSCTL_CHILDREN(rack_tlp),
1215             OID_AUTO, "use_rrr", CTLFLAG_RW,
1216             &use_rack_rr, 1,
1217             "Do we use Rack Rapid Recovery");
1218         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1219             SYSCTL_CHILDREN(rack_tlp),
1220             OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1221             &rack_max_abc_post_recovery, 2,
1222             "Since we do early recovery, do we override the l_abc to a value, if so what?");
1223         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1224             SYSCTL_CHILDREN(rack_tlp),
1225             OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1226             &rack_non_rxt_use_cr, 0,
1227             "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1228         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229             SYSCTL_CHILDREN(rack_tlp),
1230             OID_AUTO, "tlpmethod", CTLFLAG_RW,
1231             &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1232             "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1233         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1234             SYSCTL_CHILDREN(rack_tlp),
1235             OID_AUTO, "limit", CTLFLAG_RW,
1236             &rack_tlp_limit, 2,
1237             "How many TLP's can be sent without sending new data");
1238         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1239             SYSCTL_CHILDREN(rack_tlp),
1240             OID_AUTO, "use_greater", CTLFLAG_RW,
1241             &rack_tlp_use_greater, 1,
1242             "Should we use the rack_rtt time if its greater than srtt");
1243         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1244             SYSCTL_CHILDREN(rack_tlp),
1245             OID_AUTO, "tlpminto", CTLFLAG_RW,
1246             &rack_tlp_min, 10000,
1247             "TLP minimum timeout per the specification (in microseconds)");
1248         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1249             SYSCTL_CHILDREN(rack_tlp),
1250             OID_AUTO, "send_oldest", CTLFLAG_RW,
1251             &rack_always_send_oldest, 0,
1252             "Should we always send the oldest TLP and RACK-TLP");
1253         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1254             SYSCTL_CHILDREN(rack_tlp),
1255             OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1256             &rack_limited_retran, 0,
1257             "How many times can a rack timeout drive out sends");
1258         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1259             SYSCTL_CHILDREN(rack_tlp),
1260             OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1261             &rack_lower_cwnd_at_tlp, 0,
1262             "When a TLP completes a retran should we enter recovery");
1263         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1264             SYSCTL_CHILDREN(rack_tlp),
1265             OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1266             &rack_reorder_thresh, 2,
1267             "What factor for rack will be added when seeing reordering (shift right)");
1268         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1269             SYSCTL_CHILDREN(rack_tlp),
1270             OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1271             &rack_tlp_thresh, 1,
1272             "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1273         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1274             SYSCTL_CHILDREN(rack_tlp),
1275             OID_AUTO, "reorder_fade", CTLFLAG_RW,
1276             &rack_reorder_fade, 60000000,
1277             "Does reorder detection fade, if so how many microseconds (0 means never)");
1278         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1279             SYSCTL_CHILDREN(rack_tlp),
1280             OID_AUTO, "pktdelay", CTLFLAG_RW,
1281             &rack_pkt_delay, 1000,
1282             "Extra RACK time (in microseconds) besides reordering thresh");
1283
1284         /* Timer related controls */
1285         rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1286             SYSCTL_CHILDREN(rack_sysctl_root),
1287             OID_AUTO,
1288             "timers",
1289             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1290             "Timer related controls");
1291         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1292             SYSCTL_CHILDREN(rack_timers),
1293             OID_AUTO, "persmin", CTLFLAG_RW,
1294             &rack_persist_min, 250000,
1295             "What is the minimum time in microseconds between persists");
1296         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1297             SYSCTL_CHILDREN(rack_timers),
1298             OID_AUTO, "persmax", CTLFLAG_RW,
1299             &rack_persist_max, 2000000,
1300             "What is the largest delay in microseconds between persists");
1301         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1302             SYSCTL_CHILDREN(rack_timers),
1303             OID_AUTO, "delayed_ack", CTLFLAG_RW,
1304             &rack_delayed_ack_time, 40000,
1305             "Delayed ack time (40ms in microseconds)");
1306         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1307             SYSCTL_CHILDREN(rack_timers),
1308             OID_AUTO, "minrto", CTLFLAG_RW,
1309             &rack_rto_min, 30000,
1310             "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1311         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1312             SYSCTL_CHILDREN(rack_timers),
1313             OID_AUTO, "maxrto", CTLFLAG_RW,
1314             &rack_rto_max, 4000000,
1315             "Maxiumum RTO in microseconds -- should be at least as large as min_rto");
1316         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1317             SYSCTL_CHILDREN(rack_timers),
1318             OID_AUTO, "minto", CTLFLAG_RW,
1319             &rack_min_to, 1000,
1320             "Minimum rack timeout in microseconds");
1321         /* Measure controls */
1322         rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1323             SYSCTL_CHILDREN(rack_sysctl_root),
1324             OID_AUTO,
1325             "measure",
1326             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1327             "Measure related controls");
1328         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1329             SYSCTL_CHILDREN(rack_measure),
1330             OID_AUTO, "wma_divisor", CTLFLAG_RW,
1331             &rack_wma_divisor, 8,
1332             "When doing b/w calculation what is the  divisor for the WMA");
1333         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1334             SYSCTL_CHILDREN(rack_measure),
1335             OID_AUTO, "end_cwnd", CTLFLAG_RW,
1336             &rack_cwnd_block_ends_measure, 0,
1337             "Does a cwnd just-return end the measurement window (app limited)");
1338         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1339             SYSCTL_CHILDREN(rack_measure),
1340             OID_AUTO, "end_rwnd", CTLFLAG_RW,
1341             &rack_rwnd_block_ends_measure, 0,
1342             "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1343         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1344             SYSCTL_CHILDREN(rack_measure),
1345             OID_AUTO, "min_target", CTLFLAG_RW,
1346             &rack_def_data_window, 20,
1347             "What is the minimum target window (in mss) for a GP measurements");
1348         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1349             SYSCTL_CHILDREN(rack_measure),
1350             OID_AUTO, "goal_bdp", CTLFLAG_RW,
1351             &rack_goal_bdp, 2,
1352             "What is the goal BDP to measure");
1353         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1354             SYSCTL_CHILDREN(rack_measure),
1355             OID_AUTO, "min_srtts", CTLFLAG_RW,
1356             &rack_min_srtts, 1,
1357             "What is the goal BDP to measure");
1358         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1359             SYSCTL_CHILDREN(rack_measure),
1360             OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1361             &rack_min_measure_usec, 0,
1362             "What is the Minimum time time for a measurement if 0, this is off");
1363         /* Misc rack controls */
1364         rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1365             SYSCTL_CHILDREN(rack_sysctl_root),
1366             OID_AUTO,
1367             "misc",
1368             CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1369             "Misc related controls");
1370 #ifdef TCP_ACCOUNTING
1371         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1372             SYSCTL_CHILDREN(rack_misc),
1373             OID_AUTO, "tcp_acct", CTLFLAG_RW,
1374             &rack_tcp_accounting, 0,
1375             "Should we turn on TCP accounting for all rack sessions?");
1376 #endif
1377         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1378             SYSCTL_CHILDREN(rack_misc),
1379             OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1380             &rack_prr_addbackmax, 2,
1381             "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1382         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1383             SYSCTL_CHILDREN(rack_misc),
1384             OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1385             &rack_stats_gets_ms_rtt, 1,
1386             "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1387         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1388             SYSCTL_CHILDREN(rack_misc),
1389             OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1390             &rack_client_low_buf, 0,
1391             "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1392         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1393             SYSCTL_CHILDREN(rack_misc),
1394             OID_AUTO, "defprofile", CTLFLAG_RW,
1395             &rack_def_profile, 0,
1396             "Should RACK use a default profile (0=no, num == profile num)?");
1397         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1398             SYSCTL_CHILDREN(rack_misc),
1399             OID_AUTO, "cmpack", CTLFLAG_RW,
1400             &rack_use_cmp_acks, 1,
1401             "Should RACK have LRO send compressed acks");
1402         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1403             SYSCTL_CHILDREN(rack_misc),
1404             OID_AUTO, "fsb", CTLFLAG_RW,
1405             &rack_use_fsb, 1,
1406             "Should RACK use the fast send block?");
1407         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1408             SYSCTL_CHILDREN(rack_misc),
1409             OID_AUTO, "rfo", CTLFLAG_RW,
1410             &rack_use_rfo, 1,
1411             "Should RACK use rack_fast_output()?");
1412         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1413             SYSCTL_CHILDREN(rack_misc),
1414             OID_AUTO, "rsmrfo", CTLFLAG_RW,
1415             &rack_use_rsm_rfo, 1,
1416             "Should RACK use rack_fast_rsm_output()?");
1417         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1418             SYSCTL_CHILDREN(rack_misc),
1419             OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1420             &rack_enable_shared_cwnd, 1,
1421             "Should RACK try to use the shared cwnd on connections where allowed");
1422         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1423             SYSCTL_CHILDREN(rack_misc),
1424             OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1425             &rack_limits_scwnd, 1,
1426             "Should RACK place low end time limits on the shared cwnd feature");
1427         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1428             SYSCTL_CHILDREN(rack_misc),
1429             OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1430             &rack_enable_mqueue_for_nonpaced, 0,
1431             "Should RACK use mbuf queuing for non-paced connections");
1432         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1433             SYSCTL_CHILDREN(rack_misc),
1434             OID_AUTO, "iMac_dack", CTLFLAG_RW,
1435             &rack_use_imac_dack, 0,
1436             "Should RACK try to emulate iMac delayed ack");
1437         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1438             SYSCTL_CHILDREN(rack_misc),
1439             OID_AUTO, "no_prr", CTLFLAG_RW,
1440             &rack_disable_prr, 0,
1441             "Should RACK not use prr and only pace (must have pacing on)");
1442         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1443             SYSCTL_CHILDREN(rack_misc),
1444             OID_AUTO, "bb_verbose", CTLFLAG_RW,
1445             &rack_verbose_logging, 0,
1446             "Should RACK black box logging be verbose");
1447         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1448             SYSCTL_CHILDREN(rack_misc),
1449             OID_AUTO, "data_after_close", CTLFLAG_RW,
1450             &rack_ignore_data_after_close, 1,
1451             "Do we hold off sending a RST until all pending data is ack'd");
1452         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1453             SYSCTL_CHILDREN(rack_misc),
1454             OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1455             &rack_sack_not_required, 1,
1456             "Do we allow rack to run on connections not supporting SACK");
1457         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1458             SYSCTL_CHILDREN(rack_misc),
1459             OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1460             &rack_send_a_lot_in_prr, 1,
1461             "Send a lot in prr");
1462         SYSCTL_ADD_S32(&rack_sysctl_ctx,
1463             SYSCTL_CHILDREN(rack_misc),
1464             OID_AUTO, "autoscale", CTLFLAG_RW,
1465             &rack_autosndbuf_inc, 20,
1466             "What percentage should rack scale up its snd buffer by?");
1467         /* Sack Attacker detection stuff */
1468         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1469             SYSCTL_CHILDREN(rack_attack),
1470             OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1471             &rack_highest_sack_thresh_seen, 0,
1472             "Highest sack to ack ratio seen");
1473         SYSCTL_ADD_U32(&rack_sysctl_ctx,
1474             SYSCTL_CHILDREN(rack_attack),
1475             OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1476             &rack_highest_move_thresh_seen, 0,
1477             "Highest move to non-move ratio seen");
1478         rack_ack_total = counter_u64_alloc(M_WAITOK);
1479         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1480             SYSCTL_CHILDREN(rack_attack),
1481             OID_AUTO, "acktotal", CTLFLAG_RD,
1482             &rack_ack_total,
1483             "Total number of Ack's");
1484         rack_express_sack = counter_u64_alloc(M_WAITOK);
1485         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1486             SYSCTL_CHILDREN(rack_attack),
1487             OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1488             &rack_express_sack,
1489             "Total expresss number of Sack's");
1490         rack_sack_total = counter_u64_alloc(M_WAITOK);
1491         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1492             SYSCTL_CHILDREN(rack_attack),
1493             OID_AUTO, "sacktotal", CTLFLAG_RD,
1494             &rack_sack_total,
1495             "Total number of SACKs");
1496         rack_move_none = counter_u64_alloc(M_WAITOK);
1497         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1498             SYSCTL_CHILDREN(rack_attack),
1499             OID_AUTO, "move_none", CTLFLAG_RD,
1500             &rack_move_none,
1501             "Total number of SACK index reuse of postions under threshold");
1502         rack_move_some = counter_u64_alloc(M_WAITOK);
1503         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1504             SYSCTL_CHILDREN(rack_attack),
1505             OID_AUTO, "move_some", CTLFLAG_RD,
1506             &rack_move_some,
1507             "Total number of SACK index reuse of postions over threshold");
1508         rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1509         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1510             SYSCTL_CHILDREN(rack_attack),
1511             OID_AUTO, "attacks", CTLFLAG_RD,
1512             &rack_sack_attacks_detected,
1513             "Total number of SACK attackers that had sack disabled");
1514         rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1515         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1516             SYSCTL_CHILDREN(rack_attack),
1517             OID_AUTO, "reversed", CTLFLAG_RD,
1518             &rack_sack_attacks_reversed,
1519             "Total number of SACK attackers that were later determined false positive");
1520         rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1521         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1522             SYSCTL_CHILDREN(rack_attack),
1523             OID_AUTO, "nextmerge", CTLFLAG_RD,
1524             &rack_sack_used_next_merge,
1525             "Total number of times we used the next merge");
1526         rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1527         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1528             SYSCTL_CHILDREN(rack_attack),
1529             OID_AUTO, "prevmerge", CTLFLAG_RD,
1530             &rack_sack_used_prev_merge,
1531             "Total number of times we used the prev merge");
1532         /* Counters */
1533         rack_fto_send = counter_u64_alloc(M_WAITOK);
1534         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1535             SYSCTL_CHILDREN(rack_counters),
1536             OID_AUTO, "fto_send", CTLFLAG_RD,
1537             &rack_fto_send, "Total number of rack_fast_output sends");
1538         rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1539         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1540             SYSCTL_CHILDREN(rack_counters),
1541             OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1542             &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1543         rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1544         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1545             SYSCTL_CHILDREN(rack_counters),
1546             OID_AUTO, "nfto_resend", CTLFLAG_RD,
1547             &rack_nfto_resend, "Total number of rack_output retransmissions");
1548         rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1549         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1550             SYSCTL_CHILDREN(rack_counters),
1551             OID_AUTO, "nfto_send", CTLFLAG_RD,
1552             &rack_non_fto_send, "Total number of rack_output first sends");
1553         rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1554         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1555             SYSCTL_CHILDREN(rack_counters),
1556             OID_AUTO, "rfo_extended", CTLFLAG_RD,
1557             &rack_extended_rfo, "Total number of times we extended rfo");
1558
1559         rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1560         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1561             SYSCTL_CHILDREN(rack_counters),
1562             OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1563             &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1564         rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1565
1566         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1567             SYSCTL_CHILDREN(rack_counters),
1568             OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1569             &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1570
1571
1572
1573         rack_badfr = counter_u64_alloc(M_WAITOK);
1574         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1575             SYSCTL_CHILDREN(rack_counters),
1576             OID_AUTO, "badfr", CTLFLAG_RD,
1577             &rack_badfr, "Total number of bad FRs");
1578         rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1579         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1580             SYSCTL_CHILDREN(rack_counters),
1581             OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1582             &rack_badfr_bytes, "Total number of bad FRs");
1583         rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1584         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1585             SYSCTL_CHILDREN(rack_counters),
1586             OID_AUTO, "prrsndret", CTLFLAG_RD,
1587             &rack_rtm_prr_retran,
1588             "Total number of prr based retransmits");
1589         rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1590         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1591             SYSCTL_CHILDREN(rack_counters),
1592             OID_AUTO, "prrsndnew", CTLFLAG_RD,
1593             &rack_rtm_prr_newdata,
1594             "Total number of prr based new transmits");
1595         rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1596         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1597             SYSCTL_CHILDREN(rack_counters),
1598             OID_AUTO, "tsnf", CTLFLAG_RD,
1599             &rack_timestamp_mismatch,
1600             "Total number of timestamps that we could not find the reported ts");
1601         rack_find_high = counter_u64_alloc(M_WAITOK);
1602         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1603             SYSCTL_CHILDREN(rack_counters),
1604             OID_AUTO, "findhigh", CTLFLAG_RD,
1605             &rack_find_high,
1606             "Total number of FIN causing find-high");
1607         rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1608         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1609             SYSCTL_CHILDREN(rack_counters),
1610             OID_AUTO, "reordering", CTLFLAG_RD,
1611             &rack_reorder_seen,
1612             "Total number of times we added delay due to reordering");
1613         rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1614         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1615             SYSCTL_CHILDREN(rack_counters),
1616             OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1617             &rack_tlp_tot,
1618             "Total number of tail loss probe expirations");
1619         rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1620         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1621             SYSCTL_CHILDREN(rack_counters),
1622             OID_AUTO, "tlp_new", CTLFLAG_RD,
1623             &rack_tlp_newdata,
1624             "Total number of tail loss probe sending new data");
1625         rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1626         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1627             SYSCTL_CHILDREN(rack_counters),
1628             OID_AUTO, "tlp_retran", CTLFLAG_RD,
1629             &rack_tlp_retran,
1630             "Total number of tail loss probe sending retransmitted data");
1631         rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1632         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1633             SYSCTL_CHILDREN(rack_counters),
1634             OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1635             &rack_tlp_retran_bytes,
1636             "Total bytes of tail loss probe sending retransmitted data");
1637         rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1638         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1639             SYSCTL_CHILDREN(rack_counters),
1640             OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1641             &rack_tlp_retran_fail,
1642             "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1643         rack_to_tot = counter_u64_alloc(M_WAITOK);
1644         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1645             SYSCTL_CHILDREN(rack_counters),
1646             OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1647             &rack_to_tot,
1648             "Total number of times the rack to expired");
1649         rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1650         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1651             SYSCTL_CHILDREN(rack_counters),
1652             OID_AUTO, "arm_rack", CTLFLAG_RD,
1653             &rack_to_arm_rack,
1654             "Total number of times the rack timer armed");
1655         rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1656         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1657             SYSCTL_CHILDREN(rack_counters),
1658             OID_AUTO, "arm_tlp", CTLFLAG_RD,
1659             &rack_to_arm_tlp,
1660             "Total number of times the tlp timer armed");
1661         rack_calc_zero = counter_u64_alloc(M_WAITOK);
1662         rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1663         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1664             SYSCTL_CHILDREN(rack_counters),
1665             OID_AUTO, "calc_zero", CTLFLAG_RD,
1666             &rack_calc_zero,
1667             "Total number of times pacing time worked out to zero");
1668         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1669             SYSCTL_CHILDREN(rack_counters),
1670             OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1671             &rack_calc_nonzero,
1672             "Total number of times pacing time worked out to non-zero");
1673         rack_paced_segments = counter_u64_alloc(M_WAITOK);
1674         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1675             SYSCTL_CHILDREN(rack_counters),
1676             OID_AUTO, "paced", CTLFLAG_RD,
1677             &rack_paced_segments,
1678             "Total number of times a segment send caused hptsi");
1679         rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1680         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1681             SYSCTL_CHILDREN(rack_counters),
1682             OID_AUTO, "unpaced", CTLFLAG_RD,
1683             &rack_unpaced_segments,
1684             "Total number of times a segment did not cause hptsi");
1685         rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1686         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1687             SYSCTL_CHILDREN(rack_counters),
1688             OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1689             &rack_saw_enobuf,
1690             "Total number of times a sends returned enobuf for non-hdwr paced connections");
1691         rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1692         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1693             SYSCTL_CHILDREN(rack_counters),
1694             OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1695             &rack_saw_enobuf_hw,
1696             "Total number of times a send returned enobuf for hdwr paced connections");
1697         rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1698         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1699             SYSCTL_CHILDREN(rack_counters),
1700             OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1701             &rack_saw_enetunreach,
1702             "Total number of times a send received a enetunreachable");
1703         rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1704         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1705             SYSCTL_CHILDREN(rack_counters),
1706             OID_AUTO, "alloc_hot", CTLFLAG_RD,
1707             &rack_hot_alloc,
1708             "Total allocations from the top of our list");
1709         rack_to_alloc = counter_u64_alloc(M_WAITOK);
1710         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1711             SYSCTL_CHILDREN(rack_counters),
1712             OID_AUTO, "allocs", CTLFLAG_RD,
1713             &rack_to_alloc,
1714             "Total allocations of tracking structures");
1715         rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1716         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1717             SYSCTL_CHILDREN(rack_counters),
1718             OID_AUTO, "allochard", CTLFLAG_RD,
1719             &rack_to_alloc_hard,
1720             "Total allocations done with sleeping the hard way");
1721         rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1722         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1723             SYSCTL_CHILDREN(rack_counters),
1724             OID_AUTO, "allocemerg", CTLFLAG_RD,
1725             &rack_to_alloc_emerg,
1726             "Total allocations done from emergency cache");
1727         rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1728         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1729             SYSCTL_CHILDREN(rack_counters),
1730             OID_AUTO, "alloc_limited", CTLFLAG_RD,
1731             &rack_to_alloc_limited,
1732             "Total allocations dropped due to limit");
1733         rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1734         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1735             SYSCTL_CHILDREN(rack_counters),
1736             OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1737             &rack_alloc_limited_conns,
1738             "Connections with allocations dropped due to limit");
1739         rack_split_limited = counter_u64_alloc(M_WAITOK);
1740         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1741             SYSCTL_CHILDREN(rack_counters),
1742             OID_AUTO, "split_limited", CTLFLAG_RD,
1743             &rack_split_limited,
1744             "Split allocations dropped due to limit");
1745
1746         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
1747                 char name[32];
1748                 sprintf(name, "cmp_ack_cnt_%d", i);
1749                 rack_proc_comp_ack[i] = counter_u64_alloc(M_WAITOK);
1750                 SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1751                                        SYSCTL_CHILDREN(rack_counters),
1752                                        OID_AUTO, name, CTLFLAG_RD,
1753                                        &rack_proc_comp_ack[i],
1754                                        "Number of compressed acks we processed");
1755         }
1756         rack_large_ackcmp = counter_u64_alloc(M_WAITOK);
1757         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1758             SYSCTL_CHILDREN(rack_counters),
1759             OID_AUTO, "cmp_large_mbufs", CTLFLAG_RD,
1760             &rack_large_ackcmp,
1761             "Number of TCP connections with large mbuf's for compressed acks");
1762         rack_small_ackcmp = counter_u64_alloc(M_WAITOK);
1763         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1764             SYSCTL_CHILDREN(rack_counters),
1765             OID_AUTO, "cmp_small_mbufs", CTLFLAG_RD,
1766             &rack_small_ackcmp,
1767             "Number of TCP connections with small mbuf's for compressed acks");
1768 #ifdef INVARIANTS
1769         rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1770         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1771             SYSCTL_CHILDREN(rack_counters),
1772             OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1773             &rack_adjust_map_bw,
1774             "Number of times we hit the case where the sb went up and down on a sendmap entry");
1775 #endif
1776         rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1777         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1778             SYSCTL_CHILDREN(rack_counters),
1779             OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1780             &rack_multi_single_eq,
1781             "Number of compressed acks total represented");
1782         rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1783         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1784             SYSCTL_CHILDREN(rack_counters),
1785             OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1786             &rack_proc_non_comp_ack,
1787             "Number of non compresseds acks that we processed");
1788
1789
1790         rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1791         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1792             SYSCTL_CHILDREN(rack_counters),
1793             OID_AUTO, "sack_long", CTLFLAG_RD,
1794             &rack_sack_proc_all,
1795             "Total times we had to walk whole list for sack processing");
1796         rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1797         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1798             SYSCTL_CHILDREN(rack_counters),
1799             OID_AUTO, "sack_restart", CTLFLAG_RD,
1800             &rack_sack_proc_restart,
1801             "Total times we had to walk whole list due to a restart");
1802         rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1803         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1804             SYSCTL_CHILDREN(rack_counters),
1805             OID_AUTO, "sack_short", CTLFLAG_RD,
1806             &rack_sack_proc_short,
1807             "Total times we took shortcut for sack processing");
1808         rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1809         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1810             SYSCTL_CHILDREN(rack_counters),
1811             OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1812             &rack_enter_tlp_calc,
1813             "Total times we called calc-tlp");
1814         rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1815         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1816             SYSCTL_CHILDREN(rack_counters),
1817             OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1818             &rack_used_tlpmethod,
1819             "Total number of runt sacks");
1820         rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1821         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1822             SYSCTL_CHILDREN(rack_counters),
1823             OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1824             &rack_used_tlpmethod2,
1825             "Total number of times we hit TLP method 2");
1826         rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1827         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1828             SYSCTL_CHILDREN(rack_attack),
1829             OID_AUTO, "skipacked", CTLFLAG_RD,
1830             &rack_sack_skipped_acked,
1831             "Total number of times we skipped previously sacked");
1832         rack_sack_splits = counter_u64_alloc(M_WAITOK);
1833         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1834             SYSCTL_CHILDREN(rack_attack),
1835             OID_AUTO, "ofsplit", CTLFLAG_RD,
1836             &rack_sack_splits,
1837             "Total number of times we did the old fashion tree split");
1838         rack_progress_drops = counter_u64_alloc(M_WAITOK);
1839         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1840             SYSCTL_CHILDREN(rack_counters),
1841             OID_AUTO, "prog_drops", CTLFLAG_RD,
1842             &rack_progress_drops,
1843             "Total number of progress drops");
1844         rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1845         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1846             SYSCTL_CHILDREN(rack_counters),
1847             OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1848             &rack_input_idle_reduces,
1849             "Total number of idle reductions on input");
1850         rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1851         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1852             SYSCTL_CHILDREN(rack_counters),
1853             OID_AUTO, "collapsed_win", CTLFLAG_RD,
1854             &rack_collapsed_win,
1855             "Total number of collapsed windows");
1856         rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1857         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1858             SYSCTL_CHILDREN(rack_counters),
1859             OID_AUTO, "tlp_nada", CTLFLAG_RD,
1860             &rack_tlp_does_nada,
1861             "Total number of nada tlp calls");
1862         rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1863         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1864             SYSCTL_CHILDREN(rack_counters),
1865             OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1866             &rack_try_scwnd,
1867             "Total number of scwnd attempts");
1868
1869         rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1870         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1871             SYSCTL_CHILDREN(rack_counters),
1872             OID_AUTO, "timer_hole", CTLFLAG_RD,
1873             &rack_per_timer_hole,
1874             "Total persists start in timer hole");
1875
1876         rack_sbsndptr_wrong = counter_u64_alloc(M_WAITOK);
1877         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1878             SYSCTL_CHILDREN(rack_counters),
1879             OID_AUTO, "sndptr_wrong", CTLFLAG_RD,
1880             &rack_sbsndptr_wrong, "Total number of times the saved sbsndptr was incorret");
1881         rack_sbsndptr_right = counter_u64_alloc(M_WAITOK);
1882         SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1883             SYSCTL_CHILDREN(rack_counters),
1884             OID_AUTO, "sndptr_right", CTLFLAG_RD,
1885             &rack_sbsndptr_right, "Total number of times the saved sbsndptr was corret");
1886
1887         COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1888         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1889             OID_AUTO, "outsize", CTLFLAG_RD,
1890             rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1891         COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1892         SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1893             OID_AUTO, "opts", CTLFLAG_RD,
1894             rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1895         SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1896             SYSCTL_CHILDREN(rack_sysctl_root),
1897             OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1898             &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1899 }
1900
1901 static __inline int
1902 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1903 {
1904         if (SEQ_GEQ(b->r_start, a->r_start) &&
1905             SEQ_LT(b->r_start, a->r_end)) {
1906                 /*
1907                  * The entry b is within the
1908                  * block a. i.e.:
1909                  * a --   |-------------|
1910                  * b --   |----|
1911                  * <or>
1912                  * b --       |------|
1913                  * <or>
1914                  * b --       |-----------|
1915                  */
1916                 return (0);
1917         } else if (SEQ_GEQ(b->r_start, a->r_end)) {
1918                 /*
1919                  * b falls as either the next
1920                  * sequence block after a so a
1921                  * is said to be smaller than b.
1922                  * i.e:
1923                  * a --   |------|
1924                  * b --          |--------|
1925                  * or
1926                  * b --              |-----|
1927                  */
1928                 return (1);
1929         }
1930         /*
1931          * Whats left is where a is
1932          * larger than b. i.e:
1933          * a --         |-------|
1934          * b --  |---|
1935          * or even possibly
1936          * b --   |--------------|
1937          */
1938         return (-1);
1939 }
1940
1941 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1942 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1943
1944 static uint32_t
1945 rc_init_window(struct tcp_rack *rack)
1946 {
1947         uint32_t win;
1948
1949         if (rack->rc_init_win == 0) {
1950                 /*
1951                  * Nothing set by the user, use the system stack
1952                  * default.
1953                  */
1954                 return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1955         }
1956         win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1957         return (win);
1958 }
1959
1960 static uint64_t
1961 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1962 {
1963         if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1964                 return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1965         else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1966                 return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1967         else
1968                 return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1969 }
1970
1971 static uint64_t
1972 rack_get_bw(struct tcp_rack *rack)
1973 {
1974         if (rack->use_fixed_rate) {
1975                 /* Return the fixed pacing rate */
1976                 return (rack_get_fixed_pacing_bw(rack));
1977         }
1978         if (rack->r_ctl.gp_bw == 0) {
1979                 /*
1980                  * We have yet no b/w measurement,
1981                  * if we have a user set initial bw
1982                  * return it. If we don't have that and
1983                  * we have an srtt, use the tcp IW (10) to
1984                  * calculate a fictional b/w over the SRTT
1985                  * which is more or less a guess. Note
1986                  * we don't use our IW from rack on purpose
1987                  * so if we have like IW=30, we are not
1988                  * calculating a "huge" b/w.
1989                  */
1990                 uint64_t bw, srtt;
1991                 if (rack->r_ctl.init_rate)
1992                         return (rack->r_ctl.init_rate);
1993
1994                 /* Has the user set a max peak rate? */
1995 #ifdef NETFLIX_PEAKRATE
1996                 if (rack->rc_tp->t_maxpeakrate)
1997                         return (rack->rc_tp->t_maxpeakrate);
1998 #endif
1999                 /* Ok lets come up with the IW guess, if we have a srtt */
2000                 if (rack->rc_tp->t_srtt == 0) {
2001                         /*
2002                          * Go with old pacing method
2003                          * i.e. burst mitigation only.
2004                          */
2005                         return (0);
2006                 }
2007                 /* Ok lets get the initial TCP win (not racks) */
2008                 bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2009                 srtt = (uint64_t)rack->rc_tp->t_srtt;
2010                 bw *= (uint64_t)USECS_IN_SECOND;
2011                 bw /= srtt;
2012                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2013                         bw = rack->r_ctl.bw_rate_cap;
2014                 return (bw);
2015         } else {
2016                 uint64_t bw;
2017
2018                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2019                         /* Averaging is done, we can return the value */
2020                         bw = rack->r_ctl.gp_bw;
2021                 } else {
2022                         /* Still doing initial average must calculate */
2023                         bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
2024                 }
2025 #ifdef NETFLIX_PEAKRATE
2026                 if ((rack->rc_tp->t_maxpeakrate) &&
2027                     (bw > rack->rc_tp->t_maxpeakrate)) {
2028                         /* The user has set a peak rate to pace at
2029                          * don't allow us to pace faster than that.
2030                          */
2031                         return (rack->rc_tp->t_maxpeakrate);
2032                 }
2033 #endif
2034                 if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2035                         bw = rack->r_ctl.bw_rate_cap;
2036                 return (bw);
2037         }
2038 }
2039
2040 static uint16_t
2041 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2042 {
2043         if (rack->use_fixed_rate) {
2044                 return (100);
2045         } else if (rack->in_probe_rtt && (rsm == NULL))
2046                 return (rack->r_ctl.rack_per_of_gp_probertt);
2047         else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2048                   rack->r_ctl.rack_per_of_gp_rec)) {
2049                 if (rsm) {
2050                         /* a retransmission always use the recovery rate */
2051                         return (rack->r_ctl.rack_per_of_gp_rec);
2052                 } else if (rack->rack_rec_nonrxt_use_cr) {
2053                         /* Directed to use the configured rate */
2054                         goto configured_rate;
2055                 } else if (rack->rack_no_prr &&
2056                            (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2057                         /* No PRR, lets just use the b/w estimate only */
2058                         return (100);
2059                 } else {
2060                         /*
2061                          * Here we may have a non-retransmit but we
2062                          * have no overrides, so just use the recovery
2063                          * rate (prr is in effect).
2064                          */
2065                         return (rack->r_ctl.rack_per_of_gp_rec);
2066                 }
2067         }
2068 configured_rate:
2069         /* For the configured rate we look at our cwnd vs the ssthresh */
2070         if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2071                 return (rack->r_ctl.rack_per_of_gp_ss);
2072         else
2073                 return (rack->r_ctl.rack_per_of_gp_ca);
2074 }
2075
2076 static void
2077 rack_log_hdwr_pacing(struct tcp_rack *rack,
2078                      uint64_t rate, uint64_t hw_rate, int line,
2079                      int error, uint16_t mod)
2080 {
2081         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2082                 union tcp_log_stackspecific log;
2083                 struct timeval tv;
2084                 const struct ifnet *ifp;
2085
2086                 memset(&log, 0, sizeof(log));
2087                 log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2088                 log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2089                 if (rack->r_ctl.crte) {
2090                         ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2091                 } else if (rack->rc_inp->inp_route.ro_nh &&
2092                            rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2093                         ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2094                 } else
2095                         ifp = NULL;
2096                 if (ifp) {
2097                         log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2098                         log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2099                 }
2100                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2101                 log.u_bbr.bw_inuse = rate;
2102                 log.u_bbr.flex5 = line;
2103                 log.u_bbr.flex6 = error;
2104                 log.u_bbr.flex7 = mod;
2105                 log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2106                 log.u_bbr.flex8 = rack->use_fixed_rate;
2107                 log.u_bbr.flex8 <<= 1;
2108                 log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2109                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2110                 log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2111                 if (rack->r_ctl.crte)
2112                         log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2113                 else
2114                         log.u_bbr.cur_del_rate = 0;
2115                 log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2116                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2117                     &rack->rc_inp->inp_socket->so_rcv,
2118                     &rack->rc_inp->inp_socket->so_snd,
2119                     BBR_LOG_HDWR_PACE, 0,
2120                     0, &log, false, &tv);
2121         }
2122 }
2123
2124 static uint64_t
2125 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2126 {
2127         /*
2128          * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2129          */
2130         uint64_t bw_est, high_rate;
2131         uint64_t gain;
2132
2133         gain = (uint64_t)rack_get_output_gain(rack, rsm);
2134         bw_est = bw * gain;
2135         bw_est /= (uint64_t)100;
2136         /* Never fall below the minimum (def 64kbps) */
2137         if (bw_est < RACK_MIN_BW)
2138                 bw_est = RACK_MIN_BW;
2139         if (rack->r_rack_hw_rate_caps) {
2140                 /* Rate caps are in place */
2141                 if (rack->r_ctl.crte != NULL) {
2142                         /* We have a hdwr rate already */
2143                         high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2144                         if (bw_est >= high_rate) {
2145                                 /* We are capping bw at the highest rate table entry */
2146                                 rack_log_hdwr_pacing(rack,
2147                                                      bw_est, high_rate, __LINE__,
2148                                                      0, 3);
2149                                 bw_est = high_rate;
2150                                 if (capped)
2151                                         *capped = 1;
2152                         }
2153                 } else if ((rack->rack_hdrw_pacing == 0) &&
2154                            (rack->rack_hdw_pace_ena) &&
2155                            (rack->rack_attempt_hdwr_pace == 0) &&
2156                            (rack->rc_inp->inp_route.ro_nh != NULL) &&
2157                            (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2158                         /*
2159                          * Special case, we have not yet attempted hardware
2160                          * pacing, and yet we may, when we do, find out if we are
2161                          * above the highest rate. We need to know the maxbw for the interface
2162                          * in question (if it supports ratelimiting). We get back
2163                          * a 0, if the interface is not found in the RL lists.
2164                          */
2165                         high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2166                         if (high_rate) {
2167                                 /* Yep, we have a rate is it above this rate? */
2168                                 if (bw_est > high_rate) {
2169                                         bw_est = high_rate;
2170                                         if (capped)
2171                                                 *capped = 1;
2172                                 }
2173                         }
2174                 }
2175         }
2176         return (bw_est);
2177 }
2178
2179 static void
2180 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2181 {
2182         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2183                 union tcp_log_stackspecific log;
2184                 struct timeval tv;
2185
2186                 if ((mod != 1) && (rack_verbose_logging == 0)) {
2187                         /*
2188                          * We get 3 values currently for mod
2189                          * 1 - We are retransmitting and this tells the reason.
2190                          * 2 - We are clearing a dup-ack count.
2191                          * 3 - We are incrementing a dup-ack count.
2192                          *
2193                          * The clear/increment are only logged
2194                          * if you have BBverbose on.
2195                          */
2196                         return;
2197                 }
2198                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2199                 log.u_bbr.flex1 = tsused;
2200                 log.u_bbr.flex2 = thresh;
2201                 log.u_bbr.flex3 = rsm->r_flags;
2202                 log.u_bbr.flex4 = rsm->r_dupack;
2203                 log.u_bbr.flex5 = rsm->r_start;
2204                 log.u_bbr.flex6 = rsm->r_end;
2205                 log.u_bbr.flex8 = mod;
2206                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2207                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2208                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2209                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2210                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2211                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2212                 log.u_bbr.pacing_gain = rack->r_must_retran;
2213                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2214                     &rack->rc_inp->inp_socket->so_rcv,
2215                     &rack->rc_inp->inp_socket->so_snd,
2216                     BBR_LOG_SETTINGS_CHG, 0,
2217                     0, &log, false, &tv);
2218         }
2219 }
2220
2221 static void
2222 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2223 {
2224         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2225                 union tcp_log_stackspecific log;
2226                 struct timeval tv;
2227
2228                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2229                 log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2230                 log.u_bbr.flex2 = to;
2231                 log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2232                 log.u_bbr.flex4 = slot;
2233                 log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2234                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2235                 log.u_bbr.flex7 = rack->rc_in_persist;
2236                 log.u_bbr.flex8 = which;
2237                 if (rack->rack_no_prr)
2238                         log.u_bbr.pkts_out = 0;
2239                 else
2240                         log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2241                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2242                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2243                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2244                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2245                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2246                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2247                 log.u_bbr.pacing_gain = rack->r_must_retran;
2248                 log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2249                 log.u_bbr.lost = rack_rto_min;
2250                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2251                     &rack->rc_inp->inp_socket->so_rcv,
2252                     &rack->rc_inp->inp_socket->so_snd,
2253                     BBR_LOG_TIMERSTAR, 0,
2254                     0, &log, false, &tv);
2255         }
2256 }
2257
2258 static void
2259 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2260 {
2261         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2262                 union tcp_log_stackspecific log;
2263                 struct timeval tv;
2264
2265                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2266                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2267                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2268                 log.u_bbr.flex8 = to_num;
2269                 log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2270                 log.u_bbr.flex2 = rack->rc_rack_rtt;
2271                 if (rsm == NULL)
2272                         log.u_bbr.flex3 = 0;
2273                 else
2274                         log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2275                 if (rack->rack_no_prr)
2276                         log.u_bbr.flex5 = 0;
2277                 else
2278                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2279                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2280                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2281                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2282                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2283                 log.u_bbr.pacing_gain = rack->r_must_retran;
2284                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2285                     &rack->rc_inp->inp_socket->so_rcv,
2286                     &rack->rc_inp->inp_socket->so_snd,
2287                     BBR_LOG_RTO, 0,
2288                     0, &log, false, &tv);
2289         }
2290 }
2291
2292 static void
2293 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2294                  struct rack_sendmap *prev,
2295                  struct rack_sendmap *rsm,
2296                  struct rack_sendmap *next,
2297                  int flag, uint32_t th_ack, int line)
2298 {
2299         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2300                 union tcp_log_stackspecific log;
2301                 struct timeval tv;
2302
2303                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2304                 log.u_bbr.flex8 = flag;
2305                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2306                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2307                 log.u_bbr.cur_del_rate = (uint64_t)prev;
2308                 log.u_bbr.delRate = (uint64_t)rsm;
2309                 log.u_bbr.rttProp = (uint64_t)next;
2310                 log.u_bbr.flex7 = 0;
2311                 if (prev) {
2312                         log.u_bbr.flex1 = prev->r_start;
2313                         log.u_bbr.flex2 = prev->r_end;
2314                         log.u_bbr.flex7 |= 0x4;
2315                 }
2316                 if (rsm) {
2317                         log.u_bbr.flex3 = rsm->r_start;
2318                         log.u_bbr.flex4 = rsm->r_end;
2319                         log.u_bbr.flex7 |= 0x2;
2320                 }
2321                 if (next) {
2322                         log.u_bbr.flex5 = next->r_start;
2323                         log.u_bbr.flex6 = next->r_end;
2324                         log.u_bbr.flex7 |= 0x1;
2325                 }
2326                 log.u_bbr.applimited = line;
2327                 log.u_bbr.pkts_out = th_ack;
2328                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2329                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2330                 if (rack->rack_no_prr)
2331                         log.u_bbr.lost = 0;
2332                 else
2333                         log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2334                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2335                     &rack->rc_inp->inp_socket->so_rcv,
2336                     &rack->rc_inp->inp_socket->so_snd,
2337                     TCP_LOG_MAPCHG, 0,
2338                     0, &log, false, &tv);
2339         }
2340 }
2341
2342 static void
2343 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2344                  struct rack_sendmap *rsm, int conf)
2345 {
2346         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2347                 union tcp_log_stackspecific log;
2348                 struct timeval tv;
2349                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2350                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2351                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2352                 log.u_bbr.flex1 = t;
2353                 log.u_bbr.flex2 = len;
2354                 log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2355                 log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2356                 log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2357                 log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2358                 log.u_bbr.flex7 = conf;
2359                 log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2360                 log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2361                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2362                 log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2363                 log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2364                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2365                 if (rsm) {
2366                         log.u_bbr.pkt_epoch = rsm->r_start;
2367                         log.u_bbr.lost = rsm->r_end;
2368                         log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2369                         log.u_bbr.pacing_gain = rsm->r_flags;
2370                 } else {
2371                         /* Its a SYN */
2372                         log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2373                         log.u_bbr.lost = 0;
2374                         log.u_bbr.cwnd_gain = 0;
2375                         log.u_bbr.pacing_gain = 0;
2376                 }
2377                 /* Write out general bits of interest rrs here */
2378                 log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2379                 log.u_bbr.use_lt_bw <<= 1;
2380                 log.u_bbr.use_lt_bw |= rack->forced_ack;
2381                 log.u_bbr.use_lt_bw <<= 1;
2382                 log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2383                 log.u_bbr.use_lt_bw <<= 1;
2384                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2385                 log.u_bbr.use_lt_bw <<= 1;
2386                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2387                 log.u_bbr.use_lt_bw <<= 1;
2388                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2389                 log.u_bbr.use_lt_bw <<= 1;
2390                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2391                 log.u_bbr.use_lt_bw <<= 1;
2392                 log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2393                 log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2394                 log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2395                 log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2396                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2397                 log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2398                 log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2399                 log.u_bbr.bw_inuse <<= 32;
2400                 if (rsm)
2401                         log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2402                 TCP_LOG_EVENTP(tp, NULL,
2403                     &rack->rc_inp->inp_socket->so_rcv,
2404                     &rack->rc_inp->inp_socket->so_snd,
2405                     BBR_LOG_BBRRTT, 0,
2406                     0, &log, false, &tv);
2407
2408
2409         }
2410 }
2411
2412 static void
2413 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2414 {
2415         /*
2416          * Log the rtt sample we are
2417          * applying to the srtt algorithm in
2418          * useconds.
2419          */
2420         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2421                 union tcp_log_stackspecific log;
2422                 struct timeval tv;
2423
2424                 /* Convert our ms to a microsecond */
2425                 memset(&log, 0, sizeof(log));
2426                 log.u_bbr.flex1 = rtt;
2427                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2428                 log.u_bbr.flex3 = rack->r_ctl.sack_count;
2429                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2430                 log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2431                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2432                 log.u_bbr.flex7 = 1;
2433                 log.u_bbr.flex8 = rack->sack_attack_disable;
2434                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2435                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2436                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2437                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2438                 log.u_bbr.pacing_gain = rack->r_must_retran;
2439                 /*
2440                  * We capture in delRate the upper 32 bits as
2441                  * the confidence level we had declared, and the
2442                  * lower 32 bits as the actual RTT using the arrival
2443                  * timestamp.
2444                  */
2445                 log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2446                 log.u_bbr.delRate <<= 32;
2447                 log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2448                 /* Lets capture all the things that make up t_rtxcur */
2449                 log.u_bbr.applimited = rack_rto_min;
2450                 log.u_bbr.epoch = rack_rto_max;
2451                 log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2452                 log.u_bbr.lost = rack_rto_min;
2453                 log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2454                 log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2455                 log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2456                 log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2457                 log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2458                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2459                     &rack->rc_inp->inp_socket->so_rcv,
2460                     &rack->rc_inp->inp_socket->so_snd,
2461                     TCP_LOG_RTT, 0,
2462                     0, &log, false, &tv);
2463         }
2464 }
2465
2466 static void
2467 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2468 {
2469         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2470                 union tcp_log_stackspecific log;
2471                 struct timeval tv;
2472
2473                 /* Convert our ms to a microsecond */
2474                 memset(&log, 0, sizeof(log));
2475                 log.u_bbr.flex1 = rtt;
2476                 log.u_bbr.flex2 = send_time;
2477                 log.u_bbr.flex3 = ack_time;
2478                 log.u_bbr.flex4 = where;
2479                 log.u_bbr.flex7 = 2;
2480                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2481                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2482                     &rack->rc_inp->inp_socket->so_rcv,
2483                     &rack->rc_inp->inp_socket->so_snd,
2484                     TCP_LOG_RTT, 0,
2485                     0, &log, false, &tv);
2486         }
2487 }
2488
2489
2490
2491 static inline void
2492 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2493 {
2494         if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2495                 union tcp_log_stackspecific log;
2496                 struct timeval tv;
2497
2498                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2499                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2500                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2501                 log.u_bbr.flex1 = line;
2502                 log.u_bbr.flex2 = tick;
2503                 log.u_bbr.flex3 = tp->t_maxunacktime;
2504                 log.u_bbr.flex4 = tp->t_acktime;
2505                 log.u_bbr.flex8 = event;
2506                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2507                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2508                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2509                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2510                 log.u_bbr.pacing_gain = rack->r_must_retran;
2511                 TCP_LOG_EVENTP(tp, NULL,
2512                     &rack->rc_inp->inp_socket->so_rcv,
2513                     &rack->rc_inp->inp_socket->so_snd,
2514                     BBR_LOG_PROGRESS, 0,
2515                     0, &log, false, &tv);
2516         }
2517 }
2518
2519 static void
2520 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2521 {
2522         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2523                 union tcp_log_stackspecific log;
2524
2525                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2526                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2527                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2528                 log.u_bbr.flex1 = slot;
2529                 if (rack->rack_no_prr)
2530                         log.u_bbr.flex2 = 0;
2531                 else
2532                         log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2533                 log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2534                 log.u_bbr.flex8 = rack->rc_in_persist;
2535                 log.u_bbr.timeStamp = cts;
2536                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2537                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2538                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2539                 log.u_bbr.pacing_gain = rack->r_must_retran;
2540                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2541                     &rack->rc_inp->inp_socket->so_rcv,
2542                     &rack->rc_inp->inp_socket->so_snd,
2543                     BBR_LOG_BBRSND, 0,
2544                     0, &log, false, tv);
2545         }
2546 }
2547
2548 static void
2549 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2550 {
2551         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2552                 union tcp_log_stackspecific log;
2553                 struct timeval tv;
2554
2555                 memset(&log, 0, sizeof(log));
2556                 log.u_bbr.flex1 = did_out;
2557                 log.u_bbr.flex2 = nxt_pkt;
2558                 log.u_bbr.flex3 = way_out;
2559                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2560                 if (rack->rack_no_prr)
2561                         log.u_bbr.flex5 = 0;
2562                 else
2563                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2564                 log.u_bbr.flex6 = nsegs;
2565                 log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2566                 log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;        /* Do we have ack-can-send set */
2567                 log.u_bbr.flex7 <<= 1;
2568                 log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */
2569                 log.u_bbr.flex7 <<= 1;
2570                 log.u_bbr.flex7 |= rack->r_wanted_output;       /* Do we want output */
2571                 log.u_bbr.flex8 = rack->rc_in_persist;
2572                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2573                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2574                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2575                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2576                 log.u_bbr.use_lt_bw <<= 1;
2577                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2578                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2579                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2580                 log.u_bbr.pacing_gain = rack->r_must_retran;
2581                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2582                     &rack->rc_inp->inp_socket->so_rcv,
2583                     &rack->rc_inp->inp_socket->so_snd,
2584                     BBR_LOG_DOSEG_DONE, 0,
2585                     0, &log, false, &tv);
2586         }
2587 }
2588
2589 static void
2590 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2591 {
2592         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2593                 union tcp_log_stackspecific log;
2594                 struct timeval tv;
2595                 uint32_t cts;
2596
2597                 memset(&log, 0, sizeof(log));
2598                 cts = tcp_get_usecs(&tv);
2599                 log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2600                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2601                 log.u_bbr.flex4 = arg1;
2602                 log.u_bbr.flex5 = arg2;
2603                 log.u_bbr.flex6 = arg3;
2604                 log.u_bbr.flex8 = frm;
2605                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2606                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2607                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2608                 log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2609                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2610                 log.u_bbr.pacing_gain = rack->r_must_retran;
2611                 TCP_LOG_EVENTP(tp, NULL,
2612                     &tp->t_inpcb->inp_socket->so_rcv,
2613                     &tp->t_inpcb->inp_socket->so_snd,
2614                     TCP_HDWR_PACE_SIZE, 0,
2615                     0, &log, false, &tv);
2616         }
2617 }
2618
2619 static void
2620 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2621                           uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2622 {
2623         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2624                 union tcp_log_stackspecific log;
2625                 struct timeval tv;
2626
2627                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2628                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2629                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2630                 log.u_bbr.flex1 = slot;
2631                 log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2632                 log.u_bbr.flex4 = reason;
2633                 if (rack->rack_no_prr)
2634                         log.u_bbr.flex5 = 0;
2635                 else
2636                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2637                 log.u_bbr.flex7 = hpts_calling;
2638                 log.u_bbr.flex8 = rack->rc_in_persist;
2639                 log.u_bbr.lt_epoch = cwnd_to_use;
2640                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2641                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2642                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2643                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2644                 log.u_bbr.pacing_gain = rack->r_must_retran;
2645                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2646                     &rack->rc_inp->inp_socket->so_rcv,
2647                     &rack->rc_inp->inp_socket->so_snd,
2648                     BBR_LOG_JUSTRET, 0,
2649                     tlen, &log, false, &tv);
2650         }
2651 }
2652
2653 static void
2654 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2655                    struct timeval *tv, uint32_t flags_on_entry)
2656 {
2657         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2658                 union tcp_log_stackspecific log;
2659
2660                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2661                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2662                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2663                 log.u_bbr.flex1 = line;
2664                 log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2665                 log.u_bbr.flex3 = flags_on_entry;
2666                 log.u_bbr.flex4 = us_cts;
2667                 if (rack->rack_no_prr)
2668                         log.u_bbr.flex5 = 0;
2669                 else
2670                         log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2671                 log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2672                 log.u_bbr.flex7 = hpts_removed;
2673                 log.u_bbr.flex8 = 1;
2674                 log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2675                 log.u_bbr.timeStamp = us_cts;
2676                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2677                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2678                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2679                 log.u_bbr.pacing_gain = rack->r_must_retran;
2680                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2681                     &rack->rc_inp->inp_socket->so_rcv,
2682                     &rack->rc_inp->inp_socket->so_snd,
2683                     BBR_LOG_TIMERCANC, 0,
2684                     0, &log, false, tv);
2685         }
2686 }
2687
2688 static void
2689 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2690                           uint32_t flex1, uint32_t flex2,
2691                           uint32_t flex3, uint32_t flex4,
2692                           uint32_t flex5, uint32_t flex6,
2693                           uint16_t flex7, uint8_t mod)
2694 {
2695         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2696                 union tcp_log_stackspecific log;
2697                 struct timeval tv;
2698
2699                 if (mod == 1) {
2700                         /* No you can't use 1, its for the real to cancel */
2701                         return;
2702                 }
2703                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2704                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2705                 log.u_bbr.flex1 = flex1;
2706                 log.u_bbr.flex2 = flex2;
2707                 log.u_bbr.flex3 = flex3;
2708                 log.u_bbr.flex4 = flex4;
2709                 log.u_bbr.flex5 = flex5;
2710                 log.u_bbr.flex6 = flex6;
2711                 log.u_bbr.flex7 = flex7;
2712                 log.u_bbr.flex8 = mod;
2713                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2714                     &rack->rc_inp->inp_socket->so_rcv,
2715                     &rack->rc_inp->inp_socket->so_snd,
2716                     BBR_LOG_TIMERCANC, 0,
2717                     0, &log, false, &tv);
2718         }
2719 }
2720
2721 static void
2722 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2723 {
2724         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2725                 union tcp_log_stackspecific log;
2726                 struct timeval tv;
2727
2728                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2729                 log.u_bbr.flex1 = timers;
2730                 log.u_bbr.flex2 = ret;
2731                 log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2732                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2733                 log.u_bbr.flex5 = cts;
2734                 if (rack->rack_no_prr)
2735                         log.u_bbr.flex6 = 0;
2736                 else
2737                         log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2738                 log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2739                 log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2740                 log.u_bbr.pacing_gain = rack->r_must_retran;
2741                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2742                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2743                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2744                     &rack->rc_inp->inp_socket->so_rcv,
2745                     &rack->rc_inp->inp_socket->so_snd,
2746                     BBR_LOG_TO_PROCESS, 0,
2747                     0, &log, false, &tv);
2748         }
2749 }
2750
2751 static void
2752 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2753 {
2754         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2755                 union tcp_log_stackspecific log;
2756                 struct timeval tv;
2757
2758                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2759                 log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2760                 log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2761                 if (rack->rack_no_prr)
2762                         log.u_bbr.flex3 = 0;
2763                 else
2764                         log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2765                 log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2766                 log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2767                 log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2768                 log.u_bbr.flex8 = frm;
2769                 log.u_bbr.pkts_out = orig_cwnd;
2770                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2771                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2772                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2773                 log.u_bbr.use_lt_bw <<= 1;
2774                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
2775                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2776                     &rack->rc_inp->inp_socket->so_rcv,
2777                     &rack->rc_inp->inp_socket->so_snd,
2778                     BBR_LOG_BBRUPD, 0,
2779                     0, &log, false, &tv);
2780         }
2781 }
2782
2783 #ifdef NETFLIX_EXP_DETECTION
2784 static void
2785 rack_log_sad(struct tcp_rack *rack, int event)
2786 {
2787         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2788                 union tcp_log_stackspecific log;
2789                 struct timeval tv;
2790
2791                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2792                 log.u_bbr.flex1 = rack->r_ctl.sack_count;
2793                 log.u_bbr.flex2 = rack->r_ctl.ack_count;
2794                 log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2795                 log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2796                 log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2797                 log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2798                 log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2799                 log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2800                 log.u_bbr.lt_epoch |= rack->do_detection;
2801                 log.u_bbr.applimited = tcp_map_minimum;
2802                 log.u_bbr.flex7 = rack->sack_attack_disable;
2803                 log.u_bbr.flex8 = event;
2804                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2805                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2806                 log.u_bbr.delivered = tcp_sad_decay_val;
2807                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
2808                     &rack->rc_inp->inp_socket->so_rcv,
2809                     &rack->rc_inp->inp_socket->so_snd,
2810                     TCP_SAD_DETECTION, 0,
2811                     0, &log, false, &tv);
2812         }
2813 }
2814 #endif
2815
2816 static void
2817 rack_counter_destroy(void)
2818 {
2819         int i;
2820
2821         counter_u64_free(rack_fto_send);
2822         counter_u64_free(rack_fto_rsm_send);
2823         counter_u64_free(rack_nfto_resend);
2824         counter_u64_free(rack_hw_pace_init_fail);
2825         counter_u64_free(rack_hw_pace_lost);
2826         counter_u64_free(rack_non_fto_send);
2827         counter_u64_free(rack_extended_rfo);
2828         counter_u64_free(rack_ack_total);
2829         counter_u64_free(rack_express_sack);
2830         counter_u64_free(rack_sack_total);
2831         counter_u64_free(rack_move_none);
2832         counter_u64_free(rack_move_some);
2833         counter_u64_free(rack_sack_attacks_detected);
2834         counter_u64_free(rack_sack_attacks_reversed);
2835         counter_u64_free(rack_sack_used_next_merge);
2836         counter_u64_free(rack_sack_used_prev_merge);
2837         counter_u64_free(rack_badfr);
2838         counter_u64_free(rack_badfr_bytes);
2839         counter_u64_free(rack_rtm_prr_retran);
2840         counter_u64_free(rack_rtm_prr_newdata);
2841         counter_u64_free(rack_timestamp_mismatch);
2842         counter_u64_free(rack_find_high);
2843         counter_u64_free(rack_reorder_seen);
2844         counter_u64_free(rack_tlp_tot);
2845         counter_u64_free(rack_tlp_newdata);
2846         counter_u64_free(rack_tlp_retran);
2847         counter_u64_free(rack_tlp_retran_bytes);
2848         counter_u64_free(rack_tlp_retran_fail);
2849         counter_u64_free(rack_to_tot);
2850         counter_u64_free(rack_to_arm_rack);
2851         counter_u64_free(rack_to_arm_tlp);
2852         counter_u64_free(rack_calc_zero);
2853         counter_u64_free(rack_calc_nonzero);
2854         counter_u64_free(rack_paced_segments);
2855         counter_u64_free(rack_unpaced_segments);
2856         counter_u64_free(rack_saw_enobuf);
2857         counter_u64_free(rack_saw_enobuf_hw);
2858         counter_u64_free(rack_saw_enetunreach);
2859         counter_u64_free(rack_hot_alloc);
2860         counter_u64_free(rack_to_alloc);
2861         counter_u64_free(rack_to_alloc_hard);
2862         counter_u64_free(rack_to_alloc_emerg);
2863         counter_u64_free(rack_to_alloc_limited);
2864         counter_u64_free(rack_alloc_limited_conns);
2865         counter_u64_free(rack_split_limited);
2866         for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
2867                 counter_u64_free(rack_proc_comp_ack[i]);
2868         }
2869         counter_u64_free(rack_multi_single_eq);
2870         counter_u64_free(rack_proc_non_comp_ack);
2871         counter_u64_free(rack_sack_proc_all);
2872         counter_u64_free(rack_sack_proc_restart);
2873         counter_u64_free(rack_sack_proc_short);
2874         counter_u64_free(rack_enter_tlp_calc);
2875         counter_u64_free(rack_used_tlpmethod);
2876         counter_u64_free(rack_used_tlpmethod2);
2877         counter_u64_free(rack_sack_skipped_acked);
2878         counter_u64_free(rack_sack_splits);
2879         counter_u64_free(rack_progress_drops);
2880         counter_u64_free(rack_input_idle_reduces);
2881         counter_u64_free(rack_collapsed_win);
2882         counter_u64_free(rack_tlp_does_nada);
2883         counter_u64_free(rack_try_scwnd);
2884         counter_u64_free(rack_per_timer_hole);
2885         counter_u64_free(rack_large_ackcmp);
2886         counter_u64_free(rack_small_ackcmp);
2887 #ifdef INVARIANTS
2888         counter_u64_free(rack_adjust_map_bw);
2889 #endif
2890         COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2891         COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2892 }
2893
2894 static struct rack_sendmap *
2895 rack_alloc(struct tcp_rack *rack)
2896 {
2897         struct rack_sendmap *rsm;
2898
2899         /*
2900          * First get the top of the list it in
2901          * theory is the "hottest" rsm we have,
2902          * possibly just freed by ack processing.
2903          */
2904         if (rack->rc_free_cnt > rack_free_cache) {
2905                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2906                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2907                 counter_u64_add(rack_hot_alloc, 1);
2908                 rack->rc_free_cnt--;
2909                 return (rsm);
2910         }
2911         /*
2912          * Once we get under our free cache we probably
2913          * no longer have a "hot" one available. Lets
2914          * get one from UMA.
2915          */
2916         rsm = uma_zalloc(rack_zone, M_NOWAIT);
2917         if (rsm) {
2918                 rack->r_ctl.rc_num_maps_alloced++;
2919                 counter_u64_add(rack_to_alloc, 1);
2920                 return (rsm);
2921         }
2922         /*
2923          * Dig in to our aux rsm's (the last two) since
2924          * UMA failed to get us one.
2925          */
2926         if (rack->rc_free_cnt) {
2927                 counter_u64_add(rack_to_alloc_emerg, 1);
2928                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2929                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2930                 rack->rc_free_cnt--;
2931                 return (rsm);
2932         }
2933         return (NULL);
2934 }
2935
2936 static struct rack_sendmap *
2937 rack_alloc_full_limit(struct tcp_rack *rack)
2938 {
2939         if ((V_tcp_map_entries_limit > 0) &&
2940             (rack->do_detection == 0) &&
2941             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2942                 counter_u64_add(rack_to_alloc_limited, 1);
2943                 if (!rack->alloc_limit_reported) {
2944                         rack->alloc_limit_reported = 1;
2945                         counter_u64_add(rack_alloc_limited_conns, 1);
2946                 }
2947                 return (NULL);
2948         }
2949         return (rack_alloc(rack));
2950 }
2951
2952 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2953 static struct rack_sendmap *
2954 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2955 {
2956         struct rack_sendmap *rsm;
2957
2958         if (limit_type) {
2959                 /* currently there is only one limit type */
2960                 if (V_tcp_map_split_limit > 0 &&
2961                     (rack->do_detection == 0) &&
2962                     rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2963                         counter_u64_add(rack_split_limited, 1);
2964                         if (!rack->alloc_limit_reported) {
2965                                 rack->alloc_limit_reported = 1;
2966                                 counter_u64_add(rack_alloc_limited_conns, 1);
2967                         }
2968                         return (NULL);
2969                 }
2970         }
2971
2972         /* allocate and mark in the limit type, if set */
2973         rsm = rack_alloc(rack);
2974         if (rsm != NULL && limit_type) {
2975                 rsm->r_limit_type = limit_type;
2976                 rack->r_ctl.rc_num_split_allocs++;
2977         }
2978         return (rsm);
2979 }
2980
2981 static void
2982 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2983 {
2984         if (rsm->r_flags & RACK_APP_LIMITED) {
2985                 if (rack->r_ctl.rc_app_limited_cnt > 0) {
2986                         rack->r_ctl.rc_app_limited_cnt--;
2987                 }
2988         }
2989         if (rsm->r_limit_type) {
2990                 /* currently there is only one limit type */
2991                 rack->r_ctl.rc_num_split_allocs--;
2992         }
2993         if (rsm == rack->r_ctl.rc_first_appl) {
2994                 if (rack->r_ctl.rc_app_limited_cnt == 0)
2995                         rack->r_ctl.rc_first_appl = NULL;
2996                 else {
2997                         /* Follow the next one out */
2998                         struct rack_sendmap fe;
2999
3000                         fe.r_start = rsm->r_nseq_appl;
3001                         rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3002                 }
3003         }
3004         if (rsm == rack->r_ctl.rc_resend)
3005                 rack->r_ctl.rc_resend = NULL;
3006         if (rsm == rack->r_ctl.rc_rsm_at_retran)
3007                 rack->r_ctl.rc_rsm_at_retran = NULL;
3008         if (rsm == rack->r_ctl.rc_end_appl)
3009                 rack->r_ctl.rc_end_appl = NULL;
3010         if (rack->r_ctl.rc_tlpsend == rsm)
3011                 rack->r_ctl.rc_tlpsend = NULL;
3012         if (rack->r_ctl.rc_sacklast == rsm)
3013                 rack->r_ctl.rc_sacklast = NULL;
3014         memset(rsm, 0, sizeof(struct rack_sendmap));
3015         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3016         rack->rc_free_cnt++;
3017 }
3018
3019 static void
3020 rack_free_trim(struct tcp_rack *rack)
3021 {
3022         struct rack_sendmap *rsm;
3023
3024         /*
3025          * Free up all the tail entries until
3026          * we get our list down to the limit.
3027          */
3028         while (rack->rc_free_cnt > rack_free_cache) {
3029                 rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3030                 TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3031                 rack->rc_free_cnt--;
3032                 uma_zfree(rack_zone, rsm);
3033         }
3034 }
3035
3036
3037 static uint32_t
3038 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3039 {
3040         uint64_t srtt, bw, len, tim;
3041         uint32_t segsiz, def_len, minl;
3042
3043         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3044         def_len = rack_def_data_window * segsiz;
3045         if (rack->rc_gp_filled == 0) {
3046                 /*
3047                  * We have no measurement (IW is in flight?) so
3048                  * we can only guess using our data_window sysctl
3049                  * value (usually 100MSS).
3050                  */
3051                 return (def_len);
3052         }
3053         /*
3054          * Now we have a number of factors to consider.
3055          *
3056          * 1) We have a desired BDP which is usually
3057          *    at least 2.
3058          * 2) We have a minimum number of rtt's usually 1 SRTT
3059          *    but we allow it too to be more.
3060          * 3) We want to make sure a measurement last N useconds (if
3061          *    we have set rack_min_measure_usec.
3062          *
3063          * We handle the first concern here by trying to create a data
3064          * window of max(rack_def_data_window, DesiredBDP). The
3065          * second concern we handle in not letting the measurement
3066          * window end normally until at least the required SRTT's
3067          * have gone by which is done further below in
3068          * rack_enough_for_measurement(). Finally the third concern
3069          * we also handle here by calculating how long that time
3070          * would take at the current BW and then return the
3071          * max of our first calculation and that length. Note
3072          * that if rack_min_measure_usec is 0, we don't deal
3073          * with concern 3. Also for both Concern 1 and 3 an
3074          * application limited period could end the measurement
3075          * earlier.
3076          *
3077          * So lets calculate the BDP with the "known" b/w using
3078          * the SRTT has our rtt and then multiply it by the
3079          * goal.
3080          */
3081         bw = rack_get_bw(rack);
3082         srtt = (uint64_t)tp->t_srtt;
3083         len = bw * srtt;
3084         len /= (uint64_t)HPTS_USEC_IN_SEC;
3085         len *= max(1, rack_goal_bdp);
3086         /* Now we need to round up to the nearest MSS */
3087         len = roundup(len, segsiz);
3088         if (rack_min_measure_usec) {
3089                 /* Now calculate our min length for this b/w */
3090                 tim = rack_min_measure_usec;
3091                 minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3092                 if (minl == 0)
3093                         minl = 1;
3094                 minl = roundup(minl, segsiz);
3095                 if (len < minl)
3096                         len = minl;
3097         }
3098         /*
3099          * Now if we have a very small window we want
3100          * to attempt to get the window that is
3101          * as small as possible. This happens on
3102          * low b/w connections and we don't want to
3103          * span huge numbers of rtt's between measurements.
3104          *
3105          * We basically include 2 over our "MIN window" so
3106          * that the measurement can be shortened (possibly) by
3107          * an ack'ed packet.
3108          */
3109         if (len < def_len)
3110                 return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3111         else
3112                 return (max((uint32_t)len, def_len));
3113
3114 }
3115
3116 static int
3117 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack)
3118 {
3119         uint32_t tim, srtts, segsiz;
3120
3121         /*
3122          * Has enough time passed for the GP measurement to be valid?
3123          */
3124         if ((tp->snd_max == tp->snd_una) ||
3125             (th_ack == tp->snd_max)){
3126                 /* All is acked */
3127                 return (1);
3128         }
3129         if (SEQ_LT(th_ack, tp->gput_seq)) {
3130                 /* Not enough bytes yet */
3131                 return (0);
3132         }
3133         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3134         if (SEQ_LT(th_ack, tp->gput_ack) &&
3135             ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3136                 /* Not enough bytes yet */
3137                 return (0);
3138         }
3139         if (rack->r_ctl.rc_first_appl &&
3140             (rack->r_ctl.rc_first_appl->r_start == th_ack)) {
3141                 /*
3142                  * We are up to the app limited point
3143                  * we have to measure irrespective of the time..
3144                  */
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                 return (1);
3152         }
3153         /* Nope not even a full SRTT has passed */
3154         return (0);
3155 }
3156
3157 static void
3158 rack_log_timely(struct tcp_rack *rack,
3159                 uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3160                 uint64_t up_bnd, int line, uint8_t method)
3161 {
3162         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3163                 union tcp_log_stackspecific log;
3164                 struct timeval tv;
3165
3166                 memset(&log, 0, sizeof(log));
3167                 log.u_bbr.flex1 = logged;
3168                 log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3169                 log.u_bbr.flex2 <<= 4;
3170                 log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3171                 log.u_bbr.flex2 <<= 4;
3172                 log.u_bbr.flex2 |= rack->rc_gp_incr;
3173                 log.u_bbr.flex2 <<= 4;
3174                 log.u_bbr.flex2 |= rack->rc_gp_bwred;
3175                 log.u_bbr.flex3 = rack->rc_gp_incr;
3176                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3177                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3178                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3179                 log.u_bbr.flex7 = rack->rc_gp_bwred;
3180                 log.u_bbr.flex8 = method;
3181                 log.u_bbr.cur_del_rate = cur_bw;
3182                 log.u_bbr.delRate = low_bnd;
3183                 log.u_bbr.bw_inuse = up_bnd;
3184                 log.u_bbr.rttProp = rack_get_bw(rack);
3185                 log.u_bbr.pkt_epoch = line;
3186                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3187                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3188                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3189                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3190                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3191                 log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3192                 log.u_bbr.cwnd_gain <<= 1;
3193                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3194                 log.u_bbr.cwnd_gain <<= 1;
3195                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3196                 log.u_bbr.cwnd_gain <<= 1;
3197                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3198                 log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3199                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3200                     &rack->rc_inp->inp_socket->so_rcv,
3201                     &rack->rc_inp->inp_socket->so_snd,
3202                     TCP_TIMELY_WORK, 0,
3203                     0, &log, false, &tv);
3204         }
3205 }
3206
3207 static int
3208 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3209 {
3210         /*
3211          * Before we increase we need to know if
3212          * the estimate just made was less than
3213          * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3214          *
3215          * If we already are pacing at a fast enough
3216          * rate to push us faster there is no sense of
3217          * increasing.
3218          *
3219          * We first caculate our actual pacing rate (ss or ca multipler
3220          * times our cur_bw).
3221          *
3222          * Then we take the last measured rate and multipy by our
3223          * maximum pacing overage to give us a max allowable rate.
3224          *
3225          * If our act_rate is smaller than our max_allowable rate
3226          * then we should increase. Else we should hold steady.
3227          *
3228          */
3229         uint64_t act_rate, max_allow_rate;
3230
3231         if (rack_timely_no_stopping)
3232                 return (1);
3233
3234         if ((cur_bw == 0) || (last_bw_est == 0)) {
3235                 /*
3236                  * Initial startup case or
3237                  * everything is acked case.
3238                  */
3239                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3240                                 __LINE__, 9);
3241                 return (1);
3242         }
3243         if (mult <= 100) {
3244                 /*
3245                  * We can always pace at or slightly above our rate.
3246                  */
3247                 rack_log_timely(rack,  mult, cur_bw, 0, 0,
3248                                 __LINE__, 9);
3249                 return (1);
3250         }
3251         act_rate = cur_bw * (uint64_t)mult;
3252         act_rate /= 100;
3253         max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3254         max_allow_rate /= 100;
3255         if (act_rate < max_allow_rate) {
3256                 /*
3257                  * Here the rate we are actually pacing at
3258                  * is smaller than 10% above our last measurement.
3259                  * This means we are pacing below what we would
3260                  * like to try to achieve (plus some wiggle room).
3261                  */
3262                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3263                                 __LINE__, 9);
3264                 return (1);
3265         } else {
3266                 /*
3267                  * Here we are already pacing at least rack_max_per_above(10%)
3268                  * what we are getting back. This indicates most likely
3269                  * that we are being limited (cwnd/rwnd/app) and can't
3270                  * get any more b/w. There is no sense of trying to
3271                  * raise up the pacing rate its not speeding us up
3272                  * and we already are pacing faster than we are getting.
3273                  */
3274                 rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3275                                 __LINE__, 8);
3276                 return (0);
3277         }
3278 }
3279
3280 static void
3281 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3282 {
3283         /*
3284          * When we drag bottom, we want to assure
3285          * that no multiplier is below 1.0, if so
3286          * we want to restore it to at least that.
3287          */
3288         if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3289                 /* This is unlikely we usually do not touch recovery */
3290                 rack->r_ctl.rack_per_of_gp_rec = 100;
3291         }
3292         if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3293                 rack->r_ctl.rack_per_of_gp_ca = 100;
3294         }
3295         if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3296                 rack->r_ctl.rack_per_of_gp_ss = 100;
3297         }
3298 }
3299
3300 static void
3301 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3302 {
3303         if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3304                 rack->r_ctl.rack_per_of_gp_ca = 100;
3305         }
3306         if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3307                 rack->r_ctl.rack_per_of_gp_ss = 100;
3308         }
3309 }
3310
3311 static void
3312 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3313 {
3314         int32_t  calc, logged, plus;
3315
3316         logged = 0;
3317
3318         if (override) {
3319                 /*
3320                  * override is passed when we are
3321                  * loosing b/w and making one last
3322                  * gasp at trying to not loose out
3323                  * to a new-reno flow.
3324                  */
3325                 goto extra_boost;
3326         }
3327         /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3328         if (rack->rc_gp_incr &&
3329             ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3330                 /*
3331                  * Reset and get 5 strokes more before the boost. Note
3332                  * that the count is 0 based so we have to add one.
3333                  */
3334 extra_boost:
3335                 plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3336                 rack->rc_gp_timely_inc_cnt = 0;
3337         } else
3338                 plus = (uint32_t)rack_gp_increase_per;
3339         /* Must be at least 1% increase for true timely increases */
3340         if ((plus < 1) &&
3341             ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3342                 plus = 1;
3343         if (rack->rc_gp_saw_rec &&
3344             (rack->rc_gp_no_rec_chg == 0) &&
3345             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3346                                   rack->r_ctl.rack_per_of_gp_rec)) {
3347                 /* We have been in recovery ding it too */
3348                 calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3349                 if (calc > 0xffff)
3350                         calc = 0xffff;
3351                 logged |= 1;
3352                 rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3353                 if (rack_per_upper_bound_ss &&
3354                     (rack->rc_dragged_bottom == 0) &&
3355                     (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3356                         rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3357         }
3358         if (rack->rc_gp_saw_ca &&
3359             (rack->rc_gp_saw_ss == 0) &&
3360             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3361                                   rack->r_ctl.rack_per_of_gp_ca)) {
3362                 /* In CA */
3363                 calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3364                 if (calc > 0xffff)
3365                         calc = 0xffff;
3366                 logged |= 2;
3367                 rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3368                 if (rack_per_upper_bound_ca &&
3369                     (rack->rc_dragged_bottom == 0) &&
3370                     (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3371                         rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3372         }
3373         if (rack->rc_gp_saw_ss &&
3374             rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3375                                   rack->r_ctl.rack_per_of_gp_ss)) {
3376                 /* In SS */
3377                 calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3378                 if (calc > 0xffff)
3379                         calc = 0xffff;
3380                 rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3381                 if (rack_per_upper_bound_ss &&
3382                     (rack->rc_dragged_bottom == 0) &&
3383                     (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3384                         rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3385                 logged |= 4;
3386         }
3387         if (logged &&
3388             (rack->rc_gp_incr == 0)){
3389                 /* Go into increment mode */
3390                 rack->rc_gp_incr = 1;
3391                 rack->rc_gp_timely_inc_cnt = 0;
3392         }
3393         if (rack->rc_gp_incr &&
3394             logged &&
3395             (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3396                 rack->rc_gp_timely_inc_cnt++;
3397         }
3398         rack_log_timely(rack,  logged, plus, 0, 0,
3399                         __LINE__, 1);
3400 }
3401
3402 static uint32_t
3403 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3404 {
3405         /*
3406          * norm_grad = rtt_diff / minrtt;
3407          * new_per = curper * (1 - B * norm_grad)
3408          *
3409          * B = rack_gp_decrease_per (default 10%)
3410          * rtt_dif = input var current rtt-diff
3411          * curper = input var current percentage
3412          * minrtt = from rack filter
3413          *
3414          */
3415         uint64_t perf;
3416
3417         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3418                     ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3419                      (((uint64_t)rtt_diff * (uint64_t)1000000)/
3420                       (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3421                      (uint64_t)1000000)) /
3422                 (uint64_t)1000000);
3423         if (perf > curper) {
3424                 /* TSNH */
3425                 perf = curper - 1;
3426         }
3427         return ((uint32_t)perf);
3428 }
3429
3430 static uint32_t
3431 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3432 {
3433         /*
3434          *                                   highrttthresh
3435          * result = curper * (1 - (B * ( 1 -  ------          ))
3436          *                                     gp_srtt
3437          *
3438          * B = rack_gp_decrease_per (default 10%)
3439          * highrttthresh = filter_min * rack_gp_rtt_maxmul
3440          */
3441         uint64_t perf;
3442         uint32_t highrttthresh;
3443
3444         highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3445
3446         perf = (((uint64_t)curper * ((uint64_t)1000000 -
3447                                      ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3448                                         ((uint64_t)highrttthresh * (uint64_t)1000000) /
3449                                                     (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3450         return (perf);
3451 }
3452
3453 static void
3454 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3455 {
3456         uint64_t logvar, logvar2, logvar3;
3457         uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3458
3459         if (rack->rc_gp_incr) {
3460                 /* Turn off increment counting */
3461                 rack->rc_gp_incr = 0;
3462                 rack->rc_gp_timely_inc_cnt = 0;
3463         }
3464         ss_red = ca_red = rec_red = 0;
3465         logged = 0;
3466         /* Calculate the reduction value */
3467         if (rtt_diff < 0) {
3468                 rtt_diff *= -1;
3469         }
3470         /* Must be at least 1% reduction */
3471         if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3472                 /* We have been in recovery ding it too */
3473                 if (timely_says == 2) {
3474                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3475                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3476                         if (alt < new_per)
3477                                 val = alt;
3478                         else
3479                                 val = new_per;
3480                 } else
3481                          val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3482                 if (rack->r_ctl.rack_per_of_gp_rec > val) {
3483                         rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3484                         rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3485                 } else {
3486                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3487                         rec_red = 0;
3488                 }
3489                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3490                         rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3491                 logged |= 1;
3492         }
3493         if (rack->rc_gp_saw_ss) {
3494                 /* Sent in SS */
3495                 if (timely_says == 2) {
3496                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3497                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3498                         if (alt < new_per)
3499                                 val = alt;
3500                         else
3501                                 val = new_per;
3502                 } else
3503                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3504                 if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3505                         ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3506                         rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3507                 } else {
3508                         ss_red = new_per;
3509                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3510                         logvar = new_per;
3511                         logvar <<= 32;
3512                         logvar |= alt;
3513                         logvar2 = (uint32_t)rtt;
3514                         logvar2 <<= 32;
3515                         logvar2 |= (uint32_t)rtt_diff;
3516                         logvar3 = rack_gp_rtt_maxmul;
3517                         logvar3 <<= 32;
3518                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3519                         rack_log_timely(rack, timely_says,
3520                                         logvar2, logvar3,
3521                                         logvar, __LINE__, 10);
3522                 }
3523                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3524                         rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3525                 logged |= 4;
3526         } else if (rack->rc_gp_saw_ca) {
3527                 /* Sent in CA */
3528                 if (timely_says == 2) {
3529                         new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3530                         alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3531                         if (alt < new_per)
3532                                 val = alt;
3533                         else
3534                                 val = new_per;
3535                 } else
3536                         val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3537                 if (rack->r_ctl.rack_per_of_gp_ca > val) {
3538                         ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3539                         rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3540                 } else {
3541                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3542                         ca_red = 0;
3543                         logvar = new_per;
3544                         logvar <<= 32;
3545                         logvar |= alt;
3546                         logvar2 = (uint32_t)rtt;
3547                         logvar2 <<= 32;
3548                         logvar2 |= (uint32_t)rtt_diff;
3549                         logvar3 = rack_gp_rtt_maxmul;
3550                         logvar3 <<= 32;
3551                         logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3552                         rack_log_timely(rack, timely_says,
3553                                         logvar2, logvar3,
3554                                         logvar, __LINE__, 10);
3555                 }
3556                 if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3557                         rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3558                 logged |= 2;
3559         }
3560         if (rack->rc_gp_timely_dec_cnt < 0x7) {
3561                 rack->rc_gp_timely_dec_cnt++;
3562                 if (rack_timely_dec_clear &&
3563                     (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3564                         rack->rc_gp_timely_dec_cnt = 0;
3565         }
3566         logvar = ss_red;
3567         logvar <<= 32;
3568         logvar |= ca_red;
3569         rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3570                         __LINE__, 2);
3571 }
3572
3573 static void
3574 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3575                      uint32_t rtt, uint32_t line, uint8_t reas)
3576 {
3577         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3578                 union tcp_log_stackspecific log;
3579                 struct timeval tv;
3580
3581                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3582                 log.u_bbr.flex1 = line;
3583                 log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3584                 log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3585                 log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3586                 log.u_bbr.flex5 = rtt;
3587                 log.u_bbr.flex6 = rack->rc_highly_buffered;
3588                 log.u_bbr.flex6 <<= 1;
3589                 log.u_bbr.flex6 |= rack->forced_ack;
3590                 log.u_bbr.flex6 <<= 1;
3591                 log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3592                 log.u_bbr.flex6 <<= 1;
3593                 log.u_bbr.flex6 |= rack->in_probe_rtt;
3594                 log.u_bbr.flex6 <<= 1;
3595                 log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3596                 log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3597                 log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3598                 log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3599                 log.u_bbr.flex8 = reas;
3600                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3601                 log.u_bbr.delRate = rack_get_bw(rack);
3602                 log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3603                 log.u_bbr.cur_del_rate <<= 32;
3604                 log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3605                 log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3606                 log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3607                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3608                 log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3609                 log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3610                 log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3611                 log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3612                 log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3613                 log.u_bbr.rttProp = us_cts;
3614                 log.u_bbr.rttProp <<= 32;
3615                 log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3616                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
3617                     &rack->rc_inp->inp_socket->so_rcv,
3618                     &rack->rc_inp->inp_socket->so_snd,
3619                     BBR_LOG_RTT_SHRINKS, 0,
3620                     0, &log, false, &rack->r_ctl.act_rcv_time);
3621         }
3622 }
3623
3624 static void
3625 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3626 {
3627         uint64_t bwdp;
3628
3629         bwdp = rack_get_bw(rack);
3630         bwdp *= (uint64_t)rtt;
3631         bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3632         rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3633         if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3634                 /*
3635                  * A window protocol must be able to have 4 packets
3636                  * outstanding as the floor in order to function
3637                  * (especially considering delayed ack :D).
3638                  */
3639                 rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3640         }
3641 }
3642
3643 static void
3644 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3645 {
3646         /**
3647          * ProbeRTT is a bit different in rack_pacing than in
3648          * BBR. It is like BBR in that it uses the lowering of
3649          * the RTT as a signal that we saw something new and
3650          * counts from there for how long between. But it is
3651          * different in that its quite simple. It does not
3652          * play with the cwnd and wait until we get down
3653          * to N segments outstanding and hold that for
3654          * 200ms. Instead it just sets the pacing reduction
3655          * rate to a set percentage (70 by default) and hold
3656          * that for a number of recent GP Srtt's.
3657          */
3658         uint32_t segsiz;
3659
3660         if (rack->rc_gp_dyn_mul == 0)
3661                 return;
3662
3663         if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3664                 /* We are idle */
3665                 return;
3666         }
3667         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3668             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3669                 /*
3670                  * Stop the goodput now, the idea here is
3671                  * that future measurements with in_probe_rtt
3672                  * won't register if they are not greater so
3673                  * we want to get what info (if any) is available
3674                  * now.
3675                  */
3676                 rack_do_goodput_measurement(rack->rc_tp, rack,
3677                                             rack->rc_tp->snd_una, __LINE__);
3678         }
3679         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3680         rack->r_ctl.rc_time_probertt_entered = us_cts;
3681         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3682                      rack->r_ctl.rc_pace_min_segs);
3683         rack->in_probe_rtt = 1;
3684         rack->measure_saw_probe_rtt = 1;
3685         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3686         rack->r_ctl.rc_time_probertt_starts = 0;
3687         rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3688         if (rack_probertt_use_min_rtt_entry)
3689                 rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3690         else
3691                 rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3692         rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3693                              __LINE__, RACK_RTTS_ENTERPROBE);
3694 }
3695
3696 static void
3697 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3698 {
3699         struct rack_sendmap *rsm;
3700         uint32_t segsiz;
3701
3702         segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3703                      rack->r_ctl.rc_pace_min_segs);
3704         rack->in_probe_rtt = 0;
3705         if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3706             SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3707                 /*
3708                  * Stop the goodput now, the idea here is
3709                  * that future measurements with in_probe_rtt
3710                  * won't register if they are not greater so
3711                  * we want to get what info (if any) is available
3712                  * now.
3713                  */
3714                 rack_do_goodput_measurement(rack->rc_tp, rack,
3715                                             rack->rc_tp->snd_una, __LINE__);
3716         } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3717                 /*
3718                  * We don't have enough data to make a measurement.
3719                  * So lets just stop and start here after exiting
3720                  * probe-rtt. We probably are not interested in
3721                  * the results anyway.
3722                  */
3723                 rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3724         }
3725         /*
3726          * Measurements through the current snd_max are going
3727          * to be limited by the slower pacing rate.
3728          *
3729          * We need to mark these as app-limited so we
3730          * don't collapse the b/w.
3731          */
3732         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3733         if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3734                 if (rack->r_ctl.rc_app_limited_cnt == 0)
3735                         rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3736                 else {
3737                         /*
3738                          * Go out to the end app limited and mark
3739                          * this new one as next and move the end_appl up
3740                          * to this guy.
3741                          */
3742                         if (rack->r_ctl.rc_end_appl)
3743                                 rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3744                         rack->r_ctl.rc_end_appl = rsm;
3745                 }
3746                 rsm->r_flags |= RACK_APP_LIMITED;
3747                 rack->r_ctl.rc_app_limited_cnt++;
3748         }
3749         /*
3750          * Now, we need to examine our pacing rate multipliers.
3751          * If its under 100%, we need to kick it back up to
3752          * 100%. We also don't let it be over our "max" above
3753          * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3754          * Note setting clamp_atexit_prtt to 0 has the effect
3755          * of setting CA/SS to 100% always at exit (which is
3756          * the default behavior).
3757          */
3758         if (rack_probertt_clear_is) {
3759                 rack->rc_gp_incr = 0;
3760                 rack->rc_gp_bwred = 0;
3761                 rack->rc_gp_timely_inc_cnt = 0;
3762                 rack->rc_gp_timely_dec_cnt = 0;
3763         }
3764         /* Do we do any clamping at exit? */
3765         if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3766                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3767                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3768         }
3769         if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3770                 rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3771                 rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3772         }
3773         /*
3774          * Lets set rtt_diff to 0, so that we will get a "boost"
3775          * after exiting.
3776          */
3777         rack->r_ctl.rc_rtt_diff = 0;
3778
3779         /* Clear all flags so we start fresh */
3780         rack->rc_tp->t_bytes_acked = 0;
3781         rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3782         /*
3783          * If configured to, set the cwnd and ssthresh to
3784          * our targets.
3785          */
3786         if (rack_probe_rtt_sets_cwnd) {
3787                 uint64_t ebdp;
3788                 uint32_t setto;
3789
3790                 /* Set ssthresh so we get into CA once we hit our target */
3791                 if (rack_probertt_use_min_rtt_exit == 1) {
3792                         /* Set to min rtt */
3793                         rack_set_prtt_target(rack, segsiz,
3794                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3795                 } else if (rack_probertt_use_min_rtt_exit == 2) {
3796                         /* Set to current gp rtt */
3797                         rack_set_prtt_target(rack, segsiz,
3798                                              rack->r_ctl.rc_gp_srtt);
3799                 } else if (rack_probertt_use_min_rtt_exit == 3) {
3800                         /* Set to entry gp rtt */
3801                         rack_set_prtt_target(rack, segsiz,
3802                                              rack->r_ctl.rc_entry_gp_rtt);
3803                 } else {
3804                         uint64_t sum;
3805                         uint32_t setval;
3806
3807                         sum = rack->r_ctl.rc_entry_gp_rtt;
3808                         sum *= 10;
3809                         sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3810                         if (sum >= 20) {
3811                                 /*
3812                                  * A highly buffered path needs
3813                                  * cwnd space for timely to work.
3814                                  * Lets set things up as if
3815                                  * we are heading back here again.
3816                                  */
3817                                 setval = rack->r_ctl.rc_entry_gp_rtt;
3818                         } else if (sum >= 15) {
3819                                 /*
3820                                  * Lets take the smaller of the
3821                                  * two since we are just somewhat
3822                                  * buffered.
3823                                  */
3824                                 setval = rack->r_ctl.rc_gp_srtt;
3825                                 if (setval > rack->r_ctl.rc_entry_gp_rtt)
3826                                         setval = rack->r_ctl.rc_entry_gp_rtt;
3827                         } else {
3828                                 /*
3829                                  * Here we are not highly buffered
3830                                  * and should pick the min we can to
3831                                  * keep from causing loss.
3832                                  */
3833                                 setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3834                         }
3835                         rack_set_prtt_target(rack, segsiz,
3836                                              setval);
3837                 }
3838                 if (rack_probe_rtt_sets_cwnd > 1) {
3839                         /* There is a percentage here to boost */
3840                         ebdp = rack->r_ctl.rc_target_probertt_flight;
3841                         ebdp *= rack_probe_rtt_sets_cwnd;
3842                         ebdp /= 100;
3843                         setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3844                 } else
3845                         setto = rack->r_ctl.rc_target_probertt_flight;
3846                 rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3847                 if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3848                         /* Enforce a min */
3849                         rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3850                 }
3851                 /* If we set in the cwnd also set the ssthresh point so we are in CA */
3852                 rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3853         }
3854         rack_log_rtt_shrinks(rack,  us_cts,
3855                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3856                              __LINE__, RACK_RTTS_EXITPROBE);
3857         /* Clear times last so log has all the info */
3858         rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3859         rack->r_ctl.rc_time_probertt_entered = us_cts;
3860         rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3861         rack->r_ctl.rc_time_of_last_probertt = us_cts;
3862 }
3863
3864 static void
3865 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3866 {
3867         /* Check in on probe-rtt */
3868         if (rack->rc_gp_filled == 0) {
3869                 /* We do not do p-rtt unless we have gp measurements */
3870                 return;
3871         }
3872         if (rack->in_probe_rtt) {
3873                 uint64_t no_overflow;
3874                 uint32_t endtime, must_stay;
3875
3876                 if (rack->r_ctl.rc_went_idle_time &&
3877                     ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3878                         /*
3879                          * We went idle during prtt, just exit now.
3880                          */
3881                         rack_exit_probertt(rack, us_cts);
3882                 } else if (rack_probe_rtt_safety_val &&
3883                     TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3884                     ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3885                         /*
3886                          * Probe RTT safety value triggered!
3887                          */
3888                         rack_log_rtt_shrinks(rack,  us_cts,
3889                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3890                                              __LINE__, RACK_RTTS_SAFETY);
3891                         rack_exit_probertt(rack, us_cts);
3892                 }
3893                 /* Calculate the max we will wait */
3894                 endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3895                 if (rack->rc_highly_buffered)
3896                         endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3897                 /* Calculate the min we must wait */
3898                 must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3899                 if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3900                     TSTMP_LT(us_cts, endtime)) {
3901                         uint32_t calc;
3902                         /* Do we lower more? */
3903 no_exit:
3904                         if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3905                                 calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3906                         else
3907                                 calc = 0;
3908                         calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3909                         if (calc) {
3910                                 /* Maybe */
3911                                 calc *= rack_per_of_gp_probertt_reduce;
3912                                 rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3913                                 /* Limit it too */
3914                                 if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3915                                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3916                         }
3917                         /* We must reach target or the time set */
3918                         return;
3919                 }
3920                 if (rack->r_ctl.rc_time_probertt_starts == 0) {
3921                         if ((TSTMP_LT(us_cts, must_stay) &&
3922                              rack->rc_highly_buffered) ||
3923                              (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3924                               rack->r_ctl.rc_target_probertt_flight)) {
3925                                 /* We are not past the must_stay time */
3926                                 goto no_exit;
3927                         }
3928                         rack_log_rtt_shrinks(rack,  us_cts,
3929                                              get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3930                                              __LINE__, RACK_RTTS_REACHTARGET);
3931                         rack->r_ctl.rc_time_probertt_starts = us_cts;
3932                         if (rack->r_ctl.rc_time_probertt_starts == 0)
3933                                 rack->r_ctl.rc_time_probertt_starts = 1;
3934                         /* Restore back to our rate we want to pace at in prtt */
3935                         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3936                 }
3937                 /*
3938                  * Setup our end time, some number of gp_srtts plus 200ms.
3939                  */
3940                 no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3941                                (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3942                 if (rack_probertt_gpsrtt_cnt_div)
3943                         endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3944                 else
3945                         endtime = 0;
3946                 endtime += rack_min_probertt_hold;
3947                 endtime += rack->r_ctl.rc_time_probertt_starts;
3948                 if (TSTMP_GEQ(us_cts,  endtime)) {
3949                         /* yes, exit probertt */
3950                         rack_exit_probertt(rack, us_cts);
3951                 }
3952
3953         } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3954                 /* Go into probertt, its been too long since we went lower */
3955                 rack_enter_probertt(rack, us_cts);
3956         }
3957 }
3958
3959 static void
3960 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3961                        uint32_t rtt, int32_t rtt_diff)
3962 {
3963         uint64_t cur_bw, up_bnd, low_bnd, subfr;
3964         uint32_t losses;
3965
3966         if ((rack->rc_gp_dyn_mul == 0) ||
3967             (rack->use_fixed_rate) ||
3968             (rack->in_probe_rtt) ||
3969             (rack->rc_always_pace == 0)) {
3970                 /* No dynamic GP multipler in play */
3971                 return;
3972         }
3973         losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3974         cur_bw = rack_get_bw(rack);
3975         /* Calculate our up and down range */
3976         up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3977         up_bnd /= 100;
3978         up_bnd += rack->r_ctl.last_gp_comp_bw;
3979
3980         subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3981         subfr /= 100;
3982         low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3983         if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3984                 /*
3985                  * This is the case where our RTT is above
3986                  * the max target and we have been configured
3987                  * to just do timely no bonus up stuff in that case.
3988                  *
3989                  * There are two configurations, set to 1, and we
3990                  * just do timely if we are over our max. If its
3991                  * set above 1 then we slam the multipliers down
3992                  * to 100 and then decrement per timely.
3993                  */
3994                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3995                                 __LINE__, 3);
3996                 if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3997                         rack_validate_multipliers_at_or_below_100(rack);
3998                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3999         } else if ((last_bw_est < low_bnd) && !losses) {
4000                 /*
4001                  * We are decreasing this is a bit complicated this
4002                  * means we are loosing ground. This could be
4003                  * because another flow entered and we are competing
4004                  * for b/w with it. This will push the RTT up which
4005                  * makes timely unusable unless we want to get shoved
4006                  * into a corner and just be backed off (the age
4007                  * old problem with delay based CC).
4008                  *
4009                  * On the other hand if it was a route change we
4010                  * would like to stay somewhat contained and not
4011                  * blow out the buffers.
4012                  */
4013                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4014                                 __LINE__, 3);
4015                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4016                 if (rack->rc_gp_bwred == 0) {
4017                         /* Go into reduction counting */
4018                         rack->rc_gp_bwred = 1;
4019                         rack->rc_gp_timely_dec_cnt = 0;
4020                 }
4021                 if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
4022                     (timely_says == 0)) {
4023                         /*
4024                          * Push another time with a faster pacing
4025                          * to try to gain back (we include override to
4026                          * get a full raise factor).
4027                          */
4028                         if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4029                             (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4030                             (timely_says == 0) ||
4031                             (rack_down_raise_thresh == 0)) {
4032                                 /*
4033                                  * Do an override up in b/w if we were
4034                                  * below the threshold or if the threshold
4035                                  * is zero we always do the raise.
4036                                  */
4037                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4038                         } else {
4039                                 /* Log it stays the same */
4040                                 rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4041                                                 __LINE__, 11);
4042                         }
4043                         rack->rc_gp_timely_dec_cnt++;
4044                         /* We are not incrementing really no-count */
4045                         rack->rc_gp_incr = 0;
4046                         rack->rc_gp_timely_inc_cnt = 0;
4047                 } else {
4048                         /*
4049                          * Lets just use the RTT
4050                          * information and give up
4051                          * pushing.
4052                          */
4053                         goto use_timely;
4054                 }
4055         } else if ((timely_says != 2) &&
4056                     !losses &&
4057                     (last_bw_est > up_bnd)) {
4058                 /*
4059                  * We are increasing b/w lets keep going, updating
4060                  * our b/w and ignoring any timely input, unless
4061                  * of course we are at our max raise (if there is one).
4062                  */
4063
4064                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4065                                 __LINE__, 3);
4066                 rack->r_ctl.last_gp_comp_bw = cur_bw;
4067                 if (rack->rc_gp_saw_ss &&
4068                     rack_per_upper_bound_ss &&
4069                      (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
4070                             /*
4071                              * In cases where we can't go higher
4072                              * we should just use timely.
4073                              */
4074                             goto use_timely;
4075                 }
4076                 if (rack->rc_gp_saw_ca &&
4077                     rack_per_upper_bound_ca &&
4078                     (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4079                             /*
4080                              * In cases where we can't go higher
4081                              * we should just use timely.
4082                              */
4083                             goto use_timely;
4084                 }
4085                 rack->rc_gp_bwred = 0;
4086                 rack->rc_gp_timely_dec_cnt = 0;
4087                 /* You get a set number of pushes if timely is trying to reduce */
4088                 if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4089                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4090                 } else {
4091                         /* Log it stays the same */
4092                         rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4093                             __LINE__, 12);
4094                 }
4095                 return;
4096         } else {
4097                 /*
4098                  * We are staying between the lower and upper range bounds
4099                  * so use timely to decide.
4100                  */
4101                 rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4102                                 __LINE__, 3);
4103 use_timely:
4104                 if (timely_says) {
4105                         rack->rc_gp_incr = 0;
4106                         rack->rc_gp_timely_inc_cnt = 0;
4107                         if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4108                             !losses &&
4109                             (last_bw_est < low_bnd)) {
4110                                 /* We are loosing ground */
4111                                 rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4112                                 rack->rc_gp_timely_dec_cnt++;
4113                                 /* We are not incrementing really no-count */
4114                                 rack->rc_gp_incr = 0;
4115                                 rack->rc_gp_timely_inc_cnt = 0;
4116                         } else
4117                                 rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4118                 } else {
4119                         rack->rc_gp_bwred = 0;
4120                         rack->rc_gp_timely_dec_cnt = 0;
4121                         rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4122                 }
4123         }
4124 }
4125
4126 static int32_t
4127 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4128 {
4129         int32_t timely_says;
4130         uint64_t log_mult, log_rtt_a_diff;
4131
4132         log_rtt_a_diff = rtt;
4133         log_rtt_a_diff <<= 32;
4134         log_rtt_a_diff |= (uint32_t)rtt_diff;
4135         if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4136                     rack_gp_rtt_maxmul)) {
4137                 /* Reduce the b/w multipler */
4138                 timely_says = 2;
4139                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4140                 log_mult <<= 32;
4141                 log_mult |= prev_rtt;
4142                 rack_log_timely(rack,  timely_says, log_mult,
4143                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4144                                 log_rtt_a_diff, __LINE__, 4);
4145         } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4146                            ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4147                             max(rack_gp_rtt_mindiv , 1)))) {
4148                 /* Increase the b/w multipler */
4149                 log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4150                         ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4151                          max(rack_gp_rtt_mindiv , 1));
4152                 log_mult <<= 32;
4153                 log_mult |= prev_rtt;
4154                 timely_says = 0;
4155                 rack_log_timely(rack,  timely_says, log_mult ,
4156                                 get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4157                                 log_rtt_a_diff, __LINE__, 5);
4158         } else {
4159                 /*
4160                  * Use a gradient to find it the timely gradient
4161                  * is:
4162                  * grad = rc_rtt_diff / min_rtt;
4163                  *
4164                  * anything below or equal to 0 will be
4165                  * a increase indication. Anything above
4166                  * zero is a decrease. Note we take care
4167                  * of the actual gradient calculation
4168                  * in the reduction (its not needed for
4169                  * increase).
4170                  */
4171                 log_mult = prev_rtt;
4172                 if (rtt_diff <= 0) {
4173                         /*
4174                          * Rttdiff is less than zero, increase the
4175                          * b/w multipler (its 0 or negative)
4176                          */
4177                         timely_says = 0;
4178                         rack_log_timely(rack,  timely_says, log_mult,
4179                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4180                 } else {
4181                         /* Reduce the b/w multipler */
4182                         timely_says = 1;
4183                         rack_log_timely(rack,  timely_says, log_mult,
4184                                         get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4185                 }
4186         }
4187         return (timely_says);
4188 }
4189
4190 static void
4191 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4192                             tcp_seq th_ack, int line)
4193 {
4194         uint64_t tim, bytes_ps, ltim, stim, utim;
4195         uint32_t segsiz, bytes, reqbytes, us_cts;
4196         int32_t gput, new_rtt_diff, timely_says;
4197         uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4198         int did_add = 0;
4199
4200         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4201         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4202         if (TSTMP_GEQ(us_cts, tp->gput_ts))
4203                 tim = us_cts - tp->gput_ts;
4204         else
4205                 tim = 0;
4206
4207         if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4208                 stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4209         else
4210                 stim = 0;
4211         /*
4212          * Use the larger of the send time or ack time. This prevents us
4213          * from being influenced by ack artifacts to come up with too
4214          * high of measurement. Note that since we are spanning over many more
4215          * bytes in most of our measurements hopefully that is less likely to
4216          * occur.
4217          */
4218         if (tim > stim)
4219                 utim = max(tim, 1);
4220         else
4221                 utim = max(stim, 1);
4222         /* Lets get a msec time ltim too for the old stuff */
4223         ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4224         gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4225         reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4226         if ((tim == 0) && (stim == 0)) {
4227                 /*
4228                  * Invalid measurement time, maybe
4229                  * all on one ack/one send?
4230                  */
4231                 bytes = 0;
4232                 bytes_ps = 0;
4233                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4234                                            0, 0, 0, 10, __LINE__, NULL);
4235                 goto skip_measurement;
4236         }
4237         if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4238                 /* We never made a us_rtt measurement? */
4239                 bytes = 0;
4240                 bytes_ps = 0;
4241                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4242                                            0, 0, 0, 10, __LINE__, NULL);
4243                 goto skip_measurement;
4244         }
4245         /*
4246          * Calculate the maximum possible b/w this connection
4247          * could have. We base our calculation on the lowest
4248          * rtt we have seen during the measurement and the
4249          * largest rwnd the client has given us in that time. This
4250          * forms a BDP that is the maximum that we could ever
4251          * get to the client. Anything larger is not valid.
4252          *
4253          * I originally had code here that rejected measurements
4254          * where the time was less than 1/2 the latest us_rtt.
4255          * But after thinking on that I realized its wrong since
4256          * say you had a 150Mbps or even 1Gbps link, and you
4257          * were a long way away.. example I am in Europe (100ms rtt)
4258          * talking to my 1Gbps link in S.C. Now measuring say 150,000
4259          * bytes my time would be 1.2ms, and yet my rtt would say
4260          * the measurement was invalid the time was < 50ms. The
4261          * same thing is true for 150Mb (8ms of time).
4262          *
4263          * A better way I realized is to look at what the maximum
4264          * the connection could possibly do. This is gated on
4265          * the lowest RTT we have seen and the highest rwnd.
4266          * We should in theory never exceed that, if we are
4267          * then something on the path is storing up packets
4268          * and then feeding them all at once to our endpoint
4269          * messing up our measurement.
4270          */
4271         rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4272         rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4273         rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4274         if (SEQ_LT(th_ack, tp->gput_seq)) {
4275                 /* No measurement can be made */
4276                 bytes = 0;
4277                 bytes_ps = 0;
4278                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4279                                            0, 0, 0, 10, __LINE__, NULL);
4280                 goto skip_measurement;
4281         } else
4282                 bytes = (th_ack - tp->gput_seq);
4283         bytes_ps = (uint64_t)bytes;
4284         /*
4285          * Don't measure a b/w for pacing unless we have gotten at least
4286          * an initial windows worth of data in this measurement interval.
4287          *
4288          * Small numbers of bytes get badly influenced by delayed ack and
4289          * other artifacts. Note we take the initial window or our
4290          * defined minimum GP (defaulting to 10 which hopefully is the
4291          * IW).
4292          */
4293         if (rack->rc_gp_filled == 0) {
4294                 /*
4295                  * The initial estimate is special. We
4296                  * have blasted out an IW worth of packets
4297                  * without a real valid ack ts results. We
4298                  * then setup the app_limited_needs_set flag,
4299                  * this should get the first ack in (probably 2
4300                  * MSS worth) to be recorded as the timestamp.
4301                  * We thus allow a smaller number of bytes i.e.
4302                  * IW - 2MSS.
4303                  */
4304                 reqbytes -= (2 * segsiz);
4305                 /* Also lets fill previous for our first measurement to be neutral */
4306                 rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4307         }
4308         if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4309                 rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4310                                            rack->r_ctl.rc_app_limited_cnt,
4311                                            0, 0, 10, __LINE__, NULL);
4312                 goto skip_measurement;
4313         }
4314         /*
4315          * We now need to calculate the Timely like status so
4316          * we can update (possibly) the b/w multipliers.
4317          */
4318         new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4319         if (rack->rc_gp_filled == 0) {
4320                 /* No previous reading */
4321                 rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4322         } else {
4323                 if (rack->measure_saw_probe_rtt == 0) {
4324                         /*
4325                          * We don't want a probertt to be counted
4326                          * since it will be negative incorrectly. We
4327                          * expect to be reducing the RTT when we
4328                          * pace at a slower rate.
4329                          */
4330                         rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4331                         rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4332                 }
4333         }
4334         timely_says = rack_make_timely_judgement(rack,
4335                 rack->r_ctl.rc_gp_srtt,
4336                 rack->r_ctl.rc_rtt_diff,
4337                 rack->r_ctl.rc_prev_gp_srtt
4338                 );
4339         bytes_ps *= HPTS_USEC_IN_SEC;
4340         bytes_ps /= utim;
4341         if (bytes_ps > rack->r_ctl.last_max_bw) {
4342                 /*
4343                  * Something is on path playing
4344                  * since this b/w is not possible based
4345                  * on our BDP (highest rwnd and lowest rtt
4346                  * we saw in the measurement window).
4347                  *
4348                  * Another option here would be to
4349                  * instead skip the measurement.
4350                  */
4351                 rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4352                                            bytes_ps, rack->r_ctl.last_max_bw, 0,
4353                                            11, __LINE__, NULL);
4354                 bytes_ps = rack->r_ctl.last_max_bw;
4355         }
4356         /* We store gp for b/w in bytes per second */
4357         if (rack->rc_gp_filled == 0) {
4358                 /* Initial measurment */
4359                 if (bytes_ps) {
4360                         rack->r_ctl.gp_bw = bytes_ps;
4361                         rack->rc_gp_filled = 1;
4362                         rack->r_ctl.num_measurements = 1;
4363                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4364                 } else {
4365                         rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4366                                                    rack->r_ctl.rc_app_limited_cnt,
4367                                                    0, 0, 10, __LINE__, NULL);
4368                 }
4369                 if (rack->rc_inp->inp_in_hpts &&
4370                     (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4371                         /*
4372                          * Ok we can't trust the pacer in this case
4373                          * where we transition from un-paced to paced.
4374                          * Or for that matter when the burst mitigation
4375                          * was making a wild guess and got it wrong.
4376                          * Stop the pacer and clear up all the aggregate
4377                          * delays etc.
4378                          */
4379                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
4380                         rack->r_ctl.rc_hpts_flags = 0;
4381                         rack->r_ctl.rc_last_output_to = 0;
4382                 }
4383                 did_add = 2;
4384         } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4385                 /* Still a small number run an average */
4386                 rack->r_ctl.gp_bw += bytes_ps;
4387                 addpart = rack->r_ctl.num_measurements;
4388                 rack->r_ctl.num_measurements++;
4389                 if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4390                         /* We have collected enought to move forward */
4391                         rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4392                 }
4393                 did_add = 3;
4394         } else {
4395                 /*
4396                  * We want to take 1/wma of the goodput and add in to 7/8th
4397                  * of the old value weighted by the srtt. So if your measurement
4398                  * period is say 2 SRTT's long you would get 1/4 as the
4399                  * value, if it was like 1/2 SRTT then you would get 1/16th.
4400                  *
4401                  * But we must be careful not to take too much i.e. if the
4402                  * srtt is say 20ms and the measurement is taken over
4403                  * 400ms our weight would be 400/20 i.e. 20. On the
4404                  * other hand if we get a measurement over 1ms with a
4405                  * 10ms rtt we only want to take a much smaller portion.
4406                  */
4407                 if (rack->r_ctl.num_measurements < 0xff) {
4408                         rack->r_ctl.num_measurements++;
4409                 }
4410                 srtt = (uint64_t)tp->t_srtt;
4411                 if (srtt == 0) {
4412                         /*
4413                          * Strange why did t_srtt go back to zero?
4414                          */
4415                         if (rack->r_ctl.rc_rack_min_rtt)
4416                                 srtt = rack->r_ctl.rc_rack_min_rtt;
4417                         else
4418                                 srtt = HPTS_USEC_IN_MSEC;
4419                 }
4420                 /*
4421                  * XXXrrs: Note for reviewers, in playing with
4422                  * dynamic pacing I discovered this GP calculation
4423                  * as done originally leads to some undesired results.
4424                  * Basically you can get longer measurements contributing
4425                  * too much to the WMA. Thus I changed it if you are doing
4426                  * dynamic adjustments to only do the aportioned adjustment
4427                  * if we have a very small (time wise) measurement. Longer
4428                  * measurements just get there weight (defaulting to 1/8)
4429                  * add to the WMA. We may want to think about changing
4430                  * this to always do that for both sides i.e. dynamic
4431                  * and non-dynamic... but considering lots of folks
4432                  * were playing with this I did not want to change the
4433                  * calculation per.se. without your thoughts.. Lawerence?
4434                  * Peter??
4435                  */
4436                 if (rack->rc_gp_dyn_mul == 0) {
4437                         subpart = rack->r_ctl.gp_bw * utim;
4438                         subpart /= (srtt * 8);
4439                         if (subpart < (rack->r_ctl.gp_bw / 2)) {
4440                                 /*
4441                                  * The b/w update takes no more
4442                                  * away then 1/2 our running total
4443                                  * so factor it in.
4444                                  */
4445                                 addpart = bytes_ps * utim;
4446                                 addpart /= (srtt * 8);
4447                         } else {
4448                                 /*
4449                                  * Don't allow a single measurement
4450                                  * to account for more than 1/2 of the
4451                                  * WMA. This could happen on a retransmission
4452                                  * where utim becomes huge compared to
4453                                  * srtt (multiple retransmissions when using
4454                                  * the sending rate which factors in all the
4455                                  * transmissions from the first one).
4456                                  */
4457                                 subpart = rack->r_ctl.gp_bw / 2;
4458                                 addpart = bytes_ps / 2;
4459                         }
4460                         resid_bw = rack->r_ctl.gp_bw - subpart;
4461                         rack->r_ctl.gp_bw = resid_bw + addpart;
4462                         did_add = 1;
4463                 } else {
4464                         if ((utim / srtt) <= 1) {
4465                                 /*
4466                                  * The b/w update was over a small period
4467                                  * of time. The idea here is to prevent a small
4468                                  * measurement time period from counting
4469                                  * too much. So we scale it based on the
4470                                  * time so it attributes less than 1/rack_wma_divisor
4471                                  * of its measurement.
4472                                  */
4473                                 subpart = rack->r_ctl.gp_bw * utim;
4474                                 subpart /= (srtt * rack_wma_divisor);
4475                                 addpart = bytes_ps * utim;
4476                                 addpart /= (srtt * rack_wma_divisor);
4477                         } else {
4478                                 /*
4479                                  * The scaled measurement was long
4480                                  * enough so lets just add in the
4481                                  * portion of the measurment i.e. 1/rack_wma_divisor
4482                                  */
4483                                 subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4484                                 addpart = bytes_ps / rack_wma_divisor;
4485                         }
4486                         if ((rack->measure_saw_probe_rtt == 0) ||
4487                             (bytes_ps > rack->r_ctl.gp_bw)) {
4488                                 /*
4489                                  * For probe-rtt we only add it in
4490                                  * if its larger, all others we just
4491                                  * add in.
4492                                  */
4493                                 did_add = 1;
4494                                 resid_bw = rack->r_ctl.gp_bw - subpart;
4495                                 rack->r_ctl.gp_bw = resid_bw + addpart;
4496                         }
4497                 }
4498         }
4499         if ((rack->gp_ready == 0) &&
4500             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4501                 /* We have enough measurements now */
4502                 rack->gp_ready = 1;
4503                 rack_set_cc_pacing(rack);
4504                 if (rack->defer_options)
4505                         rack_apply_deferred_options(rack);
4506         }
4507         rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4508                                    rack_get_bw(rack), 22, did_add, NULL);
4509         /* We do not update any multipliers if we are in or have seen a probe-rtt */
4510         if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4511                 rack_update_multiplier(rack, timely_says, bytes_ps,
4512                                        rack->r_ctl.rc_gp_srtt,
4513                                        rack->r_ctl.rc_rtt_diff);
4514         rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4515                                    rack_get_bw(rack), 3, line, NULL);
4516         /* reset the gp srtt and setup the new prev */
4517         rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4518         /* Record the lost count for the next measurement */
4519         rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4520         /*
4521          * We restart our diffs based on the gpsrtt in the
4522          * measurement window.
4523          */
4524         rack->rc_gp_rtt_set = 0;
4525         rack->rc_gp_saw_rec = 0;
4526         rack->rc_gp_saw_ca = 0;
4527         rack->rc_gp_saw_ss = 0;
4528         rack->rc_dragged_bottom = 0;
4529 skip_measurement:
4530
4531 #ifdef STATS
4532         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4533                                  gput);
4534         /*
4535          * XXXLAS: This is a temporary hack, and should be
4536          * chained off VOI_TCP_GPUT when stats(9) grows an
4537          * API to deal with chained VOIs.
4538          */
4539         if (tp->t_stats_gput_prev > 0)
4540                 stats_voi_update_abs_s32(tp->t_stats,
4541                                          VOI_TCP_GPUT_ND,
4542                                          ((gput - tp->t_stats_gput_prev) * 100) /
4543                                          tp->t_stats_gput_prev);
4544 #endif
4545         tp->t_flags &= ~TF_GPUTINPROG;
4546         tp->t_stats_gput_prev = gput;
4547         /*
4548          * Now are we app limited now and there is space from where we
4549          * were to where we want to go?
4550          *
4551          * We don't do the other case i.e. non-applimited here since
4552          * the next send will trigger us picking up the missing data.
4553          */
4554         if (rack->r_ctl.rc_first_appl &&
4555             TCPS_HAVEESTABLISHED(tp->t_state) &&
4556             rack->r_ctl.rc_app_limited_cnt &&
4557             (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4558             ((rack->r_ctl.rc_first_appl->r_start - th_ack) >
4559              max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4560                 /*
4561                  * Yep there is enough outstanding to make a measurement here.
4562                  */
4563                 struct rack_sendmap *rsm, fe;
4564
4565                 tp->t_flags |= TF_GPUTINPROG;
4566                 rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4567                 rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4568                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4569                 rack->app_limited_needs_set = 0;
4570                 tp->gput_seq = th_ack;
4571                 if (rack->in_probe_rtt)
4572                         rack->measure_saw_probe_rtt = 1;
4573                 else if ((rack->measure_saw_probe_rtt) &&
4574                          (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4575                         rack->measure_saw_probe_rtt = 0;
4576                 if ((rack->r_ctl.rc_first_appl->r_start - th_ack) >= rack_get_measure_window(tp, rack)) {
4577                         /* There is a full window to gain info from */
4578                         tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4579                 } else {
4580                         /* We can only measure up to the applimited point */
4581                         tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_start - th_ack);
4582                 }
4583                 /*
4584                  * Now we need to find the timestamp of the send at tp->gput_seq
4585                  * for the send based measurement.
4586                  */
4587                 fe.r_start = tp->gput_seq;
4588                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4589                 if (rsm) {
4590                         /* Ok send-based limit is set */
4591                         if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4592                                 /*
4593                                  * Move back to include the earlier part
4594                                  * so our ack time lines up right (this may
4595                                  * make an overlapping measurement but thats
4596                                  * ok).
4597                                  */
4598                                 tp->gput_seq = rsm->r_start;
4599                         }
4600                         if (rsm->r_flags & RACK_ACKED)
4601                                 tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4602                         else
4603                                 rack->app_limited_needs_set = 1;
4604                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4605                 } else {
4606                         /*
4607                          * If we don't find the rsm due to some
4608                          * send-limit set the current time, which
4609                          * basically disables the send-limit.
4610                          */
4611                         struct timeval tv;
4612
4613                         microuptime(&tv);
4614                         rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4615                 }
4616                 rack_log_pacing_delay_calc(rack,
4617                                            tp->gput_seq,
4618                                            tp->gput_ack,
4619                                            (uint64_t)rsm,
4620                                            tp->gput_ts,
4621                                            rack->r_ctl.rc_app_limited_cnt,
4622                                            9,
4623                                            __LINE__, NULL);
4624         }
4625 }
4626
4627 /*
4628  * CC wrapper hook functions
4629  */
4630 static void
4631 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4632     uint16_t type, int32_t recovery)
4633 {
4634         uint32_t prior_cwnd, acked;
4635         struct tcp_log_buffer *lgb = NULL;
4636         uint8_t labc_to_use;
4637
4638         INP_WLOCK_ASSERT(tp->t_inpcb);
4639         tp->ccv->nsegs = nsegs;
4640         acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4641         if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4642                 uint32_t max;
4643
4644                 max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4645                 if (tp->ccv->bytes_this_ack > max) {
4646                         tp->ccv->bytes_this_ack = max;
4647                 }
4648         }
4649 #ifdef STATS
4650         stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4651             ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4652 #endif
4653         if ((tp->t_flags & TF_GPUTINPROG) &&
4654             rack_enough_for_measurement(tp, rack, th_ack)) {
4655                 /* Measure the Goodput */
4656                 rack_do_goodput_measurement(tp, rack, th_ack, __LINE__);
4657 #ifdef NETFLIX_PEAKRATE
4658                 if ((type == CC_ACK) &&
4659                     (tp->t_maxpeakrate)) {
4660                         /*
4661                          * We update t_peakrate_thr. This gives us roughly
4662                          * one update per round trip time. Note
4663                          * it will only be used if pace_always is off i.e
4664                          * we don't do this for paced flows.
4665                          */
4666                         rack_update_peakrate_thr(tp);
4667                 }
4668 #endif
4669         }
4670         /* Which way our we limited, if not cwnd limited no advance in CA */
4671         if (tp->snd_cwnd <= tp->snd_wnd)
4672                 tp->ccv->flags |= CCF_CWND_LIMITED;
4673         else
4674                 tp->ccv->flags &= ~CCF_CWND_LIMITED;
4675         if (tp->snd_cwnd > tp->snd_ssthresh) {
4676                 tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4677                          nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4678                 /* For the setting of a window past use the actual scwnd we are using */
4679                 if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4680                         tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4681                         tp->ccv->flags |= CCF_ABC_SENTAWND;
4682                 }
4683         } else {
4684                 tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4685                 tp->t_bytes_acked = 0;
4686         }
4687         prior_cwnd = tp->snd_cwnd;
4688         if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4689             (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4690                 labc_to_use = rack->rc_labc;
4691         else
4692                 labc_to_use = rack_max_abc_post_recovery;
4693         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4694                 union tcp_log_stackspecific log;
4695                 struct timeval tv;
4696
4697                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4698                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4699                 log.u_bbr.flex1 = th_ack;
4700                 log.u_bbr.flex2 = tp->ccv->flags;
4701                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4702                 log.u_bbr.flex4 = tp->ccv->nsegs;
4703                 log.u_bbr.flex5 = labc_to_use;
4704                 log.u_bbr.flex6 = prior_cwnd;
4705                 log.u_bbr.flex7 = V_tcp_do_newsack;
4706                 log.u_bbr.flex8 = 1;
4707                 lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4708                                      0, &log, false, NULL, NULL, 0, &tv);
4709         }
4710         if (CC_ALGO(tp)->ack_received != NULL) {
4711                 /* XXXLAS: Find a way to live without this */
4712                 tp->ccv->curack = th_ack;
4713                 tp->ccv->labc = labc_to_use;
4714                 tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4715                 CC_ALGO(tp)->ack_received(tp->ccv, type);
4716         }
4717         if (lgb) {
4718                 lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4719         }
4720         if (rack->r_must_retran) {
4721                 if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4722                         /*
4723                          * We now are beyond the rxt point so lets disable
4724                          * the flag.
4725                          */
4726                         rack->r_ctl.rc_out_at_rto = 0;
4727                         rack->r_must_retran = 0;
4728                 } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4729                         /*
4730                          * Only decrement the rc_out_at_rto if the cwnd advances
4731                          * at least a whole segment. Otherwise next time the peer
4732                          * acks, we won't be able to send this generaly happens
4733                          * when we are in Congestion Avoidance.
4734                          */
4735                         if (acked <= rack->r_ctl.rc_out_at_rto){
4736                                 rack->r_ctl.rc_out_at_rto -= acked;
4737                         } else {
4738                                 rack->r_ctl.rc_out_at_rto = 0;
4739                         }
4740                 }
4741         }
4742 #ifdef STATS
4743         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4744 #endif
4745         if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4746                 rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4747         }
4748 #ifdef NETFLIX_PEAKRATE
4749         /* we enforce max peak rate if it is set and we are not pacing */
4750         if ((rack->rc_always_pace == 0) &&
4751             tp->t_peakrate_thr &&
4752             (tp->snd_cwnd > tp->t_peakrate_thr)) {
4753                 tp->snd_cwnd = tp->t_peakrate_thr;
4754         }
4755 #endif
4756 }
4757
4758 static void
4759 tcp_rack_partialack(struct tcpcb *tp)
4760 {
4761         struct tcp_rack *rack;
4762
4763         rack = (struct tcp_rack *)tp->t_fb_ptr;
4764         INP_WLOCK_ASSERT(tp->t_inpcb);
4765         /*
4766          * If we are doing PRR and have enough
4767          * room to send <or> we are pacing and prr
4768          * is disabled we will want to see if we
4769          * can send data (by setting r_wanted_output to
4770          * true).
4771          */
4772         if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4773             rack->rack_no_prr)
4774                 rack->r_wanted_output = 1;
4775 }
4776
4777 static void
4778 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4779 {
4780         struct tcp_rack *rack;
4781         uint32_t orig_cwnd;
4782
4783         orig_cwnd = tp->snd_cwnd;
4784         INP_WLOCK_ASSERT(tp->t_inpcb);
4785         rack = (struct tcp_rack *)tp->t_fb_ptr;
4786         /* only alert CC if we alerted when we entered */
4787         if (CC_ALGO(tp)->post_recovery != NULL) {
4788                 tp->ccv->curack = th_ack;
4789                 CC_ALGO(tp)->post_recovery(tp->ccv);
4790                 if (tp->snd_cwnd < tp->snd_ssthresh) {
4791                         /*
4792                          * Rack has burst control and pacing
4793                          * so lets not set this any lower than
4794                          * snd_ssthresh per RFC-6582 (option 2).
4795                          */
4796                         tp->snd_cwnd = tp->snd_ssthresh;
4797                 }
4798         }
4799         if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4800                 union tcp_log_stackspecific log;
4801                 struct timeval tv;
4802
4803                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4804                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4805                 log.u_bbr.flex1 = th_ack;
4806                 log.u_bbr.flex2 = tp->ccv->flags;
4807                 log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4808                 log.u_bbr.flex4 = tp->ccv->nsegs;
4809                 log.u_bbr.flex5 = V_tcp_abc_l_var;
4810                 log.u_bbr.flex6 = orig_cwnd;
4811                 log.u_bbr.flex7 = V_tcp_do_newsack;
4812                 log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4813                 log.u_bbr.flex8 = 2;
4814                 tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4815                                0, &log, false, NULL, NULL, 0, &tv);
4816         }
4817         if ((rack->rack_no_prr == 0) &&
4818             (rack->no_prr_addback == 0) &&
4819             (rack->r_ctl.rc_prr_sndcnt > 0)) {
4820                 /*
4821                  * Suck the next prr cnt back into cwnd, but
4822                  * only do that if we are not application limited.
4823                  */
4824                 if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4825                         /*
4826                          * We are allowed to add back to the cwnd the amount we did
4827                          * not get out if:
4828                          * a) no_prr_addback is off.
4829                          * b) we are not app limited
4830                          * c) we are doing prr
4831                          * <and>
4832                          * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4833                          */
4834                         tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4835                                             rack->r_ctl.rc_prr_sndcnt);
4836                 }
4837                 rack->r_ctl.rc_prr_sndcnt = 0;
4838                 rack_log_to_prr(rack, 1, 0);
4839         }
4840         rack_log_to_prr(rack, 14, orig_cwnd);
4841         tp->snd_recover = tp->snd_una;
4842         EXIT_RECOVERY(tp->t_flags);
4843 }
4844
4845 static void
4846 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack)
4847 {
4848         struct tcp_rack *rack;
4849         uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4850
4851         INP_WLOCK_ASSERT(tp->t_inpcb);
4852 #ifdef STATS
4853         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4854 #endif
4855         if (IN_RECOVERY(tp->t_flags) == 0) {
4856                 in_rec_at_entry = 0;
4857                 ssthresh_enter = tp->snd_ssthresh;
4858                 cwnd_enter = tp->snd_cwnd;
4859         } else
4860                 in_rec_at_entry = 1;
4861         rack = (struct tcp_rack *)tp->t_fb_ptr;
4862         switch (type) {
4863         case CC_NDUPACK:
4864                 tp->t_flags &= ~TF_WASFRECOVERY;
4865                 tp->t_flags &= ~TF_WASCRECOVERY;
4866                 if (!IN_FASTRECOVERY(tp->t_flags)) {
4867                         rack->r_ctl.rc_prr_delivered = 0;
4868                         rack->r_ctl.rc_prr_out = 0;
4869                         if (rack->rack_no_prr == 0) {
4870                                 rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4871                                 rack_log_to_prr(rack, 2, in_rec_at_entry);
4872                         }
4873                         rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4874                         tp->snd_recover = tp->snd_max;
4875                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4876                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4877                 }
4878                 break;
4879         case CC_ECN:
4880                 if (!IN_CONGRECOVERY(tp->t_flags) ||
4881                     /*
4882                      * Allow ECN reaction on ACK to CWR, if
4883                      * that data segment was also CE marked.
4884                      */
4885                     SEQ_GEQ(ack, tp->snd_recover)) {
4886                         EXIT_CONGRECOVERY(tp->t_flags);
4887                         KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4888                         tp->snd_recover = tp->snd_max + 1;
4889                         if (tp->t_flags2 & TF2_ECN_PERMIT)
4890                                 tp->t_flags2 |= TF2_ECN_SND_CWR;
4891                 }
4892                 break;
4893         case CC_RTO:
4894                 tp->t_dupacks = 0;
4895                 tp->t_bytes_acked = 0;
4896                 EXIT_RECOVERY(tp->t_flags);
4897                 tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4898                     ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4899                 orig_cwnd = tp->snd_cwnd;
4900                 tp->snd_cwnd = ctf_fixed_maxseg(tp);
4901                 rack_log_to_prr(rack, 16, orig_cwnd);
4902                 if (tp->t_flags2 & TF2_ECN_PERMIT)
4903                         tp->t_flags2 |= TF2_ECN_SND_CWR;
4904                 break;
4905         case CC_RTO_ERR:
4906                 KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4907                 /* RTO was unnecessary, so reset everything. */
4908                 tp->snd_cwnd = tp->snd_cwnd_prev;
4909                 tp->snd_ssthresh = tp->snd_ssthresh_prev;
4910                 tp->snd_recover = tp->snd_recover_prev;
4911                 if (tp->t_flags & TF_WASFRECOVERY) {
4912                         ENTER_FASTRECOVERY(tp->t_flags);
4913                         tp->t_flags &= ~TF_WASFRECOVERY;
4914                 }
4915                 if (tp->t_flags & TF_WASCRECOVERY) {
4916                         ENTER_CONGRECOVERY(tp->t_flags);
4917                         tp->t_flags &= ~TF_WASCRECOVERY;
4918                 }
4919                 tp->snd_nxt = tp->snd_max;
4920                 tp->t_badrxtwin = 0;
4921                 break;
4922         }
4923         if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4924             (type != CC_RTO)){
4925                 tp->ccv->curack = ack;
4926                 CC_ALGO(tp)->cong_signal(tp->ccv, type);
4927         }
4928         if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4929                 rack_log_to_prr(rack, 15, cwnd_enter);
4930                 rack->r_ctl.dsack_byte_cnt = 0;
4931                 rack->r_ctl.retran_during_recovery = 0;
4932                 rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4933                 rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4934                 rack->r_ent_rec_ns = 1;
4935         }
4936 }
4937
4938 static inline void
4939 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4940 {
4941         uint32_t i_cwnd;
4942
4943         INP_WLOCK_ASSERT(tp->t_inpcb);
4944
4945 #ifdef NETFLIX_STATS
4946         KMOD_TCPSTAT_INC(tcps_idle_restarts);
4947         if (tp->t_state == TCPS_ESTABLISHED)
4948                 KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4949 #endif
4950         if (CC_ALGO(tp)->after_idle != NULL)
4951                 CC_ALGO(tp)->after_idle(tp->ccv);
4952
4953         if (tp->snd_cwnd == 1)
4954                 i_cwnd = tp->t_maxseg;          /* SYN(-ACK) lost */
4955         else
4956                 i_cwnd = rc_init_window(rack);
4957
4958         /*
4959          * Being idle is no differnt than the initial window. If the cc
4960          * clamps it down below the initial window raise it to the initial
4961          * window.
4962          */
4963         if (tp->snd_cwnd < i_cwnd) {
4964                 tp->snd_cwnd = i_cwnd;
4965         }
4966 }
4967
4968 /*
4969  * Indicate whether this ack should be delayed.  We can delay the ack if
4970  * following conditions are met:
4971  *      - There is no delayed ack timer in progress.
4972  *      - Our last ack wasn't a 0-sized window. We never want to delay
4973  *        the ack that opens up a 0-sized window.
4974  *      - LRO wasn't used for this segment. We make sure by checking that the
4975  *        segment size is not larger than the MSS.
4976  *      - Delayed acks are enabled or this is a half-synchronized T/TCP
4977  *        connection.
4978  */
4979 #define DELAY_ACK(tp, tlen)                      \
4980         (((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4981         ((tp->t_flags & TF_DELACK) == 0) &&      \
4982         (tlen <= tp->t_maxseg) &&                \
4983         (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4984
4985 static struct rack_sendmap *
4986 rack_find_lowest_rsm(struct tcp_rack *rack)
4987 {
4988         struct rack_sendmap *rsm;
4989
4990         /*
4991          * Walk the time-order transmitted list looking for an rsm that is
4992          * not acked. This will be the one that was sent the longest time
4993          * ago that is still outstanding.
4994          */
4995         TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4996                 if (rsm->r_flags & RACK_ACKED) {
4997                         continue;
4998                 }
4999                 goto finish;
5000         }
5001 finish:
5002         return (rsm);
5003 }
5004
5005 static struct rack_sendmap *
5006 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5007 {
5008         struct rack_sendmap *prsm;
5009
5010         /*
5011          * Walk the sequence order list backward until we hit and arrive at
5012          * the highest seq not acked. In theory when this is called it
5013          * should be the last segment (which it was not).
5014          */
5015         counter_u64_add(rack_find_high, 1);
5016         prsm = rsm;
5017         RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
5018                 if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5019                         continue;
5020                 }
5021                 return (prsm);
5022         }
5023         return (NULL);
5024 }
5025
5026 static uint32_t
5027 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
5028 {
5029         int32_t lro;
5030         uint32_t thresh;
5031
5032         /*
5033          * lro is the flag we use to determine if we have seen reordering.
5034          * If it gets set we have seen reordering. The reorder logic either
5035          * works in one of two ways:
5036          *
5037          * If reorder-fade is configured, then we track the last time we saw
5038          * re-ordering occur. If we reach the point where enough time as
5039          * passed we no longer consider reordering has occuring.
5040          *
5041          * Or if reorder-face is 0, then once we see reordering we consider
5042          * the connection to alway be subject to reordering and just set lro
5043          * to 1.
5044          *
5045          * In the end if lro is non-zero we add the extra time for
5046          * reordering in.
5047          */
5048         if (srtt == 0)
5049                 srtt = 1;
5050         if (rack->r_ctl.rc_reorder_ts) {
5051                 if (rack->r_ctl.rc_reorder_fade) {
5052                         if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5053                                 lro = cts - rack->r_ctl.rc_reorder_ts;
5054                                 if (lro == 0) {
5055                                         /*
5056                                          * No time as passed since the last
5057                                          * reorder, mark it as reordering.
5058                                          */
5059                                         lro = 1;
5060                                 }
5061                         } else {
5062                                 /* Negative time? */
5063                                 lro = 0;
5064                         }
5065                         if (lro > rack->r_ctl.rc_reorder_fade) {
5066                                 /* Turn off reordering seen too */
5067                                 rack->r_ctl.rc_reorder_ts = 0;
5068                                 lro = 0;
5069                         }
5070                 } else {
5071                         /* Reodering does not fade */
5072                         lro = 1;
5073                 }
5074         } else {
5075                 lro = 0;
5076         }
5077         thresh = srtt + rack->r_ctl.rc_pkt_delay;
5078         if (lro) {
5079                 /* It must be set, if not you get 1/4 rtt */
5080                 if (rack->r_ctl.rc_reorder_shift)
5081                         thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5082                 else
5083                         thresh += (srtt >> 2);
5084         } else {
5085                 thresh += 1;
5086         }
5087         /* We don't let the rack timeout be above a RTO */
5088         if (thresh > rack->rc_tp->t_rxtcur) {
5089                 thresh = rack->rc_tp->t_rxtcur;
5090         }
5091         /* And we don't want it above the RTO max either */
5092         if (thresh > rack_rto_max) {
5093                 thresh = rack_rto_max;
5094         }
5095         return (thresh);
5096 }
5097
5098 static uint32_t
5099 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5100                      struct rack_sendmap *rsm, uint32_t srtt)
5101 {
5102         struct rack_sendmap *prsm;
5103         uint32_t thresh, len;
5104         int segsiz;
5105
5106         if (srtt == 0)
5107                 srtt = 1;
5108         if (rack->r_ctl.rc_tlp_threshold)
5109                 thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5110         else
5111                 thresh = (srtt * 2);
5112
5113         /* Get the previous sent packet, if any */
5114         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5115         counter_u64_add(rack_enter_tlp_calc, 1);
5116         len = rsm->r_end - rsm->r_start;
5117         if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5118                 /* Exactly like the ID */
5119                 if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5120                         uint32_t alt_thresh;
5121                         /*
5122                          * Compensate for delayed-ack with the d-ack time.
5123                          */
5124                         counter_u64_add(rack_used_tlpmethod, 1);
5125                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5126                         if (alt_thresh > thresh)
5127                                 thresh = alt_thresh;
5128                 }
5129         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5130                 /* 2.1 behavior */
5131                 prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5132                 if (prsm && (len <= segsiz)) {
5133                         /*
5134                          * Two packets outstanding, thresh should be (2*srtt) +
5135                          * possible inter-packet delay (if any).
5136                          */
5137                         uint32_t inter_gap = 0;
5138                         int idx, nidx;
5139
5140                         counter_u64_add(rack_used_tlpmethod, 1);
5141                         idx = rsm->r_rtr_cnt - 1;
5142                         nidx = prsm->r_rtr_cnt - 1;
5143                         if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5144                                 /* Yes it was sent later (or at the same time) */
5145                                 inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5146                         }
5147                         thresh += inter_gap;
5148                 } else if (len <= segsiz) {
5149                         /*
5150                          * Possibly compensate for delayed-ack.
5151                          */
5152                         uint32_t alt_thresh;
5153
5154                         counter_u64_add(rack_used_tlpmethod2, 1);
5155                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5156                         if (alt_thresh > thresh)
5157                                 thresh = alt_thresh;
5158                 }
5159         } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5160                 /* 2.2 behavior */
5161                 if (len <= segsiz) {
5162                         uint32_t alt_thresh;
5163                         /*
5164                          * Compensate for delayed-ack with the d-ack time.
5165                          */
5166                         counter_u64_add(rack_used_tlpmethod, 1);
5167                         alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5168                         if (alt_thresh > thresh)
5169                                 thresh = alt_thresh;
5170                 }
5171         }
5172         /* Not above an RTO */
5173         if (thresh > tp->t_rxtcur) {
5174                 thresh = tp->t_rxtcur;
5175         }
5176         /* Not above a RTO max */
5177         if (thresh > rack_rto_max) {
5178                 thresh = rack_rto_max;
5179         }
5180         /* Apply user supplied min TLP */
5181         if (thresh < rack_tlp_min) {
5182                 thresh = rack_tlp_min;
5183         }
5184         return (thresh);
5185 }
5186
5187 static uint32_t
5188 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5189 {
5190         /*
5191          * We want the rack_rtt which is the
5192          * last rtt we measured. However if that
5193          * does not exist we fallback to the srtt (which
5194          * we probably will never do) and then as a last
5195          * resort we use RACK_INITIAL_RTO if no srtt is
5196          * yet set.
5197          */
5198         if (rack->rc_rack_rtt)
5199                 return (rack->rc_rack_rtt);
5200         else if (tp->t_srtt == 0)
5201                 return (RACK_INITIAL_RTO);
5202         return (tp->t_srtt);
5203 }
5204
5205 static struct rack_sendmap *
5206 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5207 {
5208         /*
5209          * Check to see that we don't need to fall into recovery. We will
5210          * need to do so if our oldest transmit is past the time we should
5211          * have had an ack.
5212          */
5213         struct tcp_rack *rack;
5214         struct rack_sendmap *rsm;
5215         int32_t idx;
5216         uint32_t srtt, thresh;
5217
5218         rack = (struct tcp_rack *)tp->t_fb_ptr;
5219         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5220                 return (NULL);
5221         }
5222         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5223         if (rsm == NULL)
5224                 return (NULL);
5225
5226         if (rsm->r_flags & RACK_ACKED) {
5227                 rsm = rack_find_lowest_rsm(rack);
5228                 if (rsm == NULL)
5229                         return (NULL);
5230         }
5231         idx = rsm->r_rtr_cnt - 1;
5232         srtt = rack_grab_rtt(tp, rack);
5233         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5234         if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5235                 return (NULL);
5236         }
5237         if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5238                 return (NULL);
5239         }
5240         /* Ok if we reach here we are over-due and this guy can be sent */
5241         if (IN_RECOVERY(tp->t_flags) == 0) {
5242                 /*
5243                  * For the one that enters us into recovery record undo
5244                  * info.
5245                  */
5246                 rack->r_ctl.rc_rsm_start = rsm->r_start;
5247                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
5248                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
5249         }
5250         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
5251         return (rsm);
5252 }
5253
5254 static uint32_t
5255 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5256 {
5257         int32_t t;
5258         int32_t tt;
5259         uint32_t ret_val;
5260
5261         t = (tp->t_srtt + (tp->t_rttvar << 2));
5262         RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5263             rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5264         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
5265                 tp->t_rxtshift++;
5266         rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5267         ret_val = (uint32_t)tt;
5268         return (ret_val);
5269 }
5270
5271 static uint32_t
5272 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5273 {
5274         /*
5275          * Start the FR timer, we do this based on getting the first one in
5276          * the rc_tmap. Note that if its NULL we must stop the timer. in all
5277          * events we need to stop the running timer (if its running) before
5278          * starting the new one.
5279          */
5280         uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5281         uint32_t srtt_cur;
5282         int32_t idx;
5283         int32_t is_tlp_timer = 0;
5284         struct rack_sendmap *rsm;
5285
5286         if (rack->t_timers_stopped) {
5287                 /* All timers have been stopped none are to run */
5288                 return (0);
5289         }
5290         if (rack->rc_in_persist) {
5291                 /* We can't start any timer in persists */
5292                 return (rack_get_persists_timer_val(tp, rack));
5293         }
5294         rack->rc_on_min_to = 0;
5295         if ((tp->t_state < TCPS_ESTABLISHED) ||
5296             ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5297                 goto activate_rxt;
5298         }
5299         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5300         if ((rsm == NULL) || sup_rack) {
5301                 /* Nothing on the send map or no rack */
5302 activate_rxt:
5303                 time_since_sent = 0;
5304                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5305                 if (rsm) {
5306                         /*
5307                          * Should we discount the RTX timer any?
5308                          *
5309                          * We want to discount it the smallest amount.
5310                          * If a timer (Rack/TLP or RXT) has gone off more
5311                          * recently thats the discount we want to use (now - timer time).
5312                          * If the retransmit of the oldest packet was more recent then
5313                          * we want to use that (now - oldest-packet-last_transmit_time).
5314                          *
5315                          */
5316                         idx = rsm->r_rtr_cnt - 1;
5317                         if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5318                                 tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5319                         else
5320                                 tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5321                         if (TSTMP_GT(cts, tstmp_touse))
5322                             time_since_sent = cts - tstmp_touse;
5323                 }
5324                 if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5325                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5326                         to = tp->t_rxtcur;
5327                         if (to > time_since_sent)
5328                                 to -= time_since_sent;
5329                         else
5330                                 to = rack->r_ctl.rc_min_to;
5331                         if (to == 0)
5332                                 to = 1;
5333                         /* Special case for KEEPINIT */
5334                         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5335                             (TP_KEEPINIT(tp) != 0) &&
5336                             rsm) {
5337                                 /*
5338                                  * We have to put a ceiling on the rxt timer
5339                                  * of the keep-init timeout.
5340                                  */
5341                                 uint32_t max_time, red;
5342
5343                                 max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5344                                 if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5345                                         red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5346                                         if (red < max_time)
5347                                                 max_time -= red;
5348                                         else
5349                                                 max_time = 1;
5350                                 }
5351                                 /* Reduce timeout to the keep value if needed */
5352                                 if (max_time < to)
5353                                         to = max_time;
5354                         }
5355                         return (to);
5356                 }
5357                 return (0);
5358         }
5359         if (rsm->r_flags & RACK_ACKED) {
5360                 rsm = rack_find_lowest_rsm(rack);
5361                 if (rsm == NULL) {
5362                         /* No lowest? */
5363                         goto activate_rxt;
5364                 }
5365         }
5366         if (rack->sack_attack_disable) {
5367                 /*
5368                  * We don't want to do
5369                  * any TLP's if you are an attacker.
5370                  * Though if you are doing what
5371                  * is expected you may still have
5372                  * SACK-PASSED marks.
5373                  */
5374                 goto activate_rxt;
5375         }
5376         /* Convert from ms to usecs */
5377         if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5378                 if ((tp->t_flags & TF_SENTFIN) &&
5379                     ((tp->snd_max - tp->snd_una) == 1) &&
5380                     (rsm->r_flags & RACK_HAS_FIN)) {
5381                         /*
5382                          * We don't start a rack timer if all we have is a
5383                          * FIN outstanding.
5384                          */
5385                         goto activate_rxt;
5386                 }
5387                 if ((rack->use_rack_rr == 0) &&
5388                     (IN_FASTRECOVERY(tp->t_flags)) &&
5389                     (rack->rack_no_prr == 0) &&
5390                      (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5391                         /*
5392                          * We are not cheating, in recovery  and
5393                          * not enough ack's to yet get our next
5394                          * retransmission out.
5395                          *
5396                          * Note that classified attackers do not
5397                          * get to use the rack-cheat.
5398                          */
5399                         goto activate_tlp;
5400                 }
5401                 srtt = rack_grab_rtt(tp, rack);
5402                 thresh = rack_calc_thresh_rack(rack, srtt, cts);
5403                 idx = rsm->r_rtr_cnt - 1;
5404                 exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5405                 if (SEQ_GEQ(exp, cts)) {
5406                         to = exp - cts;
5407                         if (to < rack->r_ctl.rc_min_to) {
5408                                 to = rack->r_ctl.rc_min_to;
5409                                 if (rack->r_rr_config == 3)
5410                                         rack->rc_on_min_to = 1;
5411                         }
5412                 } else {
5413                         to = rack->r_ctl.rc_min_to;
5414                         if (rack->r_rr_config == 3)
5415                                 rack->rc_on_min_to = 1;
5416                 }
5417         } else {
5418                 /* Ok we need to do a TLP not RACK */
5419 activate_tlp:
5420                 if ((rack->rc_tlp_in_progress != 0) &&
5421                     (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5422                         /*
5423                          * The previous send was a TLP and we have sent
5424                          * N TLP's without sending new data.
5425                          */
5426                         goto activate_rxt;
5427                 }
5428                 rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5429                 if (rsm == NULL) {
5430                         /* We found no rsm to TLP with. */
5431                         goto activate_rxt;
5432                 }
5433                 if (rsm->r_flags & RACK_HAS_FIN) {
5434                         /* If its a FIN we dont do TLP */
5435                         rsm = NULL;
5436                         goto activate_rxt;
5437                 }
5438                 idx = rsm->r_rtr_cnt - 1;
5439                 time_since_sent = 0;
5440                 if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5441                         tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5442                 else
5443                         tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5444                 if (TSTMP_GT(cts, tstmp_touse))
5445                     time_since_sent = cts - tstmp_touse;
5446                 is_tlp_timer = 1;
5447                 if (tp->t_srtt) {
5448                         if ((rack->rc_srtt_measure_made == 0) &&
5449                             (tp->t_srtt == 1)) {
5450                                 /*
5451                                  * If another stack as run and set srtt to 1,
5452                                  * then the srtt was 0, so lets use the initial.
5453                                  */
5454                                 srtt = RACK_INITIAL_RTO;
5455                         } else {
5456                                 srtt_cur = tp->t_srtt;
5457                                 srtt = srtt_cur;
5458                         }
5459                 } else
5460                         srtt = RACK_INITIAL_RTO;
5461                 /*
5462                  * If the SRTT is not keeping up and the
5463                  * rack RTT has spiked we want to use
5464                  * the last RTT not the smoothed one.
5465                  */
5466                 if (rack_tlp_use_greater &&
5467                     tp->t_srtt &&
5468                     (srtt < rack_grab_rtt(tp, rack))) {
5469                         srtt = rack_grab_rtt(tp, rack);
5470                 }
5471                 thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5472                 if (thresh > time_since_sent) {
5473                         to = thresh - time_since_sent;
5474                 } else {
5475                         to = rack->r_ctl.rc_min_to;
5476                         rack_log_alt_to_to_cancel(rack,
5477                                                   thresh,               /* flex1 */
5478                                                   time_since_sent,      /* flex2 */
5479                                                   tstmp_touse,          /* flex3 */
5480                                                   rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5481                                                   (uint32_t)rsm->r_tim_lastsent[idx],
5482                                                   srtt,
5483                                                   idx, 99);
5484                 }
5485                 if (to < rack_tlp_min) {
5486                         to = rack_tlp_min;
5487                 }
5488                 if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5489                         /*
5490                          * If the TLP time works out to larger than the max
5491                          * RTO lets not do TLP.. just RTO.
5492                          */
5493                         goto activate_rxt;
5494                 }
5495         }
5496         if (is_tlp_timer == 0) {
5497                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5498         } else {
5499                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5500         }
5501         if (to == 0)
5502                 to = 1;
5503         return (to);
5504 }
5505
5506 static void
5507 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5508 {
5509         if (rack->rc_in_persist == 0) {
5510                 if (tp->t_flags & TF_GPUTINPROG) {
5511                         /*
5512                          * Stop the goodput now, the calling of the
5513                          * measurement function clears the flag.
5514                          */
5515                         rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__);
5516                 }
5517 #ifdef NETFLIX_SHARED_CWND
5518                 if (rack->r_ctl.rc_scw) {
5519                         tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5520                         rack->rack_scwnd_is_idle = 1;
5521                 }
5522 #endif
5523                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5524                 if (rack->r_ctl.rc_went_idle_time == 0)
5525                         rack->r_ctl.rc_went_idle_time = 1;
5526                 rack_timer_cancel(tp, rack, cts, __LINE__);
5527                 tp->t_rxtshift = 0;
5528                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5529                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5530                 rack->rc_in_persist = 1;
5531         }
5532 }
5533
5534 static void
5535 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5536 {
5537         if (rack->rc_inp->inp_in_hpts) {
5538                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5539                 rack->r_ctl.rc_hpts_flags = 0;
5540         }
5541 #ifdef NETFLIX_SHARED_CWND
5542         if (rack->r_ctl.rc_scw) {
5543                 tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5544                 rack->rack_scwnd_is_idle = 0;
5545         }
5546 #endif
5547         if (rack->rc_gp_dyn_mul &&
5548             (rack->use_fixed_rate == 0) &&
5549             (rack->rc_always_pace)) {
5550                 /*
5551                  * Do we count this as if a probe-rtt just
5552                  * finished?
5553                  */
5554                 uint32_t time_idle, idle_min;
5555
5556                 time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5557                 idle_min = rack_min_probertt_hold;
5558                 if (rack_probertt_gpsrtt_cnt_div) {
5559                         uint64_t extra;
5560                         extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5561                                 (uint64_t)rack_probertt_gpsrtt_cnt_mul;
5562                         extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5563                         idle_min += (uint32_t)extra;
5564                 }
5565                 if (time_idle >= idle_min) {
5566                         /* Yes, we count it as a probe-rtt. */
5567                         uint32_t us_cts;
5568
5569                         us_cts = tcp_get_usecs(NULL);
5570                         if (rack->in_probe_rtt == 0) {
5571                                 rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5572                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5573                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5574                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5575                         } else {
5576                                 rack_exit_probertt(rack, us_cts);
5577                         }
5578                 }
5579         }
5580         rack->rc_in_persist = 0;
5581         rack->r_ctl.rc_went_idle_time = 0;
5582         tp->t_rxtshift = 0;
5583         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5584            rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5585         rack->r_ctl.rc_agg_delayed = 0;
5586         rack->r_early = 0;
5587         rack->r_late = 0;
5588         rack->r_ctl.rc_agg_early = 0;
5589 }
5590
5591 static void
5592 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5593                    struct hpts_diag *diag, struct timeval *tv)
5594 {
5595         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5596                 union tcp_log_stackspecific log;
5597
5598                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5599                 log.u_bbr.flex1 = diag->p_nxt_slot;
5600                 log.u_bbr.flex2 = diag->p_cur_slot;
5601                 log.u_bbr.flex3 = diag->slot_req;
5602                 log.u_bbr.flex4 = diag->inp_hptsslot;
5603                 log.u_bbr.flex5 = diag->slot_remaining;
5604                 log.u_bbr.flex6 = diag->need_new_to;
5605                 log.u_bbr.flex7 = diag->p_hpts_active;
5606                 log.u_bbr.flex8 = diag->p_on_min_sleep;
5607                 /* Hijack other fields as needed */
5608                 log.u_bbr.epoch = diag->have_slept;
5609                 log.u_bbr.lt_epoch = diag->yet_to_sleep;
5610                 log.u_bbr.pkts_out = diag->co_ret;
5611                 log.u_bbr.applimited = diag->hpts_sleep_time;
5612                 log.u_bbr.delivered = diag->p_prev_slot;
5613                 log.u_bbr.inflight = diag->p_runningtick;
5614                 log.u_bbr.bw_inuse = diag->wheel_tick;
5615                 log.u_bbr.rttProp = diag->wheel_cts;
5616                 log.u_bbr.timeStamp = cts;
5617                 log.u_bbr.delRate = diag->maxticks;
5618                 log.u_bbr.cur_del_rate = diag->p_curtick;
5619                 log.u_bbr.cur_del_rate <<= 32;
5620                 log.u_bbr.cur_del_rate |= diag->p_lasttick;
5621                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5622                     &rack->rc_inp->inp_socket->so_rcv,
5623                     &rack->rc_inp->inp_socket->so_snd,
5624                     BBR_LOG_HPTSDIAG, 0,
5625                     0, &log, false, tv);
5626         }
5627
5628 }
5629
5630 static void
5631 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5632 {
5633         if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5634                 union tcp_log_stackspecific log;
5635                 struct timeval tv;
5636
5637                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5638                 log.u_bbr.flex1 = sb->sb_flags;
5639                 log.u_bbr.flex2 = len;
5640                 log.u_bbr.flex3 = sb->sb_state;
5641                 log.u_bbr.flex8 = type;
5642                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5643                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
5644                     &rack->rc_inp->inp_socket->so_rcv,
5645                     &rack->rc_inp->inp_socket->so_snd,
5646                     TCP_LOG_SB_WAKE, 0,
5647                     len, &log, false, &tv);
5648         }
5649 }
5650
5651 static void
5652 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5653       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5654 {
5655         struct hpts_diag diag;
5656         struct inpcb *inp;
5657         struct timeval tv;
5658         uint32_t delayed_ack = 0;
5659         uint32_t hpts_timeout;
5660         uint32_t entry_slot = slot;
5661         uint8_t stopped;
5662         uint32_t left = 0;
5663         uint32_t us_cts;
5664
5665         inp = tp->t_inpcb;
5666         if ((tp->t_state == TCPS_CLOSED) ||
5667             (tp->t_state == TCPS_LISTEN)) {
5668                 return;
5669         }
5670         if (inp->inp_in_hpts) {
5671                 /* Already on the pacer */
5672                 return;
5673         }
5674         stopped = rack->rc_tmr_stopped;
5675         if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5676                 left = rack->r_ctl.rc_timer_exp - cts;
5677         }
5678         rack->r_ctl.rc_timer_exp = 0;
5679         rack->r_ctl.rc_hpts_flags = 0;
5680         us_cts = tcp_get_usecs(&tv);
5681         /* Now early/late accounting */
5682         rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL);
5683         if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5684                 /*
5685                  * We have a early carry over set,
5686                  * we can always add more time so we
5687                  * can always make this compensation.
5688                  *
5689                  * Note if ack's are allowed to wake us do not
5690                  * penalize the next timer for being awoke
5691                  * by an ack aka the rc_agg_early (non-paced mode).
5692                  */
5693                 slot += rack->r_ctl.rc_agg_early;
5694                 rack->r_early = 0;
5695                 rack->r_ctl.rc_agg_early = 0;
5696         }
5697         if (rack->r_late) {
5698                 /*
5699                  * This is harder, we can
5700                  * compensate some but it
5701                  * really depends on what
5702                  * the current pacing time is.
5703                  */
5704                 if (rack->r_ctl.rc_agg_delayed >= slot) {
5705                         /*
5706                          * We can't compensate for it all.
5707                          * And we have to have some time
5708                          * on the clock. We always have a min
5709                          * 10 slots (10 x 10 i.e. 100 usecs).
5710                          */
5711                         if (slot <= HPTS_TICKS_PER_USEC) {
5712                                 /* We gain delay */
5713                                 rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_USEC - slot);
5714                                 slot = HPTS_TICKS_PER_USEC;
5715                         } else {
5716                                 /* We take off some */
5717                                 rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_USEC);
5718                                 slot = HPTS_TICKS_PER_USEC;
5719                         }
5720                 } else {
5721                         slot -= rack->r_ctl.rc_agg_delayed;
5722                         rack->r_ctl.rc_agg_delayed = 0;
5723                         /* Make sure we have 100 useconds at minimum */
5724                         if (slot < HPTS_TICKS_PER_USEC) {
5725                                 rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_USEC - slot;
5726                                 slot = HPTS_TICKS_PER_USEC;
5727                         }
5728                         if (rack->r_ctl.rc_agg_delayed == 0)
5729                                 rack->r_late = 0;
5730                 }
5731         }
5732         if (slot) {
5733                 /* We are pacing too */
5734                 rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5735         }
5736         hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5737 #ifdef NETFLIX_EXP_DETECTION
5738         if (rack->sack_attack_disable &&
5739             (slot < tcp_sad_pacing_interval)) {
5740                 /*
5741                  * We have a potential attacker on
5742                  * the line. We have possibly some
5743                  * (or now) pacing time set. We want to
5744                  * slow down the processing of sacks by some
5745                  * amount (if it is an attacker). Set the default
5746                  * slot for attackers in place (unless the orginal
5747                  * interval is longer). Its stored in
5748                  * micro-seconds, so lets convert to msecs.
5749                  */
5750                 slot = tcp_sad_pacing_interval;
5751         }
5752 #endif
5753         if (tp->t_flags & TF_DELACK) {
5754                 delayed_ack = TICKS_2_USEC(tcp_delacktime);
5755                 rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5756         }
5757         if (delayed_ack && ((hpts_timeout == 0) ||
5758                             (delayed_ack < hpts_timeout)))
5759                 hpts_timeout = delayed_ack;
5760         else
5761                 rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5762         /*
5763          * If no timers are going to run and we will fall off the hptsi
5764          * wheel, we resort to a keep-alive timer if its configured.
5765          */
5766         if ((hpts_timeout == 0) &&
5767             (slot == 0)) {
5768                 if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5769                     (tp->t_state <= TCPS_CLOSING)) {
5770                         /*
5771                          * Ok we have no timer (persists, rack, tlp, rxt  or
5772                          * del-ack), we don't have segments being paced. So
5773                          * all that is left is the keepalive timer.
5774                          */
5775                         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5776                                 /* Get the established keep-alive time */
5777                                 hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5778                         } else {
5779                                 /*
5780                                  * Get the initial setup keep-alive time,
5781                                  * note that this is probably not going to
5782                                  * happen, since rack will be running a rxt timer
5783                                  * if a SYN of some sort is outstanding. It is
5784                                  * actually handled in rack_timeout_rxt().
5785                                  */
5786                                 hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5787                         }
5788                         rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5789                         if (rack->in_probe_rtt) {
5790                                 /*
5791                                  * We want to instead not wake up a long time from
5792                                  * now but to wake up about the time we would
5793                                  * exit probe-rtt and initiate a keep-alive ack.
5794                                  * This will get us out of probe-rtt and update
5795                                  * our min-rtt.
5796                                  */
5797                                 hpts_timeout = rack_min_probertt_hold;
5798                         }
5799                 }
5800         }
5801         if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5802             (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5803                 /*
5804                  * RACK, TLP, persists and RXT timers all are restartable
5805                  * based on actions input .. i.e we received a packet (ack
5806                  * or sack) and that changes things (rw, or snd_una etc).
5807                  * Thus we can restart them with a new value. For
5808                  * keep-alive, delayed_ack we keep track of what was left
5809                  * and restart the timer with a smaller value.
5810                  */
5811                 if (left < hpts_timeout)
5812                         hpts_timeout = left;
5813         }
5814         if (hpts_timeout) {
5815                 /*
5816                  * Hack alert for now we can't time-out over 2,147,483
5817                  * seconds (a bit more than 596 hours), which is probably ok
5818                  * :).
5819                  */
5820                 if (hpts_timeout > 0x7ffffffe)
5821                         hpts_timeout = 0x7ffffffe;
5822                 rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5823         }
5824         rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL);
5825         if ((rack->gp_ready == 0) &&
5826             (rack->use_fixed_rate == 0) &&
5827             (hpts_timeout < slot) &&
5828             (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5829                 /*
5830                  * We have no good estimate yet for the
5831                  * old clunky burst mitigation or the
5832                  * real pacing. And the tlp or rxt is smaller
5833                  * than the pacing calculation. Lets not
5834                  * pace that long since we know the calculation
5835                  * so far is not accurate.
5836                  */
5837                 slot = hpts_timeout;
5838         }
5839         rack->r_ctl.last_pacing_time = slot;
5840         /**
5841          * Turn off all the flags for queuing by default. The
5842          * flags have important meanings to what happens when
5843          * LRO interacts with the transport. Most likely (by default now)
5844          * mbuf_queueing and ack compression are on. So the transport
5845          * has a couple of flags that control what happens (if those
5846          * are not on then these flags won't have any effect since it
5847          * won't go through the queuing LRO path).
5848          *
5849          * INP_MBUF_QUEUE_READY - This flags says that I am busy
5850          *                        pacing output, so don't disturb. But
5851          *                        it also means LRO can wake me if there
5852          *                        is a SACK arrival.
5853          *
5854          * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5855          *                       with the above flag (QUEUE_READY) and
5856          *                       when present it says don't even wake me
5857          *                       if a SACK arrives.
5858          *
5859          * The idea behind these flags is that if we are pacing we
5860          * set the MBUF_QUEUE_READY and only get woken up if
5861          * a SACK arrives (which could change things) or if
5862          * our pacing timer expires. If, however, we have a rack
5863          * timer running, then we don't even want a sack to wake
5864          * us since the rack timer has to expire before we can send.
5865          *
5866          * Other cases should usually have none of the flags set
5867          * so LRO can call into us.
5868          */
5869         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5870         if (slot) {
5871                 rack->r_ctl.rc_last_output_to = us_cts + slot;
5872                 /*
5873                  * A pacing timer (slot) is being set, in
5874                  * such a case we cannot send (we are blocked by
5875                  * the timer). So lets tell LRO that it should not
5876                  * wake us unless there is a SACK. Note this only
5877                  * will be effective if mbuf queueing is on or
5878                  * compressed acks are being processed.
5879                  */
5880                 inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5881                 /*
5882                  * But wait if we have a Rack timer running
5883                  * even a SACK should not disturb us (with
5884                  * the exception of r_rr_config 3).
5885                  */
5886                 if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5887                     (rack->r_rr_config != 3))
5888                         inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5889                 if (rack->rc_ack_can_sendout_data) {
5890                         /*
5891                          * Ahh but wait, this is that special case
5892                          * where the pacing timer can be disturbed
5893                          * backout the changes (used for non-paced
5894                          * burst limiting).
5895                          */
5896                         inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5897                 }
5898                 if ((rack->use_rack_rr) &&
5899                     (rack->r_rr_config < 2) &&
5900                     ((hpts_timeout) && (hpts_timeout < slot))) {
5901                         /*
5902                          * Arrange for the hpts to kick back in after the
5903                          * t-o if the t-o does not cause a send.
5904                          */
5905                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5906                                                    __LINE__, &diag);
5907                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5908                         rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5909                 } else {
5910                         (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
5911                                                    __LINE__, &diag);
5912                         rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5913                         rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5914                 }
5915         } else if (hpts_timeout) {
5916                 /*
5917                  * With respect to inp_flags2 here, lets let any new acks wake
5918                  * us up here. Since we are not pacing (no pacing timer), output
5919                  * can happen so we should let it. If its a Rack timer, then any inbound
5920                  * packet probably won't change the sending (we will be blocked)
5921                  * but it may change the prr stats so letting it in (the set defaults
5922                  * at the start of this block) are good enough.
5923                  */
5924                 (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
5925                                            __LINE__, &diag);
5926                 rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5927                 rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5928         } else {
5929                 /* No timer starting */
5930 #ifdef INVARIANTS
5931                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5932                         panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5933                             tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5934                 }
5935 #endif
5936         }
5937         rack->rc_tmr_stopped = 0;
5938         if (slot)
5939                 rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5940 }
5941
5942 /*
5943  * RACK Timer, here we simply do logging and house keeping.
5944  * the normal rack_output() function will call the
5945  * appropriate thing to check if we need to do a RACK retransmit.
5946  * We return 1, saying don't proceed with rack_output only
5947  * when all timers have been stopped (destroyed PCB?).
5948  */
5949 static int
5950 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5951 {
5952         /*
5953          * This timer simply provides an internal trigger to send out data.
5954          * The check_recovery_mode call will see if there are needed
5955          * retransmissions, if so we will enter fast-recovery. The output
5956          * call may or may not do the same thing depending on sysctl
5957          * settings.
5958          */
5959         struct rack_sendmap *rsm;
5960
5961         if (tp->t_timers->tt_flags & TT_STOPPED) {
5962                 return (1);
5963         }
5964         counter_u64_add(rack_to_tot, 1);
5965         if (rack->r_state && (rack->r_state != tp->t_state))
5966                 rack_set_state(tp, rack);
5967         rack->rc_on_min_to = 0;
5968         rsm = rack_check_recovery_mode(tp, cts);
5969         rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5970         if (rsm) {
5971                 rack->r_ctl.rc_resend = rsm;
5972                 rack->r_timer_override = 1;
5973                 if (rack->use_rack_rr) {
5974                         /*
5975                          * Don't accumulate extra pacing delay
5976                          * we are allowing the rack timer to
5977                          * over-ride pacing i.e. rrr takes precedence
5978                          * if the pacing interval is longer than the rrr
5979                          * time (in other words we get the min pacing
5980                          * time versus rrr pacing time).
5981                          */
5982                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5983                 }
5984         }
5985         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5986         if (rsm == NULL) {
5987                 /* restart a timer and return 1 */
5988                 rack_start_hpts_timer(rack, tp, cts,
5989                                       0, 0, 0);
5990                 return (1);
5991         }
5992         return (0);
5993 }
5994
5995 static void
5996 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5997 {
5998         if (rsm->m->m_len > rsm->orig_m_len) {
5999                 /*
6000                  * Mbuf grew, caused by sbcompress, our offset does
6001                  * not change.
6002                  */
6003                 rsm->orig_m_len = rsm->m->m_len;
6004         } else if (rsm->m->m_len < rsm->orig_m_len) {
6005                 /*
6006                  * Mbuf shrank, trimmed off the top by an ack, our
6007                  * offset changes.
6008                  */
6009                 rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6010                 rsm->orig_m_len = rsm->m->m_len;
6011         }
6012 }
6013
6014 static void
6015 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6016 {
6017         struct mbuf *m;
6018         uint32_t soff;
6019
6020         if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
6021                 /* Fix up the orig_m_len and possibly the mbuf offset */
6022                 rack_adjust_orig_mlen(src_rsm);
6023         }
6024         m = src_rsm->m;
6025         soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6026         while (soff >= m->m_len) {
6027                 /* Move out past this mbuf */
6028                 soff -= m->m_len;
6029                 m = m->m_next;
6030                 KASSERT((m != NULL),
6031                         ("rsm:%p nrsm:%p hit at soff:%u null m",
6032                          src_rsm, rsm, soff));
6033         }
6034         rsm->m = m;
6035         rsm->soff = soff;
6036         rsm->orig_m_len = m->m_len;
6037 }
6038
6039 static __inline void
6040 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
6041                struct rack_sendmap *rsm, uint32_t start)
6042 {
6043         int idx;
6044
6045         nrsm->r_start = start;
6046         nrsm->r_end = rsm->r_end;
6047         nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6048         nrsm->r_flags = rsm->r_flags;
6049         nrsm->r_dupack = rsm->r_dupack;
6050         nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6051         nrsm->r_rtr_bytes = 0;
6052         rsm->r_end = nrsm->r_start;
6053         nrsm->r_just_ret = rsm->r_just_ret;
6054         for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6055                 nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6056         }
6057         /* Now if we have SYN flag we keep it on the left edge */
6058         if (nrsm->r_flags & RACK_HAS_SYN)
6059                 nrsm->r_flags &= ~RACK_HAS_SYN;
6060         /* Now if we have a FIN flag we keep it on the right edge */
6061         if (rsm->r_flags & RACK_HAS_FIN)
6062                 rsm->r_flags &= ~RACK_HAS_FIN;
6063         /* Push bit must go to the right edge as well */
6064         if (rsm->r_flags & RACK_HAD_PUSH)
6065                 rsm->r_flags &= ~RACK_HAD_PUSH;
6066
6067         /*
6068          * Now we need to find nrsm's new location in the mbuf chain
6069          * we basically calculate a new offset, which is soff +
6070          * how much is left in original rsm. Then we walk out the mbuf
6071          * chain to find the righ postion, it may be the same mbuf
6072          * or maybe not.
6073          */
6074         KASSERT(((rsm->m != NULL) ||
6075                  (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6076                 ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6077         if (rsm->m)
6078                 rack_setup_offset_for_rsm(rsm, nrsm);
6079 }
6080
6081 static struct rack_sendmap *
6082 rack_merge_rsm(struct tcp_rack *rack,
6083                struct rack_sendmap *l_rsm,
6084                struct rack_sendmap *r_rsm)
6085 {
6086         /*
6087          * We are merging two ack'd RSM's,
6088          * the l_rsm is on the left (lower seq
6089          * values) and the r_rsm is on the right
6090          * (higher seq value). The simplest way
6091          * to merge these is to move the right
6092          * one into the left. I don't think there
6093          * is any reason we need to try to find
6094          * the oldest (or last oldest retransmitted).
6095          */
6096         struct rack_sendmap *rm;
6097
6098         rack_log_map_chg(rack->rc_tp, rack, NULL,
6099                          l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6100         l_rsm->r_end = r_rsm->r_end;
6101         if (l_rsm->r_dupack < r_rsm->r_dupack)
6102                 l_rsm->r_dupack = r_rsm->r_dupack;
6103         if (r_rsm->r_rtr_bytes)
6104                 l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6105         if (r_rsm->r_in_tmap) {
6106                 /* This really should not happen */
6107                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6108                 r_rsm->r_in_tmap = 0;
6109         }
6110
6111         /* Now the flags */
6112         if (r_rsm->r_flags & RACK_HAS_FIN)
6113                 l_rsm->r_flags |= RACK_HAS_FIN;
6114         if (r_rsm->r_flags & RACK_TLP)
6115                 l_rsm->r_flags |= RACK_TLP;
6116         if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6117                 l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6118         if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6119             ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6120                 /*
6121                  * If both are app-limited then let the
6122                  * free lower the count. If right is app
6123                  * limited and left is not, transfer.
6124                  */
6125                 l_rsm->r_flags |= RACK_APP_LIMITED;
6126                 r_rsm->r_flags &= ~RACK_APP_LIMITED;
6127                 if (r_rsm == rack->r_ctl.rc_first_appl)
6128                         rack->r_ctl.rc_first_appl = l_rsm;
6129         }
6130         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6131 #ifdef INVARIANTS
6132         if (rm != r_rsm) {
6133                 panic("removing head in rack:%p rsm:%p rm:%p",
6134                       rack, r_rsm, rm);
6135         }
6136 #endif
6137         if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6138                 /* Transfer the split limit to the map we free */
6139                 r_rsm->r_limit_type = l_rsm->r_limit_type;
6140                 l_rsm->r_limit_type = 0;
6141         }
6142         rack_free(rack, r_rsm);
6143         return (l_rsm);
6144 }
6145
6146 /*
6147  * TLP Timer, here we simply setup what segment we want to
6148  * have the TLP expire on, the normal rack_output() will then
6149  * send it out.
6150  *
6151  * We return 1, saying don't proceed with rack_output only
6152  * when all timers have been stopped (destroyed PCB?).
6153  */
6154 static int
6155 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6156 {
6157         /*
6158          * Tail Loss Probe.
6159          */
6160         struct rack_sendmap *rsm = NULL;
6161         struct rack_sendmap *insret;
6162         struct socket *so;
6163         uint32_t amm;
6164         uint32_t out, avail;
6165         int collapsed_win = 0;
6166
6167         if (tp->t_timers->tt_flags & TT_STOPPED) {
6168                 return (1);
6169         }
6170         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6171                 /* Its not time yet */
6172                 return (0);
6173         }
6174         if (ctf_progress_timeout_check(tp, true)) {
6175                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6176                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6177                 return (1);
6178         }
6179         /*
6180          * A TLP timer has expired. We have been idle for 2 rtts. So we now
6181          * need to figure out how to force a full MSS segment out.
6182          */
6183         rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6184         rack->r_ctl.retran_during_recovery = 0;
6185         rack->r_ctl.dsack_byte_cnt = 0;
6186         counter_u64_add(rack_tlp_tot, 1);
6187         if (rack->r_state && (rack->r_state != tp->t_state))
6188                 rack_set_state(tp, rack);
6189         so = tp->t_inpcb->inp_socket;
6190         avail = sbavail(&so->so_snd);
6191         out = tp->snd_max - tp->snd_una;
6192         if (out > tp->snd_wnd) {
6193                 /* special case, we need a retransmission */
6194                 collapsed_win = 1;
6195                 goto need_retran;
6196         }
6197         /*
6198          * Check our send oldest always settings, and if
6199          * there is an oldest to send jump to the need_retran.
6200          */
6201         if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6202                 goto need_retran;
6203
6204         if (avail > out) {
6205                 /* New data is available */
6206                 amm = avail - out;
6207                 if (amm > ctf_fixed_maxseg(tp)) {
6208                         amm = ctf_fixed_maxseg(tp);
6209                         if ((amm + out) > tp->snd_wnd) {
6210                                 /* We are rwnd limited */
6211                                 goto need_retran;
6212                         }
6213                 } else if (amm < ctf_fixed_maxseg(tp)) {
6214                         /* not enough to fill a MTU */
6215                         goto need_retran;
6216                 }
6217                 if (IN_FASTRECOVERY(tp->t_flags)) {
6218                         /* Unlikely */
6219                         if (rack->rack_no_prr == 0) {
6220                                 if (out + amm <= tp->snd_wnd) {
6221                                         rack->r_ctl.rc_prr_sndcnt = amm;
6222                                         rack_log_to_prr(rack, 4, 0);
6223                                 }
6224                         } else
6225                                 goto need_retran;
6226                 } else {
6227                         /* Set the send-new override */
6228                         if (out + amm <= tp->snd_wnd)
6229                                 rack->r_ctl.rc_tlp_new_data = amm;
6230                         else
6231                                 goto need_retran;
6232                 }
6233                 rack->r_ctl.rc_tlpsend = NULL;
6234                 counter_u64_add(rack_tlp_newdata, 1);
6235                 goto send;
6236         }
6237 need_retran:
6238         /*
6239          * Ok we need to arrange the last un-acked segment to be re-sent, or
6240          * optionally the first un-acked segment.
6241          */
6242         if (collapsed_win == 0) {
6243                 if (rack_always_send_oldest)
6244                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6245                 else {
6246                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6247                         if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6248                                 rsm = rack_find_high_nonack(rack, rsm);
6249                         }
6250                 }
6251                 if (rsm == NULL) {
6252                         counter_u64_add(rack_tlp_does_nada, 1);
6253 #ifdef TCP_BLACKBOX
6254                         tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6255 #endif
6256                         goto out;
6257                 }
6258         } else {
6259                 /*
6260                  * We must find the last segment
6261                  * that was acceptable by the client.
6262                  */
6263                 RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6264                         if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6265                                 /* Found one */
6266                                 break;
6267                         }
6268                 }
6269                 if (rsm == NULL) {
6270                         /* None? if so send the first */
6271                         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6272                         if (rsm == NULL) {
6273                                 counter_u64_add(rack_tlp_does_nada, 1);
6274 #ifdef TCP_BLACKBOX
6275                                 tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6276 #endif
6277                                 goto out;
6278                         }
6279                 }
6280         }
6281         if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6282                 /*
6283                  * We need to split this the last segment in two.
6284                  */
6285                 struct rack_sendmap *nrsm;
6286
6287                 nrsm = rack_alloc_full_limit(rack);
6288                 if (nrsm == NULL) {
6289                         /*
6290                          * No memory to split, we will just exit and punt
6291                          * off to the RXT timer.
6292                          */
6293                         counter_u64_add(rack_tlp_does_nada, 1);
6294                         goto out;
6295                 }
6296                 rack_clone_rsm(rack, nrsm, rsm,
6297                                (rsm->r_end - ctf_fixed_maxseg(tp)));
6298                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6299                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6300 #ifdef INVARIANTS
6301                 if (insret != NULL) {
6302                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6303                               nrsm, insret, rack, rsm);
6304                 }
6305 #endif
6306                 if (rsm->r_in_tmap) {
6307                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6308                         nrsm->r_in_tmap = 1;
6309                 }
6310                 rsm->r_flags &= (~RACK_HAS_FIN);
6311                 rsm = nrsm;
6312         }
6313         rack->r_ctl.rc_tlpsend = rsm;
6314 send:
6315         rack->r_timer_override = 1;
6316         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6317         return (0);
6318 out:
6319         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6320         return (0);
6321 }
6322
6323 /*
6324  * Delayed ack Timer, here we simply need to setup the
6325  * ACK_NOW flag and remove the DELACK flag. From there
6326  * the output routine will send the ack out.
6327  *
6328  * We only return 1, saying don't proceed, if all timers
6329  * are stopped (destroyed PCB?).
6330  */
6331 static int
6332 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6333 {
6334         if (tp->t_timers->tt_flags & TT_STOPPED) {
6335                 return (1);
6336         }
6337         rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6338         tp->t_flags &= ~TF_DELACK;
6339         tp->t_flags |= TF_ACKNOW;
6340         KMOD_TCPSTAT_INC(tcps_delack);
6341         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6342         return (0);
6343 }
6344
6345 /*
6346  * Persists timer, here we simply send the
6347  * same thing as a keepalive will.
6348  * the one byte send.
6349  *
6350  * We only return 1, saying don't proceed, if all timers
6351  * are stopped (destroyed PCB?).
6352  */
6353 static int
6354 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6355 {
6356         struct tcptemp *t_template;
6357         struct inpcb *inp;
6358         int32_t retval = 1;
6359
6360         inp = tp->t_inpcb;
6361
6362         if (tp->t_timers->tt_flags & TT_STOPPED) {
6363                 return (1);
6364         }
6365         if (rack->rc_in_persist == 0)
6366                 return (0);
6367         if (ctf_progress_timeout_check(tp, false)) {
6368                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6369                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6370                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6371                 return (1);
6372         }
6373         KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6374         /*
6375          * Persistence timer into zero window. Force a byte to be output, if
6376          * possible.
6377          */
6378         KMOD_TCPSTAT_INC(tcps_persisttimeo);
6379         /*
6380          * Hack: if the peer is dead/unreachable, we do not time out if the
6381          * window is closed.  After a full backoff, drop the connection if
6382          * the idle time (no responses to probes) reaches the maximum
6383          * backoff that we would use if retransmitting.
6384          */
6385         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6386             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6387              TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6388                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6389                 retval = 1;
6390                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6391                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6392                 goto out;
6393         }
6394         if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6395             tp->snd_una == tp->snd_max)
6396                 rack_exit_persist(tp, rack, cts);
6397         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6398         /*
6399          * If the user has closed the socket then drop a persisting
6400          * connection after a much reduced timeout.
6401          */
6402         if (tp->t_state > TCPS_CLOSE_WAIT &&
6403             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6404                 retval = 1;
6405                 KMOD_TCPSTAT_INC(tcps_persistdrop);
6406                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6407                 tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6408                 goto out;
6409         }
6410         t_template = tcpip_maketemplate(rack->rc_inp);
6411         if (t_template) {
6412                 /* only set it if we were answered */
6413                 if (rack->forced_ack == 0) {
6414                         rack->forced_ack = 1;
6415                         rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6416                 }
6417                 tcp_respond(tp, t_template->tt_ipgen,
6418                             &t_template->tt_t, (struct mbuf *)NULL,
6419                             tp->rcv_nxt, tp->snd_una - 1, 0);
6420                 /* This sends an ack */
6421                 if (tp->t_flags & TF_DELACK)
6422                         tp->t_flags &= ~TF_DELACK;
6423                 free(t_template, M_TEMP);
6424         }
6425         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6426                 tp->t_rxtshift++;
6427 out:
6428         rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6429         rack_start_hpts_timer(rack, tp, cts,
6430                               0, 0, 0);
6431         return (retval);
6432 }
6433
6434 /*
6435  * If a keepalive goes off, we had no other timers
6436  * happening. We always return 1 here since this
6437  * routine either drops the connection or sends
6438  * out a segment with respond.
6439  */
6440 static int
6441 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6442 {
6443         struct tcptemp *t_template;
6444         struct inpcb *inp;
6445
6446         if (tp->t_timers->tt_flags & TT_STOPPED) {
6447                 return (1);
6448         }
6449         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6450         inp = tp->t_inpcb;
6451         rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6452         /*
6453          * Keep-alive timer went off; send something or drop connection if
6454          * idle for too long.
6455          */
6456         KMOD_TCPSTAT_INC(tcps_keeptimeo);
6457         if (tp->t_state < TCPS_ESTABLISHED)
6458                 goto dropit;
6459         if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6460             tp->t_state <= TCPS_CLOSING) {
6461                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6462                         goto dropit;
6463                 /*
6464                  * Send a packet designed to force a response if the peer is
6465                  * up and reachable: either an ACK if the connection is
6466                  * still alive, or an RST if the peer has closed the
6467                  * connection due to timeout or reboot. Using sequence
6468                  * number tp->snd_una-1 causes the transmitted zero-length
6469                  * segment to lie outside the receive window; by the
6470                  * protocol spec, this requires the correspondent TCP to
6471                  * respond.
6472                  */
6473                 KMOD_TCPSTAT_INC(tcps_keepprobe);
6474                 t_template = tcpip_maketemplate(inp);
6475                 if (t_template) {
6476                         if (rack->forced_ack == 0) {
6477                                 rack->forced_ack = 1;
6478                                 rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6479                         }
6480                         tcp_respond(tp, t_template->tt_ipgen,
6481                             &t_template->tt_t, (struct mbuf *)NULL,
6482                             tp->rcv_nxt, tp->snd_una - 1, 0);
6483                         free(t_template, M_TEMP);
6484                 }
6485         }
6486         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6487         return (1);
6488 dropit:
6489         KMOD_TCPSTAT_INC(tcps_keepdrops);
6490         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6491         tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6492         return (1);
6493 }
6494
6495 /*
6496  * Retransmit helper function, clear up all the ack
6497  * flags and take care of important book keeping.
6498  */
6499 static void
6500 rack_remxt_tmr(struct tcpcb *tp)
6501 {
6502         /*
6503          * The retransmit timer went off, all sack'd blocks must be
6504          * un-acked.
6505          */
6506         struct rack_sendmap *rsm, *trsm = NULL;
6507         struct tcp_rack *rack;
6508
6509         rack = (struct tcp_rack *)tp->t_fb_ptr;
6510         rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6511         rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6512         if (rack->r_state && (rack->r_state != tp->t_state))
6513                 rack_set_state(tp, rack);
6514         /*
6515          * Ideally we would like to be able to
6516          * mark SACK-PASS on anything not acked here.
6517          *
6518          * However, if we do that we would burst out
6519          * all that data 1ms apart. This would be unwise,
6520          * so for now we will just let the normal rxt timer
6521          * and tlp timer take care of it.
6522          *
6523          * Also we really need to stick them back in sequence
6524          * order. This way we send in the proper order and any
6525          * sacks that come floating in will "re-ack" the data.
6526          * To do this we zap the tmap with an INIT and then
6527          * walk through and place every rsm in the RB tree
6528          * back in its seq ordered place.
6529          */
6530         TAILQ_INIT(&rack->r_ctl.rc_tmap);
6531         RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6532                 rsm->r_dupack = 0;
6533                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6534                 /* We must re-add it back to the tlist */
6535                 if (trsm == NULL) {
6536                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6537                 } else {
6538                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6539                 }
6540                 rsm->r_in_tmap = 1;
6541                 trsm = rsm;
6542                 if (rsm->r_flags & RACK_ACKED)
6543                         rsm->r_flags |= RACK_WAS_ACKED;
6544                 rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6545         }
6546         /* Clear the count (we just un-acked them) */
6547         rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6548         rack->r_ctl.rc_sacked = 0;
6549         rack->r_ctl.rc_sacklast = NULL;
6550         rack->r_ctl.rc_agg_delayed = 0;
6551         rack->r_early = 0;
6552         rack->r_ctl.rc_agg_early = 0;
6553         rack->r_late = 0;
6554         /* Clear the tlp rtx mark */
6555         rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6556         if (rack->r_ctl.rc_resend != NULL)
6557                 rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6558         rack->r_ctl.rc_prr_sndcnt = 0;
6559         rack_log_to_prr(rack, 6, 0);
6560         rack->r_timer_override = 1;
6561         if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6562 #ifdef NETFLIX_EXP_DETECTION
6563             || (rack->sack_attack_disable != 0)
6564 #endif
6565                     ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6566                 /*
6567                  * For non-sack customers new data
6568                  * needs to go out as retransmits until
6569                  * we retransmit up to snd_max.
6570                  */
6571                 rack->r_must_retran = 1;
6572                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6573                                                 rack->r_ctl.rc_sacked);
6574         }
6575         rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6576 }
6577
6578 static void
6579 rack_convert_rtts(struct tcpcb *tp)
6580 {
6581         if (tp->t_srtt > 1) {
6582                 uint32_t val, frac;
6583
6584                 val = tp->t_srtt >> TCP_RTT_SHIFT;
6585                 frac = tp->t_srtt & 0x1f;
6586                 tp->t_srtt = TICKS_2_USEC(val);
6587                 /*
6588                  * frac is the fractional part of the srtt (if any)
6589                  * but its in ticks and every bit represents
6590                  * 1/32nd of a hz.
6591                  */
6592                 if (frac) {
6593                         if (hz == 1000) {
6594                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6595                         } else {
6596                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6597                         }
6598                         tp->t_srtt += frac;
6599                 }
6600         }
6601         if (tp->t_rttvar) {
6602                 uint32_t val, frac;
6603
6604                 val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6605                 frac = tp->t_rttvar & 0x1f;
6606                 tp->t_rttvar = TICKS_2_USEC(val);
6607                 /*
6608                  * frac is the fractional part of the srtt (if any)
6609                  * but its in ticks and every bit represents
6610                  * 1/32nd of a hz.
6611                  */
6612                 if (frac) {
6613                         if (hz == 1000) {
6614                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6615                         } else {
6616                                 frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6617                         }
6618                         tp->t_rttvar += frac;
6619                 }
6620         }
6621         tp->t_rxtcur = RACK_REXMTVAL(tp);
6622         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6623                 tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6624         }
6625         if (tp->t_rxtcur > rack_rto_max) {
6626                 tp->t_rxtcur = rack_rto_max;
6627         }
6628 }
6629
6630 static void
6631 rack_cc_conn_init(struct tcpcb *tp)
6632 {
6633         struct tcp_rack *rack;
6634         uint32_t srtt;
6635
6636         rack = (struct tcp_rack *)tp->t_fb_ptr;
6637         srtt = tp->t_srtt;
6638         cc_conn_init(tp);
6639         /*
6640          * Now convert to rack's internal format,
6641          * if required.
6642          */
6643         if ((srtt == 0) && (tp->t_srtt != 0))
6644                 rack_convert_rtts(tp);
6645         /*
6646          * We want a chance to stay in slowstart as
6647          * we create a connection. TCP spec says that
6648          * initially ssthresh is infinite. For our
6649          * purposes that is the snd_wnd.
6650          */
6651         if (tp->snd_ssthresh < tp->snd_wnd) {
6652                 tp->snd_ssthresh = tp->snd_wnd;
6653         }
6654         /*
6655          * We also want to assure a IW worth of
6656          * data can get inflight.
6657          */
6658         if (rc_init_window(rack) < tp->snd_cwnd)
6659                 tp->snd_cwnd = rc_init_window(rack);
6660 }
6661
6662 /*
6663  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6664  * we will setup to retransmit the lowest seq number outstanding.
6665  */
6666 static int
6667 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6668 {
6669         int32_t rexmt;
6670         struct inpcb *inp;
6671         int32_t retval = 0;
6672         bool isipv6;
6673
6674         inp = tp->t_inpcb;
6675         if (tp->t_timers->tt_flags & TT_STOPPED) {
6676                 return (1);
6677         }
6678         if (ctf_progress_timeout_check(tp, false)) {
6679                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6680                 rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6681                 tcp_set_inp_to_drop(inp, ETIMEDOUT);
6682                 return (1);
6683         }
6684         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6685         rack->r_ctl.retran_during_recovery = 0;
6686         rack->r_ctl.dsack_byte_cnt = 0;
6687         if (IN_FASTRECOVERY(tp->t_flags))
6688                 tp->t_flags |= TF_WASFRECOVERY;
6689         else
6690                 tp->t_flags &= ~TF_WASFRECOVERY;
6691         if (IN_CONGRECOVERY(tp->t_flags))
6692                 tp->t_flags |= TF_WASCRECOVERY;
6693         else
6694                 tp->t_flags &= ~TF_WASCRECOVERY;
6695         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6696             (tp->snd_una == tp->snd_max)) {
6697                 /* Nothing outstanding .. nothing to do */
6698                 return (0);
6699         }
6700         /*
6701          * Rack can only run one timer  at a time, so we cannot
6702          * run a KEEPINIT (gating SYN sending) and a retransmit
6703          * timer for the SYN. So if we are in a front state and
6704          * have a KEEPINIT timer we need to check the first transmit
6705          * against now to see if we have exceeded the KEEPINIT time
6706          * (if one is set).
6707          */
6708         if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6709             (TP_KEEPINIT(tp) != 0)) {
6710                 struct rack_sendmap *rsm;
6711
6712                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6713                 if (rsm) {
6714                         /* Ok we have something outstanding to test keepinit with */
6715                         if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6716                             ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6717                                 /* We have exceeded the KEEPINIT time */
6718                                 tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6719                                 goto drop_it;
6720                         }
6721                 }
6722         }
6723         /*
6724          * Retransmission timer went off.  Message has not been acked within
6725          * retransmit interval.  Back off to a longer retransmit interval
6726          * and retransmit one segment.
6727          */
6728         rack_remxt_tmr(tp);
6729         if ((rack->r_ctl.rc_resend == NULL) ||
6730             ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6731                 /*
6732                  * If the rwnd collapsed on
6733                  * the one we are retransmitting
6734                  * it does not count against the
6735                  * rxt count.
6736                  */
6737                 tp->t_rxtshift++;
6738         }
6739         if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6740                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6741 drop_it:
6742                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
6743                 KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6744                 retval = 1;
6745                 tcp_set_inp_to_drop(rack->rc_inp,
6746                     (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
6747                 goto out;
6748         }
6749         if (tp->t_state == TCPS_SYN_SENT) {
6750                 /*
6751                  * If the SYN was retransmitted, indicate CWND to be limited
6752                  * to 1 segment in cc_conn_init().
6753                  */
6754                 tp->snd_cwnd = 1;
6755         } else if (tp->t_rxtshift == 1) {
6756                 /*
6757                  * first retransmit; record ssthresh and cwnd so they can be
6758                  * recovered if this turns out to be a "bad" retransmit. A
6759                  * retransmit is considered "bad" if an ACK for this segment
6760                  * is received within RTT/2 interval; the assumption here is
6761                  * that the ACK was already in flight.  See "On Estimating
6762                  * End-to-End Network Path Properties" by Allman and Paxson
6763                  * for more details.
6764                  */
6765                 tp->snd_cwnd_prev = tp->snd_cwnd;
6766                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
6767                 tp->snd_recover_prev = tp->snd_recover;
6768                 tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6769                 tp->t_flags |= TF_PREVVALID;
6770         } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6771                 tp->t_flags &= ~TF_PREVVALID;
6772         KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6773         if ((tp->t_state == TCPS_SYN_SENT) ||
6774             (tp->t_state == TCPS_SYN_RECEIVED))
6775                 rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6776         else
6777                 rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6778
6779         RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6780            max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6781         /*
6782          * We enter the path for PLMTUD if connection is established or, if
6783          * connection is FIN_WAIT_1 status, reason for the last is that if
6784          * amount of data we send is very small, we could send it in couple
6785          * of packets and process straight to FIN. In that case we won't
6786          * catch ESTABLISHED state.
6787          */
6788 #ifdef INET6
6789         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6790 #else
6791         isipv6 = false;
6792 #endif
6793         if (((V_tcp_pmtud_blackhole_detect == 1) ||
6794             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6795             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6796             ((tp->t_state == TCPS_ESTABLISHED) ||
6797             (tp->t_state == TCPS_FIN_WAIT_1))) {
6798                 /*
6799                  * Idea here is that at each stage of mtu probe (usually,
6800                  * 1448 -> 1188 -> 524) should be given 2 chances to recover
6801                  * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6802                  * should take care of that.
6803                  */
6804                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6805                     (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6806                     (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6807                     tp->t_rxtshift % 2 == 0)) {
6808                         /*
6809                          * Enter Path MTU Black-hole Detection mechanism: -
6810                          * Disable Path MTU Discovery (IP "DF" bit). -
6811                          * Reduce MTU to lower value than what we negotiated
6812                          * with peer.
6813                          */
6814                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6815                                 /* Record that we may have found a black hole. */
6816                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6817                                 /* Keep track of previous MSS. */
6818                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6819                         }
6820
6821                         /*
6822                          * Reduce the MSS to blackhole value or to the
6823                          * default in an attempt to retransmit.
6824                          */
6825 #ifdef INET6
6826                         if (isipv6 &&
6827                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6828                                 /* Use the sysctl tuneable blackhole MSS. */
6829                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6830                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6831                         } else if (isipv6) {
6832                                 /* Use the default MSS. */
6833                                 tp->t_maxseg = V_tcp_v6mssdflt;
6834                                 /*
6835                                  * Disable Path MTU Discovery when we switch
6836                                  * to minmss.
6837                                  */
6838                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6839                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6840                         }
6841 #endif
6842 #if defined(INET6) && defined(INET)
6843                         else
6844 #endif
6845 #ifdef INET
6846                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6847                                 /* Use the sysctl tuneable blackhole MSS. */
6848                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6849                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6850                         } else {
6851                                 /* Use the default MSS. */
6852                                 tp->t_maxseg = V_tcp_mssdflt;
6853                                 /*
6854                                  * Disable Path MTU Discovery when we switch
6855                                  * to minmss.
6856                                  */
6857                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6858                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6859                         }
6860 #endif
6861                 } else {
6862                         /*
6863                          * If further retransmissions are still unsuccessful
6864                          * with a lowered MTU, maybe this isn't a blackhole
6865                          * and we restore the previous MSS and blackhole
6866                          * detection flags. The limit '6' is determined by
6867                          * giving each probe stage (1448, 1188, 524) 2
6868                          * chances to recover.
6869                          */
6870                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6871                             (tp->t_rxtshift >= 6)) {
6872                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6873                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6874                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6875                                 KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6876                         }
6877                 }
6878         }
6879         /*
6880          * Disable RFC1323 and SACK if we haven't got any response to
6881          * our third SYN to work-around some broken terminal servers
6882          * (most of which have hopefully been retired) that have bad VJ
6883          * header compression code which trashes TCP segments containing
6884          * unknown-to-them TCP options.
6885          */
6886         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6887             (tp->t_rxtshift == 3))
6888                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6889         /*
6890          * If we backed off this far, our srtt estimate is probably bogus.
6891          * Clobber it so we'll take the next rtt measurement as our srtt;
6892          * move the current srtt into rttvar to keep the current retransmit
6893          * times until then.
6894          */
6895         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6896 #ifdef INET6
6897                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6898                         in6_losing(tp->t_inpcb);
6899                 else
6900 #endif
6901                         in_losing(tp->t_inpcb);
6902                 tp->t_rttvar += tp->t_srtt;
6903                 tp->t_srtt = 0;
6904         }
6905         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6906         tp->snd_recover = tp->snd_max;
6907         tp->t_flags |= TF_ACKNOW;
6908         tp->t_rtttime = 0;
6909         rack_cong_signal(tp, CC_RTO, tp->snd_una);
6910 out:
6911         return (retval);
6912 }
6913
6914 static int
6915 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling)
6916 {
6917         int32_t ret = 0;
6918         int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6919
6920         if (timers == 0) {
6921                 return (0);
6922         }
6923         if (tp->t_state == TCPS_LISTEN) {
6924                 /* no timers on listen sockets */
6925                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6926                         return (0);
6927                 return (1);
6928         }
6929         if ((timers & PACE_TMR_RACK) &&
6930             rack->rc_on_min_to) {
6931                 /*
6932                  * For the rack timer when we
6933                  * are on a min-timeout (which means rrr_conf = 3)
6934                  * we don't want to check the timer. It may
6935                  * be going off for a pace and thats ok we
6936                  * want to send the retransmit (if its ready).
6937                  *
6938                  * If its on a normal rack timer (non-min) then
6939                  * we will check if its expired.
6940                  */
6941                 goto skip_time_check;
6942         }
6943         if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6944                 uint32_t left;
6945
6946                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6947                         ret = -1;
6948                         rack_log_to_processing(rack, cts, ret, 0);
6949                         return (0);
6950                 }
6951                 if (hpts_calling == 0) {
6952                         /*
6953                          * A user send or queued mbuf (sack) has called us? We
6954                          * return 0 and let the pacing guards
6955                          * deal with it if they should or
6956                          * should not cause a send.
6957                          */
6958                         ret = -2;
6959                         rack_log_to_processing(rack, cts, ret, 0);
6960                         return (0);
6961                 }
6962                 /*
6963                  * Ok our timer went off early and we are not paced false
6964                  * alarm, go back to sleep.
6965                  */
6966                 ret = -3;
6967                 left = rack->r_ctl.rc_timer_exp - cts;
6968                 tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
6969                 rack_log_to_processing(rack, cts, ret, left);
6970                 return (1);
6971         }
6972 skip_time_check:
6973         rack->rc_tmr_stopped = 0;
6974         rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6975         if (timers & PACE_TMR_DELACK) {
6976                 ret = rack_timeout_delack(tp, rack, cts);
6977         } else if (timers & PACE_TMR_RACK) {
6978                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6979                 rack->r_fast_output = 0;
6980                 ret = rack_timeout_rack(tp, rack, cts);
6981         } else if (timers & PACE_TMR_TLP) {
6982                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6983                 ret = rack_timeout_tlp(tp, rack, cts);
6984         } else if (timers & PACE_TMR_RXT) {
6985                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
6986                 rack->r_fast_output = 0;
6987                 ret = rack_timeout_rxt(tp, rack, cts);
6988         } else if (timers & PACE_TMR_PERSIT) {
6989                 ret = rack_timeout_persist(tp, rack, cts);
6990         } else if (timers & PACE_TMR_KEEP) {
6991                 ret = rack_timeout_keepalive(tp, rack, cts);
6992         }
6993         rack_log_to_processing(rack, cts, ret, timers);
6994         return (ret);
6995 }
6996
6997 static void
6998 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6999 {
7000         struct timeval tv;
7001         uint32_t us_cts, flags_on_entry;
7002         uint8_t hpts_removed = 0;
7003
7004         flags_on_entry = rack->r_ctl.rc_hpts_flags;
7005         us_cts = tcp_get_usecs(&tv);
7006         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7007             ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7008              ((tp->snd_max - tp->snd_una) == 0))) {
7009                 tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7010                 hpts_removed = 1;
7011                 /* If we were not delayed cancel out the flag. */
7012                 if ((tp->snd_max - tp->snd_una) == 0)
7013                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7014                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7015         }
7016         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7017                 rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7018                 if (rack->rc_inp->inp_in_hpts &&
7019                     ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7020                         /*
7021                          * Canceling timer's when we have no output being
7022                          * paced. We also must remove ourselves from the
7023                          * hpts.
7024                          */
7025                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7026                         hpts_removed = 1;
7027                 }
7028                 rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7029         }
7030         if (hpts_removed == 0)
7031                 rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7032 }
7033
7034 static void
7035 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7036 {
7037         return;
7038 }
7039
7040 static int
7041 rack_stopall(struct tcpcb *tp)
7042 {
7043         struct tcp_rack *rack;
7044         rack = (struct tcp_rack *)tp->t_fb_ptr;
7045         rack->t_timers_stopped = 1;
7046         return (0);
7047 }
7048
7049 static void
7050 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7051 {
7052         return;
7053 }
7054
7055 static int
7056 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7057 {
7058         return (0);
7059 }
7060
7061 static void
7062 rack_stop_all_timers(struct tcpcb *tp)
7063 {
7064         struct tcp_rack *rack;
7065
7066         /*
7067          * Assure no timers are running.
7068          */
7069         if (tcp_timer_active(tp, TT_PERSIST)) {
7070                 /* We enter in persists, set the flag appropriately */
7071                 rack = (struct tcp_rack *)tp->t_fb_ptr;
7072                 rack->rc_in_persist = 1;
7073         }
7074         tcp_timer_suspend(tp, TT_PERSIST);
7075         tcp_timer_suspend(tp, TT_REXMT);
7076         tcp_timer_suspend(tp, TT_KEEP);
7077         tcp_timer_suspend(tp, TT_DELACK);
7078 }
7079
7080 static void
7081 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7082     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7083 {
7084         int32_t idx;
7085         uint16_t stripped_flags;
7086
7087         rsm->r_rtr_cnt++;
7088         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7089         rsm->r_dupack = 0;
7090         if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7091                 rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7092                 rsm->r_flags |= RACK_OVERMAX;
7093         }
7094         if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7095                 rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7096                 rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7097         }
7098         idx = rsm->r_rtr_cnt - 1;
7099         rsm->r_tim_lastsent[idx] = ts;
7100         stripped_flags = rsm->r_flags & ~(RACK_SENT_SP|RACK_SENT_FP);
7101         if (rsm->r_flags & RACK_ACKED) {
7102                 /* Problably MTU discovery messing with us */
7103                 rsm->r_flags &= ~RACK_ACKED;
7104                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7105         }
7106         if (rsm->r_in_tmap) {
7107                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7108                 rsm->r_in_tmap = 0;
7109         }
7110         TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7111         rsm->r_in_tmap = 1;
7112         if (rsm->r_flags & RACK_SACK_PASSED) {
7113                 /* We have retransmitted due to the SACK pass */
7114                 rsm->r_flags &= ~RACK_SACK_PASSED;
7115                 rsm->r_flags |= RACK_WAS_SACKPASS;
7116         }
7117 }
7118
7119 static uint32_t
7120 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7121     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7122 {
7123         /*
7124          * We (re-)transmitted starting at rsm->r_start for some length
7125          * (possibly less than r_end.
7126          */
7127         struct rack_sendmap *nrsm, *insret;
7128         uint32_t c_end;
7129         int32_t len;
7130
7131         len = *lenp;
7132         c_end = rsm->r_start + len;
7133         if (SEQ_GEQ(c_end, rsm->r_end)) {
7134                 /*
7135                  * We retransmitted the whole piece or more than the whole
7136                  * slopping into the next rsm.
7137                  */
7138                 rack_update_rsm(tp, rack, rsm, ts, add_flag);
7139                 if (c_end == rsm->r_end) {
7140                         *lenp = 0;
7141                         return (0);
7142                 } else {
7143                         int32_t act_len;
7144
7145                         /* Hangs over the end return whats left */
7146                         act_len = rsm->r_end - rsm->r_start;
7147                         *lenp = (len - act_len);
7148                         return (rsm->r_end);
7149                 }
7150                 /* We don't get out of this block. */
7151         }
7152         /*
7153          * Here we retransmitted less than the whole thing which means we
7154          * have to split this into what was transmitted and what was not.
7155          */
7156         nrsm = rack_alloc_full_limit(rack);
7157         if (nrsm == NULL) {
7158                 /*
7159                  * We can't get memory, so lets not proceed.
7160                  */
7161                 *lenp = 0;
7162                 return (0);
7163         }
7164         /*
7165          * So here we are going to take the original rsm and make it what we
7166          * retransmitted. nrsm will be the tail portion we did not
7167          * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7168          * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7169          * 1, 6 and the new piece will be 6, 11.
7170          */
7171         rack_clone_rsm(rack, nrsm, rsm, c_end);
7172         nrsm->r_dupack = 0;
7173         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7174         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7175 #ifdef INVARIANTS
7176         if (insret != NULL) {
7177                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7178                       nrsm, insret, rack, rsm);
7179         }
7180 #endif
7181         if (rsm->r_in_tmap) {
7182                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7183                 nrsm->r_in_tmap = 1;
7184         }
7185         rsm->r_flags &= (~RACK_HAS_FIN);
7186         rack_update_rsm(tp, rack, rsm, ts, add_flag);
7187         /* Log a split of rsm into rsm and nrsm */
7188         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7189         *lenp = 0;
7190         return (0);
7191 }
7192
7193 static void
7194 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7195                 uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts,
7196                 struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff)
7197 {
7198         struct tcp_rack *rack;
7199         struct rack_sendmap *rsm, *nrsm, *insret, fe;
7200         register uint32_t snd_max, snd_una;
7201
7202         /*
7203          * Add to the RACK log of packets in flight or retransmitted. If
7204          * there is a TS option we will use the TS echoed, if not we will
7205          * grab a TS.
7206          *
7207          * Retransmissions will increment the count and move the ts to its
7208          * proper place. Note that if options do not include TS's then we
7209          * won't be able to effectively use the ACK for an RTT on a retran.
7210          *
7211          * Notes about r_start and r_end. Lets consider a send starting at
7212          * sequence 1 for 10 bytes. In such an example the r_start would be
7213          * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7214          * This means that r_end is actually the first sequence for the next
7215          * slot (11).
7216          *
7217          */
7218         /*
7219          * If err is set what do we do XXXrrs? should we not add the thing?
7220          * -- i.e. return if err != 0 or should we pretend we sent it? --
7221          * i.e. proceed with add ** do this for now.
7222          */
7223         INP_WLOCK_ASSERT(tp->t_inpcb);
7224         if (err)
7225                 /*
7226                  * We don't log errors -- we could but snd_max does not
7227                  * advance in this case either.
7228                  */
7229                 return;
7230
7231         if (th_flags & TH_RST) {
7232                 /*
7233                  * We don't log resets and we return immediately from
7234                  * sending
7235                  */
7236                 return;
7237         }
7238         rack = (struct tcp_rack *)tp->t_fb_ptr;
7239         snd_una = tp->snd_una;
7240         snd_max = tp->snd_max;
7241         if (th_flags & (TH_SYN | TH_FIN)) {
7242                 /*
7243                  * The call to rack_log_output is made before bumping
7244                  * snd_max. This means we can record one extra byte on a SYN
7245                  * or FIN if seq_out is adding more on and a FIN is present
7246                  * (and we are not resending).
7247                  */
7248                 if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7249                         len++;
7250                 if (th_flags & TH_FIN)
7251                         len++;
7252                 if (SEQ_LT(snd_max, tp->snd_nxt)) {
7253                         /*
7254                          * The add/update as not been done for the FIN/SYN
7255                          * yet.
7256                          */
7257                         snd_max = tp->snd_nxt;
7258                 }
7259         }
7260         if (SEQ_LEQ((seq_out + len), snd_una)) {
7261                 /* Are sending an old segment to induce an ack (keep-alive)? */
7262                 return;
7263         }
7264         if (SEQ_LT(seq_out, snd_una)) {
7265                 /* huh? should we panic? */
7266                 uint32_t end;
7267
7268                 end = seq_out + len;
7269                 seq_out = snd_una;
7270                 if (SEQ_GEQ(end, seq_out))
7271                         len = end - seq_out;
7272                 else
7273                         len = 0;
7274         }
7275         if (len == 0) {
7276                 /* We don't log zero window probes */
7277                 return;
7278         }
7279         rack->r_ctl.rc_time_last_sent = cts;
7280         if (IN_FASTRECOVERY(tp->t_flags)) {
7281                 rack->r_ctl.rc_prr_out += len;
7282         }
7283         /* First question is it a retransmission or new? */
7284         if (seq_out == snd_max) {
7285                 /* Its new */
7286 again:
7287                 rsm = rack_alloc(rack);
7288                 if (rsm == NULL) {
7289                         /*
7290                          * Hmm out of memory and the tcb got destroyed while
7291                          * we tried to wait.
7292                          */
7293                         return;
7294                 }
7295                 if (th_flags & TH_FIN) {
7296                         rsm->r_flags = RACK_HAS_FIN|add_flag;
7297                 } else {
7298                         rsm->r_flags = add_flag;
7299                 }
7300                 rsm->r_tim_lastsent[0] = cts;
7301                 rsm->r_rtr_cnt = 1;
7302                 rsm->r_rtr_bytes = 0;
7303                 if (th_flags & TH_SYN) {
7304                         /* The data space is one beyond snd_una */
7305                         rsm->r_flags |= RACK_HAS_SYN;
7306                 }
7307                 rsm->r_start = seq_out;
7308                 rsm->r_end = rsm->r_start + len;
7309                 rsm->r_dupack = 0;
7310                 /*
7311                  * save off the mbuf location that
7312                  * sndmbuf_noadv returned (which is
7313                  * where we started copying from)..
7314                  */
7315                 rsm->m = s_mb;
7316                 rsm->soff = s_moff;
7317                 /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7318                 if (rsm->m) {
7319                         if (rsm->m->m_len <= rsm->soff) {
7320                                 /*
7321                                  * XXXrrs Question, will this happen?
7322                                  *
7323                                  * If sbsndptr is set at the correct place
7324                                  * then s_moff should always be somewhere
7325                                  * within rsm->m. But if the sbsndptr was
7326                                  * off then that won't be true. If it occurs
7327                                  * we need to walkout to the correct location.
7328                                  */
7329                                 struct mbuf *lm;
7330
7331                                 lm = rsm->m;
7332                                 while (lm->m_len <= rsm->soff) {
7333                                         rsm->soff -= lm->m_len;
7334                                         lm = lm->m_next;
7335                                         KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7336                                                              __func__, rack, s_moff, s_mb, rsm->soff));
7337                                 }
7338                                 rsm->m = lm;
7339                                 counter_u64_add(rack_sbsndptr_wrong, 1);
7340                         } else
7341                                 counter_u64_add(rack_sbsndptr_right, 1);
7342                         rsm->orig_m_len = rsm->m->m_len;
7343                 } else
7344                         rsm->orig_m_len = 0;
7345                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7346                 /* Log a new rsm */
7347                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7348                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7349 #ifdef INVARIANTS
7350                 if (insret != NULL) {
7351                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7352                               nrsm, insret, rack, rsm);
7353                 }
7354 #endif
7355                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7356                 rsm->r_in_tmap = 1;
7357                 /*
7358                  * Special case detection, is there just a single
7359                  * packet outstanding when we are not in recovery?
7360                  *
7361                  * If this is true mark it so.
7362                  */
7363                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7364                     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7365                         struct rack_sendmap *prsm;
7366
7367                         prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7368                         if (prsm)
7369                                 prsm->r_one_out_nr = 1;
7370                 }
7371                 return;
7372         }
7373         /*
7374          * If we reach here its a retransmission and we need to find it.
7375          */
7376         memset(&fe, 0, sizeof(fe));
7377 more:
7378         if (hintrsm && (hintrsm->r_start == seq_out)) {
7379                 rsm = hintrsm;
7380                 hintrsm = NULL;
7381         } else {
7382                 /* No hints sorry */
7383                 rsm = NULL;
7384         }
7385         if ((rsm) && (rsm->r_start == seq_out)) {
7386                 seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7387                 if (len == 0) {
7388                         return;
7389                 } else {
7390                         goto more;
7391                 }
7392         }
7393         /* Ok it was not the last pointer go through it the hard way. */
7394 refind:
7395         fe.r_start = seq_out;
7396         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7397         if (rsm) {
7398                 if (rsm->r_start == seq_out) {
7399                         seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7400                         if (len == 0) {
7401                                 return;
7402                         } else {
7403                                 goto refind;
7404                         }
7405                 }
7406                 if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7407                         /* Transmitted within this piece */
7408                         /*
7409                          * Ok we must split off the front and then let the
7410                          * update do the rest
7411                          */
7412                         nrsm = rack_alloc_full_limit(rack);
7413                         if (nrsm == NULL) {
7414                                 rack_update_rsm(tp, rack, rsm, cts, add_flag);
7415                                 return;
7416                         }
7417                         /*
7418                          * copy rsm to nrsm and then trim the front of rsm
7419                          * to not include this part.
7420                          */
7421                         rack_clone_rsm(rack, nrsm, rsm, seq_out);
7422                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7423                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7424 #ifdef INVARIANTS
7425                         if (insret != NULL) {
7426                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7427                                       nrsm, insret, rack, rsm);
7428                         }
7429 #endif
7430                         if (rsm->r_in_tmap) {
7431                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7432                                 nrsm->r_in_tmap = 1;
7433                         }
7434                         rsm->r_flags &= (~RACK_HAS_FIN);
7435                         seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7436                         if (len == 0) {
7437                                 return;
7438                         } else if (len > 0)
7439                                 goto refind;
7440                 }
7441         }
7442         /*
7443          * Hmm not found in map did they retransmit both old and on into the
7444          * new?
7445          */
7446         if (seq_out == tp->snd_max) {
7447                 goto again;
7448         } else if (SEQ_LT(seq_out, tp->snd_max)) {
7449 #ifdef INVARIANTS
7450                 printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7451                        seq_out, len, tp->snd_una, tp->snd_max);
7452                 printf("Starting Dump of all rack entries\n");
7453                 RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7454                         printf("rsm:%p start:%u end:%u\n",
7455                                rsm, rsm->r_start, rsm->r_end);
7456                 }
7457                 printf("Dump complete\n");
7458                 panic("seq_out not found rack:%p tp:%p",
7459                       rack, tp);
7460 #endif
7461         } else {
7462 #ifdef INVARIANTS
7463                 /*
7464                  * Hmm beyond sndmax? (only if we are using the new rtt-pack
7465                  * flag)
7466                  */
7467                 panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7468                       seq_out, len, tp->snd_max, tp);
7469 #endif
7470         }
7471 }
7472
7473 /*
7474  * Record one of the RTT updates from an ack into
7475  * our sample structure.
7476  */
7477
7478 static void
7479 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7480                     int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7481 {
7482         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7483             (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7484                 rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7485         }
7486         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7487             (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7488                 rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7489         }
7490         if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7491             if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7492                 rack->r_ctl.rc_gp_lowrtt = us_rtt;
7493             if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7494                     rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7495         }
7496         if ((confidence == 1) &&
7497             ((rsm == NULL) ||
7498              (rsm->r_just_ret) ||
7499              (rsm->r_one_out_nr &&
7500               len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7501                 /*
7502                  * If the rsm had a just return
7503                  * hit it then we can't trust the
7504                  * rtt measurement for buffer deterimination
7505                  * Note that a confidence of 2, indicates
7506                  * SACK'd which overrides the r_just_ret or
7507                  * the r_one_out_nr. If it was a CUM-ACK and
7508                  * we had only two outstanding, but get an
7509                  * ack for only 1. Then that also lowers our
7510                  * confidence.
7511                  */
7512                 confidence = 0;
7513         }
7514         if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7515             (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7516                 if (rack->r_ctl.rack_rs.confidence == 0) {
7517                         /*
7518                          * We take anything with no current confidence
7519                          * saved.
7520                          */
7521                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7522                         rack->r_ctl.rack_rs.confidence = confidence;
7523                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7524                 } else if (confidence || rack->r_ctl.rack_rs.confidence) {
7525                         /*
7526                          * Once we have a confident number,
7527                          * we can update it with a smaller
7528                          * value since this confident number
7529                          * may include the DSACK time until
7530                          * the next segment (the second one) arrived.
7531                          */
7532                         rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7533                         rack->r_ctl.rack_rs.confidence = confidence;
7534                         rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7535                 }
7536         }
7537         rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7538         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7539         rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7540         rack->r_ctl.rack_rs.rs_rtt_cnt++;
7541 }
7542
7543 /*
7544  * Collect new round-trip time estimate
7545  * and update averages and current timeout.
7546  */
7547 static void
7548 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7549 {
7550         int32_t delta;
7551         uint32_t o_srtt, o_var;
7552         int32_t hrtt_up = 0;
7553         int32_t rtt;
7554
7555         if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7556                 /* No valid sample */
7557                 return;
7558         if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7559                 /* We are to use the lowest RTT seen in a single ack */
7560                 rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7561         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7562                 /* We are to use the highest RTT seen in a single ack */
7563                 rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7564         } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7565                 /* We are to use the average RTT seen in a single ack */
7566                 rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7567                                 (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7568         } else {
7569 #ifdef INVARIANTS
7570                 panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7571 #endif
7572                 return;
7573         }
7574         if (rtt == 0)
7575                 rtt = 1;
7576         if (rack->rc_gp_rtt_set == 0) {
7577                 /*
7578                  * With no RTT we have to accept
7579                  * even one we are not confident of.
7580                  */
7581                 rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7582                 rack->rc_gp_rtt_set = 1;
7583         } else if (rack->r_ctl.rack_rs.confidence) {
7584                 /* update the running gp srtt */
7585                 rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7586                 rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7587         }
7588         if (rack->r_ctl.rack_rs.confidence) {
7589                 /*
7590                  * record the low and high for highly buffered path computation,
7591                  * we only do this if we are confident (not a retransmission).
7592                  */
7593                 if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7594                         rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7595                         hrtt_up = 1;
7596                 }
7597                 if (rack->rc_highly_buffered == 0) {
7598                         /*
7599                          * Currently once we declare a path has
7600                          * highly buffered there is no going
7601                          * back, which may be a problem...
7602                          */
7603                         if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7604                                 rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7605                                                      rack->r_ctl.rc_highest_us_rtt,
7606                                                      rack->r_ctl.rc_lowest_us_rtt,
7607                                                      RACK_RTTS_SEEHBP);
7608                                 rack->rc_highly_buffered = 1;
7609                         }
7610                 }
7611         }
7612         if ((rack->r_ctl.rack_rs.confidence) ||
7613             (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7614                 /*
7615                  * If we are highly confident of it <or> it was
7616                  * never retransmitted we accept it as the last us_rtt.
7617                  */
7618                 rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7619                 /* The lowest rtt can be set if its was not retransmited */
7620                 if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7621                         rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7622                         if (rack->r_ctl.rc_lowest_us_rtt == 0)
7623                                 rack->r_ctl.rc_lowest_us_rtt = 1;
7624                 }
7625         }
7626         o_srtt = tp->t_srtt;
7627         o_var = tp->t_rttvar;
7628         rack = (struct tcp_rack *)tp->t_fb_ptr;
7629         if (tp->t_srtt != 0) {
7630                 /*
7631                  * We keep a simple srtt in microseconds, like our rtt
7632                  * measurement. We don't need to do any tricks with shifting
7633                  * etc. Instead we just add in 1/8th of the new measurement
7634                  * and subtract out 1/8 of the old srtt. We do the same with
7635                  * the variance after finding the absolute value of the
7636                  * difference between this sample and the current srtt.
7637                  */
7638                 delta = tp->t_srtt - rtt;
7639                 /* Take off 1/8th of the current sRTT */
7640                 tp->t_srtt -= (tp->t_srtt >> 3);
7641                 /* Add in 1/8th of the new RTT just measured */
7642                 tp->t_srtt += (rtt >> 3);
7643                 if (tp->t_srtt <= 0)
7644                         tp->t_srtt = 1;
7645                 /* Now lets make the absolute value of the variance */
7646                 if (delta < 0)
7647                         delta = -delta;
7648                 /* Subtract out 1/8th */
7649                 tp->t_rttvar -= (tp->t_rttvar >> 3);
7650                 /* Add in 1/8th of the new variance we just saw */
7651                 tp->t_rttvar += (delta >> 3);
7652                 if (tp->t_rttvar <= 0)
7653                         tp->t_rttvar = 1;
7654                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7655                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7656         } else {
7657                 /*
7658                  * No rtt measurement yet - use the unsmoothed rtt. Set the
7659                  * variance to half the rtt (so our first retransmit happens
7660                  * at 3*rtt).
7661                  */
7662                 tp->t_srtt = rtt;
7663                 tp->t_rttvar = rtt >> 1;
7664                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7665         }
7666         rack->rc_srtt_measure_made = 1;
7667         KMOD_TCPSTAT_INC(tcps_rttupdated);
7668         tp->t_rttupdated++;
7669 #ifdef STATS
7670         if (rack_stats_gets_ms_rtt == 0) {
7671                 /* Send in the microsecond rtt used for rxt timeout purposes */
7672                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7673         } else if (rack_stats_gets_ms_rtt == 1) {
7674                 /* Send in the millisecond rtt used for rxt timeout purposes */
7675                 int32_t ms_rtt;
7676
7677                 /* Round up */
7678                 ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7679                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7680         } else if (rack_stats_gets_ms_rtt == 2) {
7681                 /* Send in the millisecond rtt has close to the path RTT as we can get  */
7682                 int32_t ms_rtt;
7683
7684                 /* Round up */
7685                 ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7686                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7687         }  else {
7688                 /* Send in the microsecond rtt has close to the path RTT as we can get  */
7689                 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7690         }
7691
7692 #endif
7693         /*
7694          * the retransmit should happen at rtt + 4 * rttvar. Because of the
7695          * way we do the smoothing, srtt and rttvar will each average +1/2
7696          * tick of bias.  When we compute the retransmit timer, we want 1/2
7697          * tick of rounding and 1 extra tick because of +-1/2 tick
7698          * uncertainty in the firing of the timer.  The bias will give us
7699          * exactly the 1.5 tick we need.  But, because the bias is
7700          * statistical, we have to test that we don't drop below the minimum
7701          * feasible timer (which is 2 ticks).
7702          */
7703         tp->t_rxtshift = 0;
7704         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7705                       max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7706         rack_log_rtt_sample(rack, rtt);
7707         tp->t_softerror = 0;
7708 }
7709
7710
7711 static void
7712 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7713 {
7714         /*
7715          * Apply to filter the inbound us-rtt at us_cts.
7716          */
7717         uint32_t old_rtt;
7718
7719         old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7720         apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7721                                us_rtt, us_cts);
7722         if (rack->r_ctl.last_pacing_time &&
7723             rack->rc_gp_dyn_mul &&
7724             (rack->r_ctl.last_pacing_time > us_rtt))
7725                 rack->pacing_longer_than_rtt = 1;
7726         else
7727                 rack->pacing_longer_than_rtt = 0;
7728         if (old_rtt > us_rtt) {
7729                 /* We just hit a new lower rtt time */
7730                 rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7731                                      __LINE__, RACK_RTTS_NEWRTT);
7732                 /*
7733                  * Only count it if its lower than what we saw within our
7734                  * calculated range.
7735                  */
7736                 if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7737                         if (rack_probertt_lower_within &&
7738                             rack->rc_gp_dyn_mul &&
7739                             (rack->use_fixed_rate == 0) &&
7740                             (rack->rc_always_pace)) {
7741                                 /*
7742                                  * We are seeing a new lower rtt very close
7743                                  * to the time that we would have entered probe-rtt.
7744                                  * This is probably due to the fact that a peer flow
7745                                  * has entered probe-rtt. Lets go in now too.
7746                                  */
7747                                 uint32_t val;
7748
7749                                 val = rack_probertt_lower_within * rack_time_between_probertt;
7750                                 val /= 100;
7751                                 if ((rack->in_probe_rtt == 0)  &&
7752                                     ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) {
7753                                         rack_enter_probertt(rack, us_cts);
7754                                 }
7755                         }
7756                         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7757                 }
7758         }
7759 }
7760
7761 static int
7762 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7763     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7764 {
7765         int32_t i, all;
7766         uint32_t t, len_acked;
7767
7768         if ((rsm->r_flags & RACK_ACKED) ||
7769             (rsm->r_flags & RACK_WAS_ACKED))
7770                 /* Already done */
7771                 return (0);
7772         if (rsm->r_no_rtt_allowed) {
7773                 /* Not allowed */
7774                 return (0);
7775         }
7776         if (ack_type == CUM_ACKED) {
7777                 if (SEQ_GT(th_ack, rsm->r_end)) {
7778                         len_acked = rsm->r_end - rsm->r_start;
7779                         all = 1;
7780                 } else {
7781                         len_acked = th_ack - rsm->r_start;
7782                         all = 0;
7783                 }
7784         } else {
7785                 len_acked = rsm->r_end - rsm->r_start;
7786                 all = 0;
7787         }
7788         if (rsm->r_rtr_cnt == 1) {
7789                 uint32_t us_rtt;
7790
7791                 t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7792                 if ((int)t <= 0)
7793                         t = 1;
7794                 if (!tp->t_rttlow || tp->t_rttlow > t)
7795                         tp->t_rttlow = t;
7796                 if (!rack->r_ctl.rc_rack_min_rtt ||
7797                     SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7798                         rack->r_ctl.rc_rack_min_rtt = t;
7799                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7800                                 rack->r_ctl.rc_rack_min_rtt = 1;
7801                         }
7802                 }
7803                 if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7804                         us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7805                 else
7806                         us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7807                 if (us_rtt == 0)
7808                         us_rtt = 1;
7809                 rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7810                 if (ack_type == SACKED) {
7811                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7812                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7813                 } else {
7814                         /*
7815                          * We need to setup what our confidence
7816                          * is in this ack.
7817                          *
7818                          * If the rsm was app limited and it is
7819                          * less than a mss in length (the end
7820                          * of the send) then we have a gap. If we
7821                          * were app limited but say we were sending
7822                          * multiple MSS's then we are more confident
7823                          * int it.
7824                          *
7825                          * When we are not app-limited then we see if
7826                          * the rsm is being included in the current
7827                          * measurement, we tell this by the app_limited_needs_set
7828                          * flag.
7829                          *
7830                          * Note that being cwnd blocked is not applimited
7831                          * as well as the pacing delay between packets which
7832                          * are sending only 1 or 2 MSS's also will show up
7833                          * in the RTT. We probably need to examine this algorithm
7834                          * a bit more and enhance it to account for the delay
7835                          * between rsm's. We could do that by saving off the
7836                          * pacing delay of each rsm (in an rsm) and then
7837                          * factoring that in somehow though for now I am
7838                          * not sure how :)
7839                          */
7840                         int calc_conf = 0;
7841
7842                         if (rsm->r_flags & RACK_APP_LIMITED) {
7843                                 if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7844                                         calc_conf = 0;
7845                                 else
7846                                         calc_conf = 1;
7847                         } else if (rack->app_limited_needs_set == 0) {
7848                                 calc_conf = 1;
7849                         } else {
7850                                 calc_conf = 0;
7851                         }
7852                         rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7853                         tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7854                                             calc_conf, rsm, rsm->r_rtr_cnt);
7855                 }
7856                 if ((rsm->r_flags & RACK_TLP) &&
7857                     (!IN_FASTRECOVERY(tp->t_flags))) {
7858                         /* Segment was a TLP and our retrans matched */
7859                         if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7860                                 rack->r_ctl.rc_rsm_start = tp->snd_max;
7861                                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
7862                                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
7863                                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
7864                         }
7865                 }
7866                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7867                         /* New more recent rack_tmit_time */
7868                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7869                         rack->rc_rack_rtt = t;
7870                 }
7871                 return (1);
7872         }
7873         /*
7874          * We clear the soft/rxtshift since we got an ack.
7875          * There is no assurance we will call the commit() function
7876          * so we need to clear these to avoid incorrect handling.
7877          */
7878         tp->t_rxtshift = 0;
7879         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7880                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7881         tp->t_softerror = 0;
7882         if (to && (to->to_flags & TOF_TS) &&
7883             (ack_type == CUM_ACKED) &&
7884             (to->to_tsecr) &&
7885             ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7886                 /*
7887                  * Now which timestamp does it match? In this block the ACK
7888                  * must be coming from a previous transmission.
7889                  */
7890                 for (i = 0; i < rsm->r_rtr_cnt; i++) {
7891                         if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7892                                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7893                                 if ((int)t <= 0)
7894                                         t = 1;
7895                                 if ((i + 1) < rsm->r_rtr_cnt) {
7896                                         /*
7897                                          * The peer ack'd from our previous
7898                                          * transmission. We have a spurious
7899                                          * retransmission and thus we dont
7900                                          * want to update our rack_rtt.
7901                                          */
7902                                         return (0);
7903                                 }
7904                                 if (!tp->t_rttlow || tp->t_rttlow > t)
7905                                         tp->t_rttlow = t;
7906                                 if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7907                                         rack->r_ctl.rc_rack_min_rtt = t;
7908                                         if (rack->r_ctl.rc_rack_min_rtt == 0) {
7909                                                 rack->r_ctl.rc_rack_min_rtt = 1;
7910                                         }
7911                                 }
7912                                 if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7913                                            (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7914                                         /* New more recent rack_tmit_time */
7915                                         rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7916                                         rack->rc_rack_rtt = t;
7917                                 }
7918                                 rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7919                                 tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7920                                                     rsm->r_rtr_cnt);
7921                                 return (1);
7922                         }
7923                 }
7924                 goto ts_not_found;
7925         } else {
7926                 /*
7927                  * Ok its a SACK block that we retransmitted. or a windows
7928                  * machine without timestamps. We can tell nothing from the
7929                  * time-stamp since its not there or the time the peer last
7930                  * recieved a segment that moved forward its cum-ack point.
7931                  */
7932 ts_not_found:
7933                 i = rsm->r_rtr_cnt - 1;
7934                 t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7935                 if ((int)t <= 0)
7936                         t = 1;
7937                 if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7938                         /*
7939                          * We retransmitted and the ack came back in less
7940                          * than the smallest rtt we have observed. We most
7941                          * likely did an improper retransmit as outlined in
7942                          * 6.2 Step 2 point 2 in the rack-draft so we
7943                          * don't want to update our rack_rtt. We in
7944                          * theory (in future) might want to think about reverting our
7945                          * cwnd state but we won't for now.
7946                          */
7947                         return (0);
7948                 } else if (rack->r_ctl.rc_rack_min_rtt) {
7949                         /*
7950                          * We retransmitted it and the retransmit did the
7951                          * job.
7952                          */
7953                         if (!rack->r_ctl.rc_rack_min_rtt ||
7954                             SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7955                                 rack->r_ctl.rc_rack_min_rtt = t;
7956                                 if (rack->r_ctl.rc_rack_min_rtt == 0) {
7957                                         rack->r_ctl.rc_rack_min_rtt = 1;
7958                                 }
7959                         }
7960                         if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7961                                 /* New more recent rack_tmit_time */
7962                                 rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7963                                 rack->rc_rack_rtt = t;
7964                         }
7965                         return (1);
7966                 }
7967         }
7968         return (0);
7969 }
7970
7971 /*
7972  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7973  */
7974 static void
7975 rack_log_sack_passed(struct tcpcb *tp,
7976     struct tcp_rack *rack, struct rack_sendmap *rsm)
7977 {
7978         struct rack_sendmap *nrsm;
7979
7980         nrsm = rsm;
7981         TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7982             rack_head, r_tnext) {
7983                 if (nrsm == rsm) {
7984                         /* Skip orginal segment he is acked */
7985                         continue;
7986                 }
7987                 if (nrsm->r_flags & RACK_ACKED) {
7988                         /*
7989                          * Skip ack'd segments, though we
7990                          * should not see these, since tmap
7991                          * should not have ack'd segments.
7992                          */
7993                         continue;
7994                 }
7995                 if (nrsm->r_flags & RACK_SACK_PASSED) {
7996                         /*
7997                          * We found one that is already marked
7998                          * passed, we have been here before and
7999                          * so all others below this are marked.
8000                          */
8001                         break;
8002                 }
8003                 nrsm->r_flags |= RACK_SACK_PASSED;
8004                 nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8005         }
8006 }
8007
8008 static void
8009 rack_need_set_test(struct tcpcb *tp,
8010                    struct tcp_rack *rack,
8011                    struct rack_sendmap *rsm,
8012                    tcp_seq th_ack,
8013                    int line,
8014                    int use_which)
8015 {
8016
8017         if ((tp->t_flags & TF_GPUTINPROG) &&
8018             SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8019                 /*
8020                  * We were app limited, and this ack
8021                  * butts up or goes beyond the point where we want
8022                  * to start our next measurement. We need
8023                  * to record the new gput_ts as here and
8024                  * possibly update the start sequence.
8025                  */
8026                 uint32_t seq, ts;
8027
8028                 if (rsm->r_rtr_cnt > 1) {
8029                         /*
8030                          * This is a retransmit, can we
8031                          * really make any assessment at this
8032                          * point?  We are not really sure of
8033                          * the timestamp, is it this or the
8034                          * previous transmission?
8035                          *
8036                          * Lets wait for something better that
8037                          * is not retransmitted.
8038                          */
8039                         return;
8040                 }
8041                 seq = tp->gput_seq;
8042                 ts = tp->gput_ts;
8043                 rack->app_limited_needs_set = 0;
8044                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8045                 /* Do we start at a new end? */
8046                 if ((use_which == RACK_USE_BEG) &&
8047                     SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8048                         /*
8049                          * When we get an ACK that just eats
8050                          * up some of the rsm, we set RACK_USE_BEG
8051                          * since whats at r_start (i.e. th_ack)
8052                          * is left unacked and thats where the
8053                          * measurement not starts.
8054                          */
8055                         tp->gput_seq = rsm->r_start;
8056                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8057                 }
8058                 if ((use_which == RACK_USE_END) &&
8059                     SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8060                             /*
8061                              * We use the end when the cumack
8062                              * is moving forward and completely
8063                              * deleting the rsm passed so basically
8064                              * r_end holds th_ack.
8065                              *
8066                              * For SACK's we also want to use the end
8067                              * since this piece just got sacked and
8068                              * we want to target anything after that
8069                              * in our measurement.
8070                              */
8071                             tp->gput_seq = rsm->r_end;
8072                             rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8073                 }
8074                 if (use_which == RACK_USE_END_OR_THACK) {
8075                         /*
8076                          * special case for ack moving forward,
8077                          * not a sack, we need to move all the
8078                          * way up to where this ack cum-ack moves
8079                          * to.
8080                          */
8081                         if (SEQ_GT(th_ack, rsm->r_end))
8082                                 tp->gput_seq = th_ack;
8083                         else
8084                                 tp->gput_seq = rsm->r_end;
8085                         rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8086                 }
8087                 if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8088                         /*
8089                          * We moved beyond this guy's range, re-calculate
8090                          * the new end point.
8091                          */
8092                         if (rack->rc_gp_filled == 0) {
8093                                 tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8094                         } else {
8095                                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8096                         }
8097                 }
8098                 /*
8099                  * We are moving the goal post, we may be able to clear the
8100                  * measure_saw_probe_rtt flag.
8101                  */
8102                 if ((rack->in_probe_rtt == 0) &&
8103                     (rack->measure_saw_probe_rtt) &&
8104                     (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8105                         rack->measure_saw_probe_rtt = 0;
8106                 rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8107                                            seq, tp->gput_seq, 0, 5, line, NULL);
8108                 if (rack->rc_gp_filled &&
8109                     ((tp->gput_ack - tp->gput_seq) <
8110                      max(rc_init_window(rack), (MIN_GP_WIN *
8111                                                 ctf_fixed_maxseg(tp))))) {
8112                         uint32_t ideal_amount;
8113
8114                         ideal_amount = rack_get_measure_window(tp, rack);
8115                         if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8116                                 /*
8117                                  * There is no sense of continuing this measurement
8118                                  * because its too small to gain us anything we
8119                                  * trust. Skip it and that way we can start a new
8120                                  * measurement quicker.
8121                                  */
8122                                 tp->t_flags &= ~TF_GPUTINPROG;
8123                                 rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8124                                                            0, 0, 0, 6, __LINE__, NULL);
8125                         } else {
8126                                 /*
8127                                  * Reset the window further out.
8128                                  */
8129                                 tp->gput_ack = tp->gput_seq + ideal_amount;
8130                         }
8131                 }
8132         }
8133 }
8134
8135 static uint32_t
8136 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8137                    struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8138 {
8139         uint32_t start, end, changed = 0;
8140         struct rack_sendmap stack_map;
8141         struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
8142         int32_t used_ref = 1;
8143         int moved = 0;
8144
8145         start = sack->start;
8146         end = sack->end;
8147         rsm = *prsm;
8148         memset(&fe, 0, sizeof(fe));
8149 do_rest_ofb:
8150         if ((rsm == NULL) ||
8151             (SEQ_LT(end, rsm->r_start)) ||
8152             (SEQ_GEQ(start, rsm->r_end)) ||
8153             (SEQ_LT(start, rsm->r_start))) {
8154                 /*
8155                  * We are not in the right spot,
8156                  * find the correct spot in the tree.
8157                  */
8158                 used_ref = 0;
8159                 fe.r_start = start;
8160                 rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8161                 moved++;
8162         }
8163         if (rsm == NULL) {
8164                 /* TSNH */
8165                 goto out;
8166         }
8167         /* Ok we have an ACK for some piece of this rsm */
8168         if (rsm->r_start != start) {
8169                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8170                         /**
8171                          * Need to split this in two pieces the before and after,
8172                          * the before remains in the map, the after must be
8173                          * added. In other words we have:
8174                          * rsm        |--------------|
8175                          * sackblk        |------->
8176                          * rsm will become
8177                          *     rsm    |---|
8178                          * and nrsm will be  the sacked piece
8179                          *     nrsm       |----------|
8180                          *
8181                          * But before we start down that path lets
8182                          * see if the sack spans over on top of
8183                          * the next guy and it is already sacked.
8184                          */
8185                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8186                         if (next && (next->r_flags & RACK_ACKED) &&
8187                             SEQ_GEQ(end, next->r_start)) {
8188                                 /**
8189                                  * So the next one is already acked, and
8190                                  * we can thus by hookery use our stack_map
8191                                  * to reflect the piece being sacked and
8192                                  * then adjust the two tree entries moving
8193                                  * the start and ends around. So we start like:
8194                                  *  rsm     |------------|             (not-acked)
8195                                  *  next                 |-----------| (acked)
8196                                  *  sackblk        |-------->
8197                                  *  We want to end like so:
8198                                  *  rsm     |------|                   (not-acked)
8199                                  *  next           |-----------------| (acked)
8200                                  *  nrsm           |-----|
8201                                  * Where nrsm is a temporary stack piece we
8202                                  * use to update all the gizmos.
8203                                  */
8204                                 /* Copy up our fudge block */
8205                                 nrsm = &stack_map;
8206                                 memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8207                                 /* Now adjust our tree blocks */
8208                                 rsm->r_end = start;
8209                                 next->r_start = start;
8210                                 /* Now we must adjust back where next->m is */
8211                                 rack_setup_offset_for_rsm(rsm, next);
8212
8213                                 /* We don't need to adjust rsm, it did not change */
8214                                 /* Clear out the dup ack count of the remainder */
8215                                 rsm->r_dupack = 0;
8216                                 rsm->r_just_ret = 0;
8217                                 rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8218                                 /* Now lets make sure our fudge block is right */
8219                                 nrsm->r_start = start;
8220                                 /* Now lets update all the stats and such */
8221                                 rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8222                                 if (rack->app_limited_needs_set)
8223                                         rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8224                                 changed += (nrsm->r_end - nrsm->r_start);
8225                                 rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8226                                 if (nrsm->r_flags & RACK_SACK_PASSED) {
8227                                         counter_u64_add(rack_reorder_seen, 1);
8228                                         rack->r_ctl.rc_reorder_ts = cts;
8229                                 }
8230                                 /*
8231                                  * Now we want to go up from rsm (the
8232                                  * one left un-acked) to the next one
8233                                  * in the tmap. We do this so when
8234                                  * we walk backwards we include marking
8235                                  * sack-passed on rsm (The one passed in
8236                                  * is skipped since it is generally called
8237                                  * on something sacked before removing it
8238                                  * from the tmap).
8239                                  */
8240                                 if (rsm->r_in_tmap) {
8241                                         nrsm = TAILQ_NEXT(rsm, r_tnext);
8242                                         /*
8243                                          * Now that we have the next
8244                                          * one walk backwards from there.
8245                                          */
8246                                         if (nrsm && nrsm->r_in_tmap)
8247                                                 rack_log_sack_passed(tp, rack, nrsm);
8248                                 }
8249                                 /* Now are we done? */
8250                                 if (SEQ_LT(end, next->r_end) ||
8251                                     (end == next->r_end)) {
8252                                         /* Done with block */
8253                                         goto out;
8254                                 }
8255                                 rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8256                                 counter_u64_add(rack_sack_used_next_merge, 1);
8257                                 /* Postion for the next block */
8258                                 start = next->r_end;
8259                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8260                                 if (rsm == NULL)
8261                                         goto out;
8262                         } else {
8263                                 /**
8264                                  * We can't use any hookery here, so we
8265                                  * need to split the map. We enter like
8266                                  * so:
8267                                  *  rsm      |--------|
8268                                  *  sackblk       |----->
8269                                  * We will add the new block nrsm and
8270                                  * that will be the new portion, and then
8271                                  * fall through after reseting rsm. So we
8272                                  * split and look like this:
8273                                  *  rsm      |----|
8274                                  *  sackblk       |----->
8275                                  *  nrsm          |---|
8276                                  * We then fall through reseting
8277                                  * rsm to nrsm, so the next block
8278                                  * picks it up.
8279                                  */
8280                                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8281                                 if (nrsm == NULL) {
8282                                         /*
8283                                          * failed XXXrrs what can we do but loose the sack
8284                                          * info?
8285                                          */
8286                                         goto out;
8287                                 }
8288                                 counter_u64_add(rack_sack_splits, 1);
8289                                 rack_clone_rsm(rack, nrsm, rsm, start);
8290                                 rsm->r_just_ret = 0;
8291                                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8292 #ifdef INVARIANTS
8293                                 if (insret != NULL) {
8294                                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8295                                               nrsm, insret, rack, rsm);
8296                                 }
8297 #endif
8298                                 if (rsm->r_in_tmap) {
8299                                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8300                                         nrsm->r_in_tmap = 1;
8301                                 }
8302                                 rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8303                                 rsm->r_flags &= (~RACK_HAS_FIN);
8304                                 /* Position us to point to the new nrsm that starts the sack blk */
8305                                 rsm = nrsm;
8306                         }
8307                 } else {
8308                         /* Already sacked this piece */
8309                         counter_u64_add(rack_sack_skipped_acked, 1);
8310                         moved++;
8311                         if (end == rsm->r_end) {
8312                                 /* Done with block */
8313                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8314                                 goto out;
8315                         } else if (SEQ_LT(end, rsm->r_end)) {
8316                                 /* A partial sack to a already sacked block */
8317                                 moved++;
8318                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8319                                 goto out;
8320                         } else {
8321                                 /*
8322                                  * The end goes beyond this guy
8323                                  * repostion the start to the
8324                                  * next block.
8325                                  */
8326                                 start = rsm->r_end;
8327                                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8328                                 if (rsm == NULL)
8329                                         goto out;
8330                         }
8331                 }
8332         }
8333         if (SEQ_GEQ(end, rsm->r_end)) {
8334                 /**
8335                  * The end of this block is either beyond this guy or right
8336                  * at this guy. I.e.:
8337                  *  rsm ---                 |-----|
8338                  *  end                     |-----|
8339                  *  <or>
8340                  *  end                     |---------|
8341                  */
8342                 if ((rsm->r_flags & RACK_ACKED) == 0) {
8343                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8344                         changed += (rsm->r_end - rsm->r_start);
8345                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8346                         if (rsm->r_in_tmap) /* should be true */
8347                                 rack_log_sack_passed(tp, rack, rsm);
8348                         /* Is Reordering occuring? */
8349                         if (rsm->r_flags & RACK_SACK_PASSED) {
8350                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8351                                 counter_u64_add(rack_reorder_seen, 1);
8352                                 rack->r_ctl.rc_reorder_ts = cts;
8353                         }
8354                         if (rack->app_limited_needs_set)
8355                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8356                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8357                         rsm->r_flags |= RACK_ACKED;
8358                         rsm->r_flags &= ~RACK_TLP;
8359                         if (rsm->r_in_tmap) {
8360                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8361                                 rsm->r_in_tmap = 0;
8362                         }
8363                         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8364                 } else {
8365                         counter_u64_add(rack_sack_skipped_acked, 1);
8366                         moved++;
8367                 }
8368                 if (end == rsm->r_end) {
8369                         /* This block only - done, setup for next */
8370                         goto out;
8371                 }
8372                 /*
8373                  * There is more not coverend by this rsm move on
8374                  * to the next block in the RB tree.
8375                  */
8376                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8377                 start = rsm->r_end;
8378                 rsm = nrsm;
8379                 if (rsm == NULL)
8380                         goto out;
8381                 goto do_rest_ofb;
8382         }
8383         /**
8384          * The end of this sack block is smaller than
8385          * our rsm i.e.:
8386          *  rsm ---                 |-----|
8387          *  end                     |--|
8388          */
8389         if ((rsm->r_flags & RACK_ACKED) == 0) {
8390                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8391                 if (prev && (prev->r_flags & RACK_ACKED)) {
8392                         /**
8393                          * Goal, we want the right remainder of rsm to shrink
8394                          * in place and span from (rsm->r_start = end) to rsm->r_end.
8395                          * We want to expand prev to go all the way
8396                          * to prev->r_end <- end.
8397                          * so in the tree we have before:
8398                          *   prev     |--------|         (acked)
8399                          *   rsm               |-------| (non-acked)
8400                          *   sackblk           |-|
8401                          * We churn it so we end up with
8402                          *   prev     |----------|       (acked)
8403                          *   rsm                 |-----| (non-acked)
8404                          *   nrsm              |-| (temporary)
8405                          */
8406                         nrsm = &stack_map;
8407                         memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8408                         prev->r_end = end;
8409                         rsm->r_start = end;
8410                         /* Now adjust nrsm (stack copy) to be
8411                          * the one that is the small
8412                          * piece that was "sacked".
8413                          */
8414                         nrsm->r_end = end;
8415                         rsm->r_dupack = 0;
8416                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8417                         /*
8418                          * Now that the rsm has had its start moved forward
8419                          * lets go ahead and get its new place in the world.
8420                          */
8421                         rack_setup_offset_for_rsm(prev, rsm);
8422                         /*
8423                          * Now nrsm is our new little piece
8424                          * that is acked (which was merged
8425                          * to prev). Update the rtt and changed
8426                          * based on that. Also check for reordering.
8427                          */
8428                         rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8429                         if (rack->app_limited_needs_set)
8430                                 rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8431                         changed += (nrsm->r_end - nrsm->r_start);
8432                         rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8433                         if (nrsm->r_flags & RACK_SACK_PASSED) {
8434                                 counter_u64_add(rack_reorder_seen, 1);
8435                                 rack->r_ctl.rc_reorder_ts = cts;
8436                         }
8437                         rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8438                         rsm = prev;
8439                         counter_u64_add(rack_sack_used_prev_merge, 1);
8440                 } else {
8441                         /**
8442                          * This is the case where our previous
8443                          * block is not acked either, so we must
8444                          * split the block in two.
8445                          */
8446                         nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8447                         if (nrsm == NULL) {
8448                                 /* failed rrs what can we do but loose the sack info? */
8449                                 goto out;
8450                         }
8451                         /**
8452                          * In this case nrsm becomes
8453                          * nrsm->r_start = end;
8454                          * nrsm->r_end = rsm->r_end;
8455                          * which is un-acked.
8456                          * <and>
8457                          * rsm->r_end = nrsm->r_start;
8458                          * i.e. the remaining un-acked
8459                          * piece is left on the left
8460                          * hand side.
8461                          *
8462                          * So we start like this
8463                          * rsm      |----------| (not acked)
8464                          * sackblk  |---|
8465                          * build it so we have
8466                          * rsm      |---|         (acked)
8467                          * nrsm         |------|  (not acked)
8468                          */
8469                         counter_u64_add(rack_sack_splits, 1);
8470                         rack_clone_rsm(rack, nrsm, rsm, end);
8471                         rsm->r_flags &= (~RACK_HAS_FIN);
8472                         rsm->r_just_ret = 0;
8473                         insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8474 #ifdef INVARIANTS
8475                         if (insret != NULL) {
8476                                 panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8477                                       nrsm, insret, rack, rsm);
8478                         }
8479 #endif
8480                         if (rsm->r_in_tmap) {
8481                                 TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8482                                 nrsm->r_in_tmap = 1;
8483                         }
8484                         nrsm->r_dupack = 0;
8485                         rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8486                         rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8487                         changed += (rsm->r_end - rsm->r_start);
8488                         rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8489                         if (rsm->r_in_tmap) /* should be true */
8490                                 rack_log_sack_passed(tp, rack, rsm);
8491                         /* Is Reordering occuring? */
8492                         if (rsm->r_flags & RACK_SACK_PASSED) {
8493                                 rsm->r_flags &= ~RACK_SACK_PASSED;
8494                                 counter_u64_add(rack_reorder_seen, 1);
8495                                 rack->r_ctl.rc_reorder_ts = cts;
8496                         }
8497                         if (rack->app_limited_needs_set)
8498                                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8499                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8500                         rsm->r_flags |= RACK_ACKED;
8501                         rsm->r_flags &= ~RACK_TLP;
8502                         rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8503                         if (rsm->r_in_tmap) {
8504                                 TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8505                                 rsm->r_in_tmap = 0;
8506                         }
8507                 }
8508         } else if (start != end){
8509                 /*
8510                  * The block was already acked.
8511                  */
8512                 counter_u64_add(rack_sack_skipped_acked, 1);
8513                 moved++;
8514         }
8515 out:
8516         if (rsm && (rsm->r_flags & RACK_ACKED)) {
8517                 /*
8518                  * Now can we merge where we worked
8519                  * with either the previous or
8520                  * next block?
8521                  */
8522                 next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8523                 while (next) {
8524                     if (next->r_flags & RACK_ACKED) {
8525                         /* yep this and next can be merged */
8526                         rsm = rack_merge_rsm(rack, rsm, next);
8527                         next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8528                     } else
8529                             break;
8530                 }
8531                 /* Now what about the previous? */
8532                 prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8533                 while (prev) {
8534                     if (prev->r_flags & RACK_ACKED) {
8535                         /* yep the previous and this can be merged */
8536                         rsm = rack_merge_rsm(rack, prev, rsm);
8537                         prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8538                     } else
8539                             break;
8540                 }
8541         }
8542         if (used_ref == 0) {
8543                 counter_u64_add(rack_sack_proc_all, 1);
8544         } else {
8545                 counter_u64_add(rack_sack_proc_short, 1);
8546         }
8547         /* Save off the next one for quick reference. */
8548         if (rsm)
8549                 nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8550         else
8551                 nrsm = NULL;
8552         *prsm = rack->r_ctl.rc_sacklast = nrsm;
8553         /* Pass back the moved. */
8554         *moved_two = moved;
8555         return (changed);
8556 }
8557
8558 static void inline
8559 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8560 {
8561         struct rack_sendmap *tmap;
8562
8563         tmap = NULL;
8564         while (rsm && (rsm->r_flags & RACK_ACKED)) {
8565                 /* Its no longer sacked, mark it so */
8566                 rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8567 #ifdef INVARIANTS
8568                 if (rsm->r_in_tmap) {
8569                         panic("rack:%p rsm:%p flags:0x%x in tmap?",
8570                               rack, rsm, rsm->r_flags);
8571                 }
8572 #endif
8573                 rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8574                 /* Rebuild it into our tmap */
8575                 if (tmap == NULL) {
8576                         TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8577                         tmap = rsm;
8578                 } else {
8579                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8580                         tmap = rsm;
8581                 }
8582                 tmap->r_in_tmap = 1;
8583                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8584         }
8585         /*
8586          * Now lets possibly clear the sack filter so we start
8587          * recognizing sacks that cover this area.
8588          */
8589         sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8590
8591 }
8592
8593 static void
8594 rack_do_decay(struct tcp_rack *rack)
8595 {
8596         struct timeval res;
8597
8598 #define timersub(tvp, uvp, vvp)                                         \
8599         do {                                                            \
8600                 (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;          \
8601                 (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;       \
8602                 if ((vvp)->tv_usec < 0) {                               \
8603                         (vvp)->tv_sec--;                                \
8604                         (vvp)->tv_usec += 1000000;                      \
8605                 }                                                       \
8606         } while (0)
8607
8608         timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8609 #undef timersub
8610
8611         rack->r_ctl.input_pkt++;
8612         if ((rack->rc_in_persist) ||
8613             (res.tv_sec >= 1) ||
8614             (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8615                 /*
8616                  * Check for decay of non-SAD,
8617                  * we want all SAD detection metrics to
8618                  * decay 1/4 per second (or more) passed.
8619                  */
8620                 uint32_t pkt_delta;
8621
8622                 pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8623                 /* Update our saved tracking values */
8624                 rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8625                 rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8626                 /* Now do we escape without decay? */
8627 #ifdef NETFLIX_EXP_DETECTION
8628                 if (rack->rc_in_persist ||
8629                     (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8630                     (pkt_delta < tcp_sad_low_pps)){
8631                         /*
8632                          * We don't decay idle connections
8633                          * or ones that have a low input pps.
8634                          */
8635                         return;
8636                 }
8637                 /* Decay the counters */
8638                 rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8639                                                         tcp_sad_decay_val);
8640                 rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8641                                                          tcp_sad_decay_val);
8642                 rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8643                                                                tcp_sad_decay_val);
8644                 rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8645                                                                 tcp_sad_decay_val);
8646 #endif
8647         }
8648 }
8649
8650 static void
8651 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8652 {
8653         struct rack_sendmap *rsm, *rm;
8654
8655         /*
8656          * The ACK point is advancing to th_ack, we must drop off
8657          * the packets in the rack log and calculate any eligble
8658          * RTT's.
8659          */
8660         rack->r_wanted_output = 1;
8661 more:
8662         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8663         if (rsm == NULL) {
8664                 if ((th_ack - 1) == tp->iss) {
8665                         /*
8666                          * For the SYN incoming case we will not
8667                          * have called tcp_output for the sending of
8668                          * the SYN, so there will be no map. All
8669                          * other cases should probably be a panic.
8670                          */
8671                         return;
8672                 }
8673                 if (tp->t_flags & TF_SENTFIN) {
8674                         /* if we sent a FIN we often will not have map */
8675                         return;
8676                 }
8677 #ifdef INVARIANTS
8678                 panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8679                       tp,
8680                       tp->t_state, th_ack, rack,
8681                       tp->snd_una, tp->snd_max, tp->snd_nxt);
8682 #endif
8683                 return;
8684         }
8685         if (SEQ_LT(th_ack, rsm->r_start)) {
8686                 /* Huh map is missing this */
8687 #ifdef INVARIANTS
8688                 printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8689                        rsm->r_start,
8690                        th_ack, tp->t_state, rack->r_state);
8691 #endif
8692                 return;
8693         }
8694         rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8695         /* Now do we consume the whole thing? */
8696         if (SEQ_GEQ(th_ack, rsm->r_end)) {
8697                 /* Its all consumed. */
8698                 uint32_t left;
8699                 uint8_t newly_acked;
8700
8701                 rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8702                 rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8703                 rsm->r_rtr_bytes = 0;
8704                 /* Record the time of highest cumack sent */
8705                 rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8706                 rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8707 #ifdef INVARIANTS
8708                 if (rm != rsm) {
8709                         panic("removing head in rack:%p rsm:%p rm:%p",
8710                               rack, rsm, rm);
8711                 }
8712 #endif
8713                 if (rsm->r_in_tmap) {
8714                         TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8715                         rsm->r_in_tmap = 0;
8716                 }
8717                 newly_acked = 1;
8718                 if (rsm->r_flags & RACK_ACKED) {
8719                         /*
8720                          * It was acked on the scoreboard -- remove
8721                          * it from total
8722                          */
8723                         rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8724                         newly_acked = 0;
8725                 } else if (rsm->r_flags & RACK_SACK_PASSED) {
8726                         /*
8727                          * There are segments ACKED on the
8728                          * scoreboard further up. We are seeing
8729                          * reordering.
8730                          */
8731                         rsm->r_flags &= ~RACK_SACK_PASSED;
8732                         counter_u64_add(rack_reorder_seen, 1);
8733                         rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8734                         rsm->r_flags |= RACK_ACKED;
8735                         rack->r_ctl.rc_reorder_ts = cts;
8736                         if (rack->r_ent_rec_ns) {
8737                                 /*
8738                                  * We have sent no more, and we saw an sack
8739                                  * then ack arrive.
8740                                  */
8741                                 rack->r_might_revert = 1;
8742                         }
8743                 }
8744                 if ((rsm->r_flags & RACK_TO_REXT) &&
8745                     (tp->t_flags & TF_RCVD_TSTMP) &&
8746                     (to->to_flags & TOF_TS) &&
8747                     (tp->t_flags & TF_PREVVALID)) {
8748                         /*
8749                          * We can use the timestamp to see
8750                          * if this retransmission was from the
8751                          * first transmit. If so we made a mistake.
8752                          */
8753                         tp->t_flags &= ~TF_PREVVALID;
8754                         if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
8755                                 /* The first transmit is what this ack is for */
8756                                 rack_cong_signal(tp, CC_RTO_ERR, th_ack);
8757                         }
8758                 }
8759                 left = th_ack - rsm->r_end;
8760                 if (rack->app_limited_needs_set && newly_acked)
8761                         rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
8762                 /* Free back to zone */
8763                 rack_free(rack, rsm);
8764                 if (left) {
8765                         goto more;
8766                 }
8767                 /* Check for reneging */
8768                 rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8769                 if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
8770                         /*
8771                          * The peer has moved snd_una up to
8772                          * the edge of this send, i.e. one
8773                          * that it had previously acked. The only
8774                          * way that can be true if the peer threw
8775                          * away data (space issues) that it had
8776                          * previously sacked (else it would have
8777                          * given us snd_una up to (rsm->r_end).
8778                          * We need to undo the acked markings here.
8779                          *
8780                          * Note we have to look to make sure th_ack is
8781                          * our rsm->r_start in case we get an old ack
8782                          * where th_ack is behind snd_una.
8783                          */
8784                         rack_peer_reneges(rack, rsm, th_ack);
8785                 }
8786                 return;
8787         }
8788         if (rsm->r_flags & RACK_ACKED) {
8789                 /*
8790                  * It was acked on the scoreboard -- remove it from
8791                  * total for the part being cum-acked.
8792                  */
8793                 rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
8794         }
8795         /*
8796          * Clear the dup ack count for
8797          * the piece that remains.
8798          */
8799         rsm->r_dupack = 0;
8800         rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8801         if (rsm->r_rtr_bytes) {
8802                 /*
8803                  * It was retransmitted adjust the
8804                  * sack holes for what was acked.
8805                  */
8806                 int ack_am;
8807
8808                 ack_am = (th_ack - rsm->r_start);
8809                 if (ack_am >= rsm->r_rtr_bytes) {
8810                         rack->r_ctl.rc_holes_rxt -= ack_am;
8811                         rsm->r_rtr_bytes -= ack_am;
8812                 }
8813         }
8814         /*
8815          * Update where the piece starts and record
8816          * the time of send of highest cumack sent.
8817          */
8818         rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8819         rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
8820         /* Now we need to move our offset forward too */
8821         if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
8822                 /* Fix up the orig_m_len and possibly the mbuf offset */
8823                 rack_adjust_orig_mlen(rsm);
8824         }
8825         rsm->soff += (th_ack - rsm->r_start);
8826         rsm->r_start = th_ack;
8827         /* Now do we need to move the mbuf fwd too? */
8828         if (rsm->m) {
8829                 while (rsm->soff >= rsm->m->m_len) {
8830                         rsm->soff -= rsm->m->m_len;
8831                         rsm->m = rsm->m->m_next;
8832                         KASSERT((rsm->m != NULL),
8833                                 (" nrsm:%p hit at soff:%u null m",
8834                                  rsm, rsm->soff));
8835                 }
8836                 rsm->orig_m_len = rsm->m->m_len;
8837         }
8838         if (rack->app_limited_needs_set)
8839                 rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
8840 }
8841
8842 static void
8843 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
8844 {
8845         struct rack_sendmap *rsm;
8846         int sack_pass_fnd = 0;
8847
8848         if (rack->r_might_revert) {
8849                 /*
8850                  * Ok we have reordering, have not sent anything, we
8851                  * might want to revert the congestion state if nothing
8852                  * further has SACK_PASSED on it. Lets check.
8853                  *
8854                  * We also get here when we have DSACKs come in for
8855                  * all the data that we FR'd. Note that a rxt or tlp
8856                  * timer clears this from happening.
8857                  */
8858
8859                 TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
8860                         if (rsm->r_flags & RACK_SACK_PASSED) {
8861                                 sack_pass_fnd = 1;
8862                                 break;
8863                         }
8864                 }
8865                 if (sack_pass_fnd == 0) {
8866                         /*
8867                          * We went into recovery
8868                          * incorrectly due to reordering!
8869                          */
8870                         int orig_cwnd;
8871
8872                         rack->r_ent_rec_ns = 0;
8873                         orig_cwnd = tp->snd_cwnd;
8874                         tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec;
8875                         tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
8876                         tp->snd_recover = tp->snd_una;
8877                         rack_log_to_prr(rack, 14, orig_cwnd);
8878                         EXIT_RECOVERY(tp->t_flags);
8879                 }
8880                 rack->r_might_revert = 0;
8881         }
8882 }
8883
8884 #ifdef NETFLIX_EXP_DETECTION
8885 static void
8886 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
8887 {
8888         if ((rack->do_detection || tcp_force_detection) &&
8889             tcp_sack_to_ack_thresh &&
8890             tcp_sack_to_move_thresh &&
8891             ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
8892                 /*
8893                  * We have thresholds set to find
8894                  * possible attackers and disable sack.
8895                  * Check them.
8896                  */
8897                 uint64_t ackratio, moveratio, movetotal;
8898
8899                 /* Log detecting */
8900                 rack_log_sad(rack, 1);
8901                 ackratio = (uint64_t)(rack->r_ctl.sack_count);
8902                 ackratio *= (uint64_t)(1000);
8903                 if (rack->r_ctl.ack_count)
8904                         ackratio /= (uint64_t)(rack->r_ctl.ack_count);
8905                 else {
8906                         /* We really should not hit here */
8907                         ackratio = 1000;
8908                 }
8909                 if ((rack->sack_attack_disable == 0) &&
8910                     (ackratio > rack_highest_sack_thresh_seen))
8911                         rack_highest_sack_thresh_seen = (uint32_t)ackratio;
8912                 movetotal = rack->r_ctl.sack_moved_extra;
8913                 movetotal += rack->r_ctl.sack_noextra_move;
8914                 moveratio = rack->r_ctl.sack_moved_extra;
8915                 moveratio *= (uint64_t)1000;
8916                 if (movetotal)
8917                         moveratio /= movetotal;
8918                 else {
8919                         /* No moves, thats pretty good */
8920                         moveratio = 0;
8921                 }
8922                 if ((rack->sack_attack_disable == 0) &&
8923                     (moveratio > rack_highest_move_thresh_seen))
8924                         rack_highest_move_thresh_seen = (uint32_t)moveratio;
8925                 if (rack->sack_attack_disable == 0) {
8926                         if ((ackratio > tcp_sack_to_ack_thresh) &&
8927                             (moveratio > tcp_sack_to_move_thresh)) {
8928                                 /* Disable sack processing */
8929                                 rack->sack_attack_disable = 1;
8930                                 if (rack->r_rep_attack == 0) {
8931                                         rack->r_rep_attack = 1;
8932                                         counter_u64_add(rack_sack_attacks_detected, 1);
8933                                 }
8934                                 if (tcp_attack_on_turns_on_logging) {
8935                                         /*
8936                                          * Turn on logging, used for debugging
8937                                          * false positives.
8938                                          */
8939                                         rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
8940                                 }
8941                                 /* Clamp the cwnd at flight size */
8942                                 rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
8943                                 rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
8944                                 rack_log_sad(rack, 2);
8945                         }
8946                 } else {
8947                         /* We are sack-disabled check for false positives */
8948                         if ((ackratio <= tcp_restoral_thresh) ||
8949                             (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
8950                                 rack->sack_attack_disable = 0;
8951                                 rack_log_sad(rack, 3);
8952                                 /* Restart counting */
8953                                 rack->r_ctl.sack_count = 0;
8954                                 rack->r_ctl.sack_moved_extra = 0;
8955                                 rack->r_ctl.sack_noextra_move = 1;
8956                                 rack->r_ctl.ack_count = max(1,
8957                                       (bytes_this_ack / segsiz));
8958
8959                                 if (rack->r_rep_reverse == 0) {
8960                                         rack->r_rep_reverse = 1;
8961                                         counter_u64_add(rack_sack_attacks_reversed, 1);
8962                                 }
8963                                 /* Restore the cwnd */
8964                                 if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
8965                                         rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
8966                         }
8967                 }
8968         }
8969 }
8970 #endif
8971
8972 static void
8973 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
8974 {
8975
8976         uint32_t am;
8977
8978         if (SEQ_GT(end, start))
8979                 am = end - start;
8980         else
8981                 am = 0;
8982         /*
8983          * We keep track of how many DSACK blocks we get
8984          * after a recovery incident.
8985          */
8986         rack->r_ctl.dsack_byte_cnt += am;
8987         if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
8988             rack->r_ctl.retran_during_recovery &&
8989             (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
8990                 /*
8991                  * False recovery most likely culprit is reordering. If
8992                  * nothing else is missing we need to revert.
8993                  */
8994                 rack->r_might_revert = 1;
8995                 rack_handle_might_revert(rack->rc_tp, rack);
8996                 rack->r_might_revert = 0;
8997                 rack->r_ctl.retran_during_recovery = 0;
8998                 rack->r_ctl.dsack_byte_cnt = 0;
8999         }
9000 }
9001
9002 static void
9003 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9004 {
9005         /* Deal with changed and PRR here (in recovery only) */
9006         uint32_t pipe, snd_una;
9007
9008         rack->r_ctl.rc_prr_delivered += changed;
9009
9010         if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9011                 /*
9012                  * It is all outstanding, we are application limited
9013                  * and thus we don't need more room to send anything.
9014                  * Note we use tp->snd_una here and not th_ack because
9015                  * the data as yet not been cut from the sb.
9016                  */
9017                 rack->r_ctl.rc_prr_sndcnt = 0;
9018                 return;
9019         }
9020         /* Compute prr_sndcnt */
9021         if (SEQ_GT(tp->snd_una, th_ack)) {
9022                 snd_una = tp->snd_una;
9023         } else {
9024                 snd_una = th_ack;
9025         }
9026         pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9027         if (pipe > tp->snd_ssthresh) {
9028                 long sndcnt;
9029
9030                 sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9031                 if (rack->r_ctl.rc_prr_recovery_fs > 0)
9032                         sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9033                 else {
9034                         rack->r_ctl.rc_prr_sndcnt = 0;
9035                         rack_log_to_prr(rack, 9, 0);
9036                         sndcnt = 0;
9037                 }
9038                 sndcnt++;
9039                 if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9040                         sndcnt -= rack->r_ctl.rc_prr_out;
9041                 else
9042                         sndcnt = 0;
9043                 rack->r_ctl.rc_prr_sndcnt = sndcnt;
9044                 rack_log_to_prr(rack, 10, 0);
9045         } else {
9046                 uint32_t limit;
9047
9048                 if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9049                         limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9050                 else
9051                         limit = 0;
9052                 if (changed > limit)
9053                         limit = changed;
9054                 limit += ctf_fixed_maxseg(tp);
9055                 if (tp->snd_ssthresh > pipe) {
9056                         rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9057                         rack_log_to_prr(rack, 11, 0);
9058                 } else {
9059                         rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9060                         rack_log_to_prr(rack, 12, 0);
9061                 }
9062         }
9063 }
9064
9065 static void
9066 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9067 {
9068         uint32_t changed;
9069         struct tcp_rack *rack;
9070         struct rack_sendmap *rsm;
9071         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9072         register uint32_t th_ack;
9073         int32_t i, j, k, num_sack_blks = 0;
9074         uint32_t cts, acked, ack_point, sack_changed = 0;
9075         int loop_start = 0, moved_two = 0;
9076         uint32_t tsused;
9077
9078
9079         INP_WLOCK_ASSERT(tp->t_inpcb);
9080         if (th->th_flags & TH_RST) {
9081                 /* We don't log resets */
9082                 return;
9083         }
9084         rack = (struct tcp_rack *)tp->t_fb_ptr;
9085         cts = tcp_get_usecs(NULL);
9086         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9087         changed = 0;
9088         th_ack = th->th_ack;
9089         if (rack->sack_attack_disable == 0)
9090                 rack_do_decay(rack);
9091         if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9092                 /*
9093                  * You only get credit for
9094                  * MSS and greater (and you get extra
9095                  * credit for larger cum-ack moves).
9096                  */
9097                 int ac;
9098
9099                 ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9100                 rack->r_ctl.ack_count += ac;
9101                 counter_u64_add(rack_ack_total, ac);
9102         }
9103         if (rack->r_ctl.ack_count > 0xfff00000) {
9104                 /*
9105                  * reduce the number to keep us under
9106                  * a uint32_t.
9107                  */
9108                 rack->r_ctl.ack_count /= 2;
9109                 rack->r_ctl.sack_count /= 2;
9110         }
9111         if (SEQ_GT(th_ack, tp->snd_una)) {
9112                 rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9113                 tp->t_acktime = ticks;
9114         }
9115         if (rsm && SEQ_GT(th_ack, rsm->r_start))
9116                 changed = th_ack - rsm->r_start;
9117         if (changed) {
9118                 rack_process_to_cumack(tp, rack, th_ack, cts, to);
9119         }
9120         if ((to->to_flags & TOF_SACK) == 0) {
9121                 /* We are done nothing left and no sack. */
9122                 rack_handle_might_revert(tp, rack);
9123                 /*
9124                  * For cases where we struck a dup-ack
9125                  * with no SACK, add to the changes so
9126                  * PRR will work right.
9127                  */
9128                 if (dup_ack_struck && (changed == 0)) {
9129                         changed += ctf_fixed_maxseg(rack->rc_tp);
9130                 }
9131                 goto out;
9132         }
9133         /* Sack block processing */
9134         if (SEQ_GT(th_ack, tp->snd_una))
9135                 ack_point = th_ack;
9136         else
9137                 ack_point = tp->snd_una;
9138         for (i = 0; i < to->to_nsacks; i++) {
9139                 bcopy((to->to_sacks + i * TCPOLEN_SACK),
9140                       &sack, sizeof(sack));
9141                 sack.start = ntohl(sack.start);
9142                 sack.end = ntohl(sack.end);
9143                 if (SEQ_GT(sack.end, sack.start) &&
9144                     SEQ_GT(sack.start, ack_point) &&
9145                     SEQ_LT(sack.start, tp->snd_max) &&
9146                     SEQ_GT(sack.end, ack_point) &&
9147                     SEQ_LEQ(sack.end, tp->snd_max)) {
9148                         sack_blocks[num_sack_blks] = sack;
9149                         num_sack_blks++;
9150 #ifdef NETFLIX_STATS
9151                 } else if (SEQ_LEQ(sack.start, th_ack) &&
9152                            SEQ_LEQ(sack.end, th_ack)) {
9153                         /*
9154                          * Its a D-SACK block.
9155                          */
9156                         tcp_record_dsack(sack.start, sack.end);
9157 #endif
9158                         rack_note_dsack(rack, sack.start, sack.end);
9159                 }
9160         }
9161         /*
9162          * Sort the SACK blocks so we can update the rack scoreboard with
9163          * just one pass.
9164          */
9165         num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9166                                          num_sack_blks, th->th_ack);
9167         ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9168         if (num_sack_blks == 0) {
9169                 /* Nothing to sack (DSACKs?) */
9170                 goto out_with_totals;
9171         }
9172         if (num_sack_blks < 2) {
9173                 /* Only one, we don't need to sort */
9174                 goto do_sack_work;
9175         }
9176         /* Sort the sacks */
9177         for (i = 0; i < num_sack_blks; i++) {
9178                 for (j = i + 1; j < num_sack_blks; j++) {
9179                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9180                                 sack = sack_blocks[i];
9181                                 sack_blocks[i] = sack_blocks[j];
9182                                 sack_blocks[j] = sack;
9183                         }
9184                 }
9185         }
9186         /*
9187          * Now are any of the sack block ends the same (yes some
9188          * implementations send these)?
9189          */
9190 again:
9191         if (num_sack_blks == 0)
9192                 goto out_with_totals;
9193         if (num_sack_blks > 1) {
9194                 for (i = 0; i < num_sack_blks; i++) {
9195                         for (j = i + 1; j < num_sack_blks; j++) {
9196                                 if (sack_blocks[i].end == sack_blocks[j].end) {
9197                                         /*
9198                                          * Ok these two have the same end we
9199                                          * want the smallest end and then
9200                                          * throw away the larger and start
9201                                          * again.
9202                                          */
9203                                         if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9204                                                 /*
9205                                                  * The second block covers
9206                                                  * more area use that
9207                                                  */
9208                                                 sack_blocks[i].start = sack_blocks[j].start;
9209                                         }
9210                                         /*
9211                                          * Now collapse out the dup-sack and
9212                                          * lower the count
9213                                          */
9214                                         for (k = (j + 1); k < num_sack_blks; k++) {
9215                                                 sack_blocks[j].start = sack_blocks[k].start;
9216                                                 sack_blocks[j].end = sack_blocks[k].end;
9217                                                 j++;
9218                                         }
9219                                         num_sack_blks--;
9220                                         goto again;
9221                                 }
9222                         }
9223                 }
9224         }
9225 do_sack_work:
9226         /*
9227          * First lets look to see if
9228          * we have retransmitted and
9229          * can use the transmit next?
9230          */
9231         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9232         if (rsm &&
9233             SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9234             SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9235                 /*
9236                  * We probably did the FR and the next
9237                  * SACK in continues as we would expect.
9238                  */
9239                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9240                 if (acked) {
9241                         rack->r_wanted_output = 1;
9242                         changed += acked;
9243                         sack_changed += acked;
9244                 }
9245                 if (num_sack_blks == 1) {
9246                         /*
9247                          * This is what we would expect from
9248                          * a normal implementation to happen
9249                          * after we have retransmitted the FR,
9250                          * i.e the sack-filter pushes down
9251                          * to 1 block and the next to be retransmitted
9252                          * is the sequence in the sack block (has more
9253                          * are acked). Count this as ACK'd data to boost
9254                          * up the chances of recovering any false positives.
9255                          */
9256                         rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9257                         counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9258                         counter_u64_add(rack_express_sack, 1);
9259                         if (rack->r_ctl.ack_count > 0xfff00000) {
9260                                 /*
9261                                  * reduce the number to keep us under
9262                                  * a uint32_t.
9263                                  */
9264                                 rack->r_ctl.ack_count /= 2;
9265                                 rack->r_ctl.sack_count /= 2;
9266                         }
9267                         goto out_with_totals;
9268                 } else {
9269                         /*
9270                          * Start the loop through the
9271                          * rest of blocks, past the first block.
9272                          */
9273                         moved_two = 0;
9274                         loop_start = 1;
9275                 }
9276         }
9277         /* Its a sack of some sort */
9278         rack->r_ctl.sack_count++;
9279         if (rack->r_ctl.sack_count > 0xfff00000) {
9280                 /*
9281                  * reduce the number to keep us under
9282                  * a uint32_t.
9283                  */
9284                 rack->r_ctl.ack_count /= 2;
9285                 rack->r_ctl.sack_count /= 2;
9286         }
9287         counter_u64_add(rack_sack_total, 1);
9288         if (rack->sack_attack_disable) {
9289                 /* An attacker disablement is in place */
9290                 if (num_sack_blks > 1) {
9291                         rack->r_ctl.sack_count += (num_sack_blks - 1);
9292                         rack->r_ctl.sack_moved_extra++;
9293                         counter_u64_add(rack_move_some, 1);
9294                         if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9295                                 rack->r_ctl.sack_moved_extra /= 2;
9296                                 rack->r_ctl.sack_noextra_move /= 2;
9297                         }
9298                 }
9299                 goto out;
9300         }
9301         rsm = rack->r_ctl.rc_sacklast;
9302         for (i = loop_start; i < num_sack_blks; i++) {
9303                 acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9304                 if (acked) {
9305                         rack->r_wanted_output = 1;
9306                         changed += acked;
9307                         sack_changed += acked;
9308                 }
9309                 if (moved_two) {
9310                         /*
9311                          * If we did not get a SACK for at least a MSS and
9312                          * had to move at all, or if we moved more than our
9313                          * threshold, it counts against the "extra" move.
9314                          */
9315                         rack->r_ctl.sack_moved_extra += moved_two;
9316                         counter_u64_add(rack_move_some, 1);
9317                 } else {
9318                         /*
9319                          * else we did not have to move
9320                          * any more than we would expect.
9321                          */
9322                         rack->r_ctl.sack_noextra_move++;
9323                         counter_u64_add(rack_move_none, 1);
9324                 }
9325                 if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9326                         /*
9327                          * If the SACK was not a full MSS then
9328                          * we add to sack_count the number of
9329                          * MSS's (or possibly more than
9330                          * a MSS if its a TSO send) we had to skip by.
9331                          */
9332                         rack->r_ctl.sack_count += moved_two;
9333                         counter_u64_add(rack_sack_total, moved_two);
9334                 }
9335                 /*
9336                  * Now we need to setup for the next
9337                  * round. First we make sure we won't
9338                  * exceed the size of our uint32_t on
9339                  * the various counts, and then clear out
9340                  * moved_two.
9341                  */
9342                 if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9343                     (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9344                         rack->r_ctl.sack_moved_extra /= 2;
9345                         rack->r_ctl.sack_noextra_move /= 2;
9346                 }
9347                 if (rack->r_ctl.sack_count > 0xfff00000) {
9348                         rack->r_ctl.ack_count /= 2;
9349                         rack->r_ctl.sack_count /= 2;
9350                 }
9351                 moved_two = 0;
9352         }
9353 out_with_totals:
9354         if (num_sack_blks > 1) {
9355                 /*
9356                  * You get an extra stroke if
9357                  * you have more than one sack-blk, this
9358                  * could be where we are skipping forward
9359                  * and the sack-filter is still working, or
9360                  * it could be an attacker constantly
9361                  * moving us.
9362                  */
9363                 rack->r_ctl.sack_moved_extra++;
9364                 counter_u64_add(rack_move_some, 1);
9365         }
9366 out:
9367 #ifdef NETFLIX_EXP_DETECTION
9368         rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9369 #endif
9370         if (changed) {
9371                 /* Something changed cancel the rack timer */
9372                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9373         }
9374         tsused = tcp_get_usecs(NULL);
9375         rsm = tcp_rack_output(tp, rack, tsused);
9376         if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9377             rsm) {
9378                 /* Enter recovery */
9379                 rack->r_ctl.rc_rsm_start = rsm->r_start;
9380                 rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
9381                 rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
9382                 entered_recovery = 1;
9383                 rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9384                 /*
9385                  * When we enter recovery we need to assure we send
9386                  * one packet.
9387                  */
9388                 if (rack->rack_no_prr == 0) {
9389                         rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9390                         rack_log_to_prr(rack, 8, 0);
9391                 }
9392                 rack->r_timer_override = 1;
9393                 rack->r_early = 0;
9394                 rack->r_ctl.rc_agg_early = 0;
9395         } else if (IN_FASTRECOVERY(tp->t_flags) &&
9396                    rsm &&
9397                    (rack->r_rr_config == 3)) {
9398                 /*
9399                  * Assure we can output and we get no
9400                  * remembered pace time except the retransmit.
9401                  */
9402                 rack->r_timer_override = 1;
9403                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9404                 rack->r_ctl.rc_resend = rsm;
9405         }
9406         if (IN_FASTRECOVERY(tp->t_flags) &&
9407             (rack->rack_no_prr == 0) &&
9408             (entered_recovery == 0)) {
9409                 rack_update_prr(tp, rack, changed, th_ack);
9410                 if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9411                      ((rack->rc_inp->inp_in_hpts == 0) &&
9412                       ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9413                         /*
9414                          * If you are pacing output you don't want
9415                          * to override.
9416                          */
9417                         rack->r_early = 0;
9418                         rack->r_ctl.rc_agg_early = 0;
9419                         rack->r_timer_override = 1;
9420                 }
9421         }
9422 }
9423
9424 static void
9425 rack_strike_dupack(struct tcp_rack *rack)
9426 {
9427         struct rack_sendmap *rsm;
9428
9429         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9430         while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9431                 rsm = TAILQ_NEXT(rsm, r_tnext);
9432         }
9433         if (rsm && (rsm->r_dupack < 0xff)) {
9434                 rsm->r_dupack++;
9435                 if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9436                         struct timeval tv;
9437                         uint32_t cts;
9438                         /*
9439                          * Here we see if we need to retransmit. For
9440                          * a SACK type connection if enough time has passed
9441                          * we will get a return of the rsm. For a non-sack
9442                          * connection we will get the rsm returned if the
9443                          * dupack value is 3 or more.
9444                          */
9445                         cts = tcp_get_usecs(&tv);
9446                         rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9447                         if (rack->r_ctl.rc_resend != NULL) {
9448                                 if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9449                                         rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9450                                                          rack->rc_tp->snd_una);
9451                                 }
9452                                 rack->r_wanted_output = 1;
9453                                 rack->r_timer_override = 1;
9454                                 rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9455                         }
9456                 } else {
9457                         rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9458                 }
9459         }
9460 }
9461
9462 static void
9463 rack_check_bottom_drag(struct tcpcb *tp,
9464                        struct tcp_rack *rack,
9465                        struct socket *so, int32_t acked)
9466 {
9467         uint32_t segsiz, minseg;
9468
9469         segsiz = ctf_fixed_maxseg(tp);
9470         minseg = segsiz;
9471
9472         if (tp->snd_max == tp->snd_una) {
9473                 /*
9474                  * We are doing dynamic pacing and we are way
9475                  * under. Basically everything got acked while
9476                  * we were still waiting on the pacer to expire.
9477                  *
9478                  * This means we need to boost the b/w in
9479                  * addition to any earlier boosting of
9480                  * the multipler.
9481                  */
9482                 rack->rc_dragged_bottom = 1;
9483                 rack_validate_multipliers_at_or_above100(rack);
9484                 /*
9485                  * Lets use the segment bytes acked plus
9486                  * the lowest RTT seen as the basis to
9487                  * form a b/w estimate. This will be off
9488                  * due to the fact that the true estimate
9489                  * should be around 1/2 the time of the RTT
9490                  * but we can settle for that.
9491                  */
9492                 if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9493                     acked) {
9494                         uint64_t bw, calc_bw, rtt;
9495
9496                         rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9497                         if (rtt == 0) {
9498                                 /* no us sample is there a ms one? */
9499                                 if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9500                                         rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9501                                 } else {
9502                                         goto no_measurement;
9503                                 }
9504                         }
9505                         bw = acked;
9506                         calc_bw = bw * 1000000;
9507                         calc_bw /= rtt;
9508                         if (rack->r_ctl.last_max_bw &&
9509                             (rack->r_ctl.last_max_bw < calc_bw)) {
9510                                 /*
9511                                  * If we have a last calculated max bw
9512                                  * enforce it.
9513                                  */
9514                                 calc_bw = rack->r_ctl.last_max_bw;
9515                         }
9516                         /* now plop it in */
9517                         if (rack->rc_gp_filled == 0) {
9518                                 if (calc_bw > ONE_POINT_TWO_MEG) {
9519                                         /*
9520                                          * If we have no measurement
9521                                          * don't let us set in more than
9522                                          * 1.2Mbps. If we are still too
9523                                          * low after pacing with this we
9524                                          * will hopefully have a max b/w
9525                                          * available to sanity check things.
9526                                          */
9527                                         calc_bw = ONE_POINT_TWO_MEG;
9528                                 }
9529                                 rack->r_ctl.rc_rtt_diff = 0;
9530                                 rack->r_ctl.gp_bw = calc_bw;
9531                                 rack->rc_gp_filled = 1;
9532                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9533                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9534                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9535                         } else if (calc_bw > rack->r_ctl.gp_bw) {
9536                                 rack->r_ctl.rc_rtt_diff = 0;
9537                                 if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9538                                         rack->r_ctl.num_measurements = RACK_REQ_AVG;
9539                                 rack->r_ctl.gp_bw = calc_bw;
9540                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9541                         } else
9542                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9543                         if ((rack->gp_ready == 0) &&
9544                             (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9545                                 /* We have enough measurements now */
9546                                 rack->gp_ready = 1;
9547                                 rack_set_cc_pacing(rack);
9548                                 if (rack->defer_options)
9549                                         rack_apply_deferred_options(rack);
9550                         }
9551                         /*
9552                          * For acks over 1mss we do a extra boost to simulate
9553                          * where we would get 2 acks (we want 110 for the mul).
9554                          */
9555                         if (acked > segsiz)
9556                                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9557                 } else {
9558                         /*
9559                          * zero rtt possibly?, settle for just an old increase.
9560                          */
9561 no_measurement:
9562                         rack_increase_bw_mul(rack, -1, 0, 0, 1);
9563                 }
9564         } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9565                    (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9566                                                minseg)) &&
9567                    (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9568                    (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9569                    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9570                     (segsiz * rack_req_segs))) {
9571                 /*
9572                  * We are doing dynamic GP pacing and
9573                  * we have everything except 1MSS or less
9574                  * bytes left out. We are still pacing away.
9575                  * And there is data that could be sent, This
9576                  * means we are inserting delayed ack time in
9577                  * our measurements because we are pacing too slow.
9578                  */
9579                 rack_validate_multipliers_at_or_above100(rack);
9580                 rack->rc_dragged_bottom = 1;
9581                 rack_increase_bw_mul(rack, -1, 0, 0, 1);
9582         }
9583 }
9584
9585
9586
9587 static void
9588 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9589 {
9590         /*
9591          * The fast output path is enabled and we
9592          * have moved the cumack forward. Lets see if
9593          * we can expand forward the fast path length by
9594          * that amount. What we would ideally like to
9595          * do is increase the number of bytes in the
9596          * fast path block (left_to_send) by the
9597          * acked amount. However we have to gate that
9598          * by two factors:
9599          * 1) The amount outstanding and the rwnd of the peer
9600          *    (i.e. we don't want to exceed the rwnd of the peer).
9601          *    <and>
9602          * 2) The amount of data left in the socket buffer (i.e.
9603          *    we can't send beyond what is in the buffer).
9604          *
9605          * Note that this does not take into account any increase
9606          * in the cwnd. We will only extend the fast path by
9607          * what was acked.
9608          */
9609         uint32_t new_total, gating_val;
9610
9611         new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9612         gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9613                          (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9614         if (new_total <= gating_val) {
9615                 /* We can increase left_to_send by the acked amount */
9616                 counter_u64_add(rack_extended_rfo, 1);
9617                 rack->r_ctl.fsb.left_to_send = new_total;
9618                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9619                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
9620                          rack, rack->r_ctl.fsb.left_to_send,
9621                          sbavail(&rack->rc_inp->inp_socket->so_snd),
9622                          (tp->snd_max - tp->snd_una)));
9623
9624         }
9625 }
9626
9627 static void
9628 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9629 {
9630         /*
9631          * Here any sendmap entry that points to the
9632          * beginning mbuf must be adjusted to the correct
9633          * offset. This must be called with:
9634          * 1) The socket buffer locked
9635          * 2) snd_una adjusted to its new postion.
9636          *
9637          * Note that (2) implies rack_ack_received has also
9638          * been called.
9639          *
9640          * We grab the first mbuf in the socket buffer and
9641          * then go through the front of the sendmap, recalculating
9642          * the stored offset for any sendmap entry that has
9643          * that mbuf. We must use the sb functions to do this
9644          * since its possible an add was done has well as
9645          * the subtraction we may have just completed. This should
9646          * not be a penalty though, since we just referenced the sb
9647          * to go in and trim off the mbufs that we freed (of course
9648          * there will be a penalty for the sendmap references though).
9649          */
9650         struct mbuf *m;
9651         struct rack_sendmap *rsm;
9652
9653         SOCKBUF_LOCK_ASSERT(sb);
9654         m = sb->sb_mb;
9655         rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9656         if ((rsm == NULL) || (m == NULL)) {
9657                 /* Nothing outstanding */
9658                 return;
9659         }
9660         while (rsm->m && (rsm->m == m)) {
9661                 /* one to adjust */
9662 #ifdef INVARIANTS
9663                 struct mbuf *tm;
9664                 uint32_t soff;
9665
9666                 tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
9667                 if (rsm->orig_m_len != m->m_len) {
9668                         rack_adjust_orig_mlen(rsm);
9669                 }
9670                 if (rsm->soff != soff) {
9671                         /*
9672                          * This is not a fatal error, we anticipate it
9673                          * might happen (the else code), so we count it here
9674                          * so that under invariant we can see that it really
9675                          * does happen.
9676                          */
9677                         counter_u64_add(rack_adjust_map_bw, 1);
9678                 }
9679                 rsm->m = tm;
9680                 rsm->soff = soff;
9681                 if (tm)
9682                         rsm->orig_m_len = rsm->m->m_len;
9683                 else
9684                         rsm->orig_m_len = 0;
9685 #else
9686                 rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
9687                 if (rsm->m)
9688                         rsm->orig_m_len = rsm->m->m_len;
9689                 else
9690                         rsm->orig_m_len = 0;
9691 #endif
9692                 rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
9693                               rsm);
9694                 if (rsm == NULL)
9695                         break;
9696         }
9697 }
9698
9699 /*
9700  * Return value of 1, we do not need to call rack_process_data().
9701  * return value of 0, rack_process_data can be called.
9702  * For ret_val if its 0 the TCP is locked, if its non-zero
9703  * its unlocked and probably unsafe to touch the TCB.
9704  */
9705 static int
9706 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
9707     struct tcpcb *tp, struct tcpopt *to,
9708     uint32_t tiwin, int32_t tlen,
9709     int32_t * ofia, int32_t thflags, int32_t *ret_val)
9710 {
9711         int32_t ourfinisacked = 0;
9712         int32_t nsegs, acked_amount;
9713         int32_t acked;
9714         struct mbuf *mfree;
9715         struct tcp_rack *rack;
9716         int32_t under_pacing = 0;
9717         int32_t recovery = 0;
9718
9719         rack = (struct tcp_rack *)tp->t_fb_ptr;
9720         if (SEQ_GT(th->th_ack, tp->snd_max)) {
9721                 __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
9722                                       &rack->r_ctl.challenge_ack_ts,
9723                                       &rack->r_ctl.challenge_ack_cnt);
9724                 rack->r_wanted_output = 1;
9725                 return (1);
9726         }
9727         if (rack->gp_ready &&
9728             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
9729                 under_pacing = 1;
9730         }
9731         if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
9732                 int in_rec, dup_ack_struck = 0;
9733
9734                 in_rec = IN_FASTRECOVERY(tp->t_flags);
9735                 if (rack->rc_in_persist) {
9736                         tp->t_rxtshift = 0;
9737                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9738                                       rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9739                 }
9740                 if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) {
9741                         rack_strike_dupack(rack);
9742                         dup_ack_struck = 1;
9743                 }
9744                 rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
9745         }
9746         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
9747                 /*
9748                  * Old ack, behind (or duplicate to) the last one rcv'd
9749                  * Note: We mark reordering is occuring if its
9750                  * less than and we have not closed our window.
9751                  */
9752                 if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
9753                         counter_u64_add(rack_reorder_seen, 1);
9754                         rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9755                 }
9756                 return (0);
9757         }
9758         /*
9759          * If we reach this point, ACK is not a duplicate, i.e., it ACKs
9760          * something we sent.
9761          */
9762         if (tp->t_flags & TF_NEEDSYN) {
9763                 /*
9764                  * T/TCP: Connection was half-synchronized, and our SYN has
9765                  * been ACK'd (so connection is now fully synchronized).  Go
9766                  * to non-starred state, increment snd_una for ACK of SYN,
9767                  * and check if we can do window scaling.
9768                  */
9769                 tp->t_flags &= ~TF_NEEDSYN;
9770                 tp->snd_una++;
9771                 /* Do window scaling? */
9772                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
9773                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
9774                         tp->rcv_scale = tp->request_r_scale;
9775                         /* Send window already scaled. */
9776                 }
9777         }
9778         nsegs = max(1, m->m_pkthdr.lro_nsegs);
9779         INP_WLOCK_ASSERT(tp->t_inpcb);
9780
9781         acked = BYTES_THIS_ACK(tp, th);
9782         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
9783         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
9784         /*
9785          * If we just performed our first retransmit, and the ACK arrives
9786          * within our recovery window, then it was a mistake to do the
9787          * retransmit in the first place.  Recover our original cwnd and
9788          * ssthresh, and proceed to transmit where we left off.
9789          */
9790         if ((tp->t_flags & TF_PREVVALID) &&
9791             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
9792                 tp->t_flags &= ~TF_PREVVALID;
9793                 if (tp->t_rxtshift == 1 &&
9794                     (int)(ticks - tp->t_badrxtwin) < 0)
9795                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
9796         }
9797         if (acked) {
9798                 /* assure we are not backed off */
9799                 tp->t_rxtshift = 0;
9800                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9801                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9802                 rack->rc_tlp_in_progress = 0;
9803                 rack->r_ctl.rc_tlp_cnt_out = 0;
9804                 /*
9805                  * If it is the RXT timer we want to
9806                  * stop it, so we can restart a TLP.
9807                  */
9808                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
9809                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9810 #ifdef NETFLIX_HTTP_LOGGING
9811                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
9812 #endif
9813         }
9814         /*
9815          * If we have a timestamp reply, update smoothed round trip time. If
9816          * no timestamp is present but transmit timer is running and timed
9817          * sequence number was acked, update smoothed round trip time. Since
9818          * we now have an rtt measurement, cancel the timer backoff (cf.,
9819          * Phil Karn's retransmit alg.). Recompute the initial retransmit
9820          * timer.
9821          *
9822          * Some boxes send broken timestamp replies during the SYN+ACK
9823          * phase, ignore timestamps of 0 or we could calculate a huge RTT
9824          * and blow up the retransmit timer.
9825          */
9826         /*
9827          * If all outstanding data is acked, stop retransmit timer and
9828          * remember to restart (more output or persist). If there is more
9829          * data to be acked, restart retransmit timer, using current
9830          * (possibly backed-off) value.
9831          */
9832         if (acked == 0) {
9833                 if (ofia)
9834                         *ofia = ourfinisacked;
9835                 return (0);
9836         }
9837         if (IN_RECOVERY(tp->t_flags)) {
9838                 if (SEQ_LT(th->th_ack, tp->snd_recover) &&
9839                     (SEQ_LT(th->th_ack, tp->snd_max))) {
9840                         tcp_rack_partialack(tp);
9841                 } else {
9842                         rack_post_recovery(tp, th->th_ack);
9843                         recovery = 1;
9844                 }
9845         }
9846         /*
9847          * Let the congestion control algorithm update congestion control
9848          * related information. This typically means increasing the
9849          * congestion window.
9850          */
9851         rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
9852         SOCKBUF_LOCK(&so->so_snd);
9853         acked_amount = min(acked, (int)sbavail(&so->so_snd));
9854         tp->snd_wnd -= acked_amount;
9855         mfree = sbcut_locked(&so->so_snd, acked_amount);
9856         if ((sbused(&so->so_snd) == 0) &&
9857             (acked > acked_amount) &&
9858             (tp->t_state >= TCPS_FIN_WAIT_1) &&
9859             (tp->t_flags & TF_SENTFIN)) {
9860                 /*
9861                  * We must be sure our fin
9862                  * was sent and acked (we can be
9863                  * in FIN_WAIT_1 without having
9864                  * sent the fin).
9865                  */
9866                 ourfinisacked = 1;
9867         }
9868         tp->snd_una = th->th_ack;
9869         if (acked_amount && sbavail(&so->so_snd))
9870                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
9871         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
9872         /* NB: sowwakeup_locked() does an implicit unlock. */
9873         sowwakeup_locked(so);
9874         m_freem(mfree);
9875         if (SEQ_GT(tp->snd_una, tp->snd_recover))
9876                 tp->snd_recover = tp->snd_una;
9877
9878         if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
9879                 tp->snd_nxt = tp->snd_una;
9880         }
9881         if (under_pacing &&
9882             (rack->use_fixed_rate == 0) &&
9883             (rack->in_probe_rtt == 0) &&
9884             rack->rc_gp_dyn_mul &&
9885             rack->rc_always_pace) {
9886                 /* Check if we are dragging bottom */
9887                 rack_check_bottom_drag(tp, rack, so, acked);
9888         }
9889         if (tp->snd_una == tp->snd_max) {
9890                 /* Nothing left outstanding */
9891                 tp->t_flags &= ~TF_PREVVALID;
9892                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
9893                 rack->r_ctl.retran_during_recovery = 0;
9894                 rack->r_ctl.dsack_byte_cnt = 0;
9895                 if (rack->r_ctl.rc_went_idle_time == 0)
9896                         rack->r_ctl.rc_went_idle_time = 1;
9897                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
9898                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
9899                         tp->t_acktime = 0;
9900                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9901                 /* Set need output so persist might get set */
9902                 rack->r_wanted_output = 1;
9903                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
9904                 if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
9905                     (sbavail(&so->so_snd) == 0) &&
9906                     (tp->t_flags2 & TF2_DROP_AF_DATA)) {
9907                         /*
9908                          * The socket was gone and the
9909                          * peer sent data (now or in the past), time to
9910                          * reset him.
9911                          */
9912                         *ret_val = 1;
9913                         /* tcp_close will kill the inp pre-log the Reset */
9914                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
9915                         tp = tcp_close(tp);
9916                         ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
9917                         return (1);
9918                 }
9919         }
9920         if (ofia)
9921                 *ofia = ourfinisacked;
9922         return (0);
9923 }
9924
9925 static void
9926 rack_collapsed_window(struct tcp_rack *rack)
9927 {
9928         /*
9929          * Now we must walk the
9930          * send map and divide the
9931          * ones left stranded. These
9932          * guys can't cause us to abort
9933          * the connection and are really
9934          * "unsent". However if a buggy
9935          * client actually did keep some
9936          * of the data i.e. collapsed the win
9937          * and refused to ack and then opened
9938          * the win and acked that data. We would
9939          * get into an ack war, the simplier
9940          * method then of just pretending we
9941          * did not send those segments something
9942          * won't work.
9943          */
9944         struct rack_sendmap *rsm, *nrsm, fe, *insret;
9945         tcp_seq max_seq;
9946
9947         max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
9948         memset(&fe, 0, sizeof(fe));
9949         fe.r_start = max_seq;
9950         /* Find the first seq past or at maxseq */
9951         rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
9952         if (rsm == NULL) {
9953                 /* Nothing to do strange */
9954                 rack->rc_has_collapsed = 0;
9955                 return;
9956         }
9957         /*
9958          * Now do we need to split at
9959          * the collapse point?
9960          */
9961         if (SEQ_GT(max_seq, rsm->r_start)) {
9962                 nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9963                 if (nrsm == NULL) {
9964                         /* We can't get a rsm, mark all? */
9965                         nrsm = rsm;
9966                         goto no_split;
9967                 }
9968                 /* Clone it */
9969                 rack_clone_rsm(rack, nrsm, rsm, max_seq);
9970                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
9971 #ifdef INVARIANTS
9972                 if (insret != NULL) {
9973                         panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
9974                               nrsm, insret, rack, rsm);
9975                 }
9976 #endif
9977                 rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
9978                 if (rsm->r_in_tmap) {
9979                         TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9980                         nrsm->r_in_tmap = 1;
9981                 }
9982                 /*
9983                  * Set in the new RSM as the
9984                  * collapsed starting point
9985                  */
9986                 rsm = nrsm;
9987         }
9988 no_split:
9989         counter_u64_add(rack_collapsed_win, 1);
9990         RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
9991                 nrsm->r_flags |= RACK_RWND_COLLAPSED;
9992         }
9993         rack->rc_has_collapsed = 1;
9994 }
9995
9996 static void
9997 rack_un_collapse_window(struct tcp_rack *rack)
9998 {
9999         struct rack_sendmap *rsm;
10000
10001         RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
10002                 if (rsm->r_flags & RACK_RWND_COLLAPSED)
10003                         rsm->r_flags &= ~RACK_RWND_COLLAPSED;
10004                 else
10005                         break;
10006         }
10007         rack->rc_has_collapsed = 0;
10008 }
10009
10010 static void
10011 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10012                         int32_t tlen, int32_t tfo_syn)
10013 {
10014         if (DELAY_ACK(tp, tlen) || tfo_syn) {
10015                 if (rack->rc_dack_mode &&
10016                     (tlen > 500) &&
10017                     (rack->rc_dack_toggle == 1)) {
10018                         goto no_delayed_ack;
10019                 }
10020                 rack_timer_cancel(tp, rack,
10021                                   rack->r_ctl.rc_rcvtime, __LINE__);
10022                 tp->t_flags |= TF_DELACK;
10023         } else {
10024 no_delayed_ack:
10025                 rack->r_wanted_output = 1;
10026                 tp->t_flags |= TF_ACKNOW;
10027                 if (rack->rc_dack_mode) {
10028                         if (tp->t_flags & TF_DELACK)
10029                                 rack->rc_dack_toggle = 1;
10030                         else
10031                                 rack->rc_dack_toggle = 0;
10032                 }
10033         }
10034 }
10035
10036 static void
10037 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10038 {
10039         /*
10040          * If fast output is in progress, lets validate that
10041          * the new window did not shrink on us and make it
10042          * so fast output should end.
10043          */
10044         if (rack->r_fast_output) {
10045                 uint32_t out;
10046
10047                 /*
10048                  * Calculate what we will send if left as is
10049                  * and compare that to our send window.
10050                  */
10051                 out = ctf_outstanding(tp);
10052                 if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10053                         /* ok we have an issue */
10054                         if (out >= tp->snd_wnd) {
10055                                 /* Turn off fast output the window is met or collapsed */
10056                                 rack->r_fast_output = 0;
10057                         } else {
10058                                 /* we have some room left */
10059                                 rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10060                                 if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10061                                         /* If not at least 1 full segment never mind */
10062                                         rack->r_fast_output = 0;
10063                                 }
10064                         }
10065                 }
10066         }
10067 }
10068
10069
10070 /*
10071  * Return value of 1, the TCB is unlocked and most
10072  * likely gone, return value of 0, the TCP is still
10073  * locked.
10074  */
10075 static int
10076 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10077     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10078     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10079 {
10080         /*
10081          * Update window information. Don't look at window if no ACK: TAC's
10082          * send garbage on first SYN.
10083          */
10084         int32_t nsegs;
10085         int32_t tfo_syn;
10086         struct tcp_rack *rack;
10087
10088         rack = (struct tcp_rack *)tp->t_fb_ptr;
10089         INP_WLOCK_ASSERT(tp->t_inpcb);
10090         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10091         if ((thflags & TH_ACK) &&
10092             (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10093             (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10094             (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10095                 /* keep track of pure window updates */
10096                 if (tlen == 0 &&
10097                     tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10098                         KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10099                 tp->snd_wnd = tiwin;
10100                 rack_validate_fo_sendwin_up(tp, rack);
10101                 tp->snd_wl1 = th->th_seq;
10102                 tp->snd_wl2 = th->th_ack;
10103                 if (tp->snd_wnd > tp->max_sndwnd)
10104                         tp->max_sndwnd = tp->snd_wnd;
10105                 rack->r_wanted_output = 1;
10106         } else if (thflags & TH_ACK) {
10107                 if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10108                         tp->snd_wnd = tiwin;
10109                         rack_validate_fo_sendwin_up(tp, rack);
10110                         tp->snd_wl1 = th->th_seq;
10111                         tp->snd_wl2 = th->th_ack;
10112                 }
10113         }
10114         if (tp->snd_wnd < ctf_outstanding(tp))
10115                 /* The peer collapsed the window */
10116                 rack_collapsed_window(rack);
10117         else if (rack->rc_has_collapsed)
10118                 rack_un_collapse_window(rack);
10119         /* Was persist timer active and now we have window space? */
10120         if ((rack->rc_in_persist != 0) &&
10121             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10122                                 rack->r_ctl.rc_pace_min_segs))) {
10123                 rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10124                 tp->snd_nxt = tp->snd_max;
10125                 /* Make sure we output to start the timer */
10126                 rack->r_wanted_output = 1;
10127         }
10128         /* Do we enter persists? */
10129         if ((rack->rc_in_persist == 0) &&
10130             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10131             TCPS_HAVEESTABLISHED(tp->t_state) &&
10132             (tp->snd_max == tp->snd_una) &&
10133             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10134             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10135                 /*
10136                  * Here the rwnd is less than
10137                  * the pacing size, we are established,
10138                  * nothing is outstanding, and there is
10139                  * data to send. Enter persists.
10140                  */
10141                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10142         }
10143         if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10144                 m_freem(m);
10145                 return (0);
10146         }
10147         /*
10148          * don't process the URG bit, ignore them drag
10149          * along the up.
10150          */
10151         tp->rcv_up = tp->rcv_nxt;
10152         INP_WLOCK_ASSERT(tp->t_inpcb);
10153
10154         /*
10155          * Process the segment text, merging it into the TCP sequencing
10156          * queue, and arranging for acknowledgment of receipt if necessary.
10157          * This process logically involves adjusting tp->rcv_wnd as data is
10158          * presented to the user (this happens in tcp_usrreq.c, case
10159          * PRU_RCVD).  If a FIN has already been received on this connection
10160          * then we just ignore the text.
10161          */
10162         tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10163                    IS_FASTOPEN(tp->t_flags));
10164         if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10165             TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10166                 tcp_seq save_start = th->th_seq;
10167                 tcp_seq save_rnxt  = tp->rcv_nxt;
10168                 int     save_tlen  = tlen;
10169
10170                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10171                 /*
10172                  * Insert segment which includes th into TCP reassembly
10173                  * queue with control block tp.  Set thflags to whether
10174                  * reassembly now includes a segment with FIN.  This handles
10175                  * the common case inline (segment is the next to be
10176                  * received on an established connection, and the queue is
10177                  * empty), avoiding linkage into and removal from the queue
10178                  * and repetition of various conversions. Set DELACK for
10179                  * segments received in order, but ack immediately when
10180                  * segments are out of order (so fast retransmit can work).
10181                  */
10182                 if (th->th_seq == tp->rcv_nxt &&
10183                     SEGQ_EMPTY(tp) &&
10184                     (TCPS_HAVEESTABLISHED(tp->t_state) ||
10185                     tfo_syn)) {
10186 #ifdef NETFLIX_SB_LIMITS
10187                         u_int mcnt, appended;
10188
10189                         if (so->so_rcv.sb_shlim) {
10190                                 mcnt = m_memcnt(m);
10191                                 appended = 0;
10192                                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10193                                     CFO_NOSLEEP, NULL) == false) {
10194                                         counter_u64_add(tcp_sb_shlim_fails, 1);
10195                                         m_freem(m);
10196                                         return (0);
10197                                 }
10198                         }
10199 #endif
10200                         rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10201                         tp->rcv_nxt += tlen;
10202                         if (tlen &&
10203                             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10204                             (tp->t_fbyte_in == 0)) {
10205                                 tp->t_fbyte_in = ticks;
10206                                 if (tp->t_fbyte_in == 0)
10207                                         tp->t_fbyte_in = 1;
10208                                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10209                                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10210                         }
10211                         thflags = th->th_flags & TH_FIN;
10212                         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10213                         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10214                         SOCKBUF_LOCK(&so->so_rcv);
10215                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10216                                 m_freem(m);
10217                         } else
10218 #ifdef NETFLIX_SB_LIMITS
10219                                 appended =
10220 #endif
10221                                         sbappendstream_locked(&so->so_rcv, m, 0);
10222
10223                         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10224                         tp->t_flags |= TF_WAKESOR;
10225 #ifdef NETFLIX_SB_LIMITS
10226                         if (so->so_rcv.sb_shlim && appended != mcnt)
10227                                 counter_fo_release(so->so_rcv.sb_shlim,
10228                                     mcnt - appended);
10229 #endif
10230                 } else {
10231                         /*
10232                          * XXX: Due to the header drop above "th" is
10233                          * theoretically invalid by now.  Fortunately
10234                          * m_adj() doesn't actually frees any mbufs when
10235                          * trimming from the head.
10236                          */
10237                         tcp_seq temp = save_start;
10238
10239                         thflags = tcp_reass(tp, th, &temp, &tlen, m);
10240                         tp->t_flags |= TF_ACKNOW;
10241
10242                 }
10243                 if ((tp->t_flags & TF_SACK_PERMIT) &&
10244                     (save_tlen > 0) &&
10245                     TCPS_HAVEESTABLISHED(tp->t_state)) {
10246                         if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10247                                 /*
10248                                  * DSACK actually handled in the fastpath
10249                                  * above.
10250                                  */
10251                                 RACK_OPTS_INC(tcp_sack_path_1);
10252                                 tcp_update_sack_list(tp, save_start,
10253                                     save_start + save_tlen);
10254                         } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10255                                 if ((tp->rcv_numsacks >= 1) &&
10256                                     (tp->sackblks[0].end == save_start)) {
10257                                         /*
10258                                          * Partial overlap, recorded at todrop
10259                                          * above.
10260                                          */
10261                                         RACK_OPTS_INC(tcp_sack_path_2a);
10262                                         tcp_update_sack_list(tp,
10263                                             tp->sackblks[0].start,
10264                                             tp->sackblks[0].end);
10265                                 } else {
10266                                         RACK_OPTS_INC(tcp_sack_path_2b);
10267                                         tcp_update_dsack_list(tp, save_start,
10268                                             save_start + save_tlen);
10269                                 }
10270                         } else if (tlen >= save_tlen) {
10271                                 /* Update of sackblks. */
10272                                 RACK_OPTS_INC(tcp_sack_path_3);
10273                                 tcp_update_dsack_list(tp, save_start,
10274                                     save_start + save_tlen);
10275                         } else if (tlen > 0) {
10276                                 RACK_OPTS_INC(tcp_sack_path_4);
10277                                 tcp_update_dsack_list(tp, save_start,
10278                                     save_start + tlen);
10279                         }
10280                 }
10281                 tcp_handle_wakeup(tp, so);
10282         } else {
10283                 m_freem(m);
10284                 thflags &= ~TH_FIN;
10285         }
10286
10287         /*
10288          * If FIN is received ACK the FIN and let the user know that the
10289          * connection is closing.
10290          */
10291         if (thflags & TH_FIN) {
10292                 if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10293                         /* The socket upcall is handled by socantrcvmore. */
10294                         socantrcvmore(so);
10295                         /*
10296                          * If connection is half-synchronized (ie NEEDSYN
10297                          * flag on) then delay ACK, so it may be piggybacked
10298                          * when SYN is sent. Otherwise, since we received a
10299                          * FIN then no more input can be expected, send ACK
10300                          * now.
10301                          */
10302                         if (tp->t_flags & TF_NEEDSYN) {
10303                                 rack_timer_cancel(tp, rack,
10304                                     rack->r_ctl.rc_rcvtime, __LINE__);
10305                                 tp->t_flags |= TF_DELACK;
10306                         } else {
10307                                 tp->t_flags |= TF_ACKNOW;
10308                         }
10309                         tp->rcv_nxt++;
10310                 }
10311                 switch (tp->t_state) {
10312                         /*
10313                          * In SYN_RECEIVED and ESTABLISHED STATES enter the
10314                          * CLOSE_WAIT state.
10315                          */
10316                 case TCPS_SYN_RECEIVED:
10317                         tp->t_starttime = ticks;
10318                         /* FALLTHROUGH */
10319                 case TCPS_ESTABLISHED:
10320                         rack_timer_cancel(tp, rack,
10321                             rack->r_ctl.rc_rcvtime, __LINE__);
10322                         tcp_state_change(tp, TCPS_CLOSE_WAIT);
10323                         break;
10324
10325                         /*
10326                          * If still in FIN_WAIT_1 STATE FIN has not been
10327                          * acked so enter the CLOSING state.
10328                          */
10329                 case TCPS_FIN_WAIT_1:
10330                         rack_timer_cancel(tp, rack,
10331                             rack->r_ctl.rc_rcvtime, __LINE__);
10332                         tcp_state_change(tp, TCPS_CLOSING);
10333                         break;
10334
10335                         /*
10336                          * In FIN_WAIT_2 state enter the TIME_WAIT state,
10337                          * starting the time-wait timer, turning off the
10338                          * other standard timers.
10339                          */
10340                 case TCPS_FIN_WAIT_2:
10341                         rack_timer_cancel(tp, rack,
10342                             rack->r_ctl.rc_rcvtime, __LINE__);
10343                         tcp_twstart(tp);
10344                         return (1);
10345                 }
10346         }
10347         /*
10348          * Return any desired output.
10349          */
10350         if ((tp->t_flags & TF_ACKNOW) ||
10351             (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10352                 rack->r_wanted_output = 1;
10353         }
10354         INP_WLOCK_ASSERT(tp->t_inpcb);
10355         return (0);
10356 }
10357
10358 /*
10359  * Here nothing is really faster, its just that we
10360  * have broken out the fast-data path also just like
10361  * the fast-ack.
10362  */
10363 static int
10364 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10365     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10366     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10367 {
10368         int32_t nsegs;
10369         int32_t newsize = 0;    /* automatic sockbuf scaling */
10370         struct tcp_rack *rack;
10371 #ifdef NETFLIX_SB_LIMITS
10372         u_int mcnt, appended;
10373 #endif
10374 #ifdef TCPDEBUG
10375         /*
10376          * The size of tcp_saveipgen must be the size of the max ip header,
10377          * now IPv6.
10378          */
10379         u_char tcp_saveipgen[IP6_HDR_LEN];
10380         struct tcphdr tcp_savetcp;
10381         short ostate = 0;
10382
10383 #endif
10384         /*
10385          * If last ACK falls within this segment's sequence numbers, record
10386          * the timestamp. NOTE that the test is modified according to the
10387          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10388          */
10389         if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10390                 return (0);
10391         }
10392         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10393                 return (0);
10394         }
10395         if (tiwin && tiwin != tp->snd_wnd) {
10396                 return (0);
10397         }
10398         if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10399                 return (0);
10400         }
10401         if (__predict_false((to->to_flags & TOF_TS) &&
10402             (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10403                 return (0);
10404         }
10405         if (__predict_false((th->th_ack != tp->snd_una))) {
10406                 return (0);
10407         }
10408         if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10409                 return (0);
10410         }
10411         if ((to->to_flags & TOF_TS) != 0 &&
10412             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10413                 tp->ts_recent_age = tcp_ts_getticks();
10414                 tp->ts_recent = to->to_tsval;
10415         }
10416         rack = (struct tcp_rack *)tp->t_fb_ptr;
10417         /*
10418          * This is a pure, in-sequence data packet with nothing on the
10419          * reassembly queue and we have enough buffer space to take it.
10420          */
10421         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10422
10423 #ifdef NETFLIX_SB_LIMITS
10424         if (so->so_rcv.sb_shlim) {
10425                 mcnt = m_memcnt(m);
10426                 appended = 0;
10427                 if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10428                     CFO_NOSLEEP, NULL) == false) {
10429                         counter_u64_add(tcp_sb_shlim_fails, 1);
10430                         m_freem(m);
10431                         return (1);
10432                 }
10433         }
10434 #endif
10435         /* Clean receiver SACK report if present */
10436         if (tp->rcv_numsacks)
10437                 tcp_clean_sackreport(tp);
10438         KMOD_TCPSTAT_INC(tcps_preddat);
10439         tp->rcv_nxt += tlen;
10440         if (tlen &&
10441             ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10442             (tp->t_fbyte_in == 0)) {
10443                 tp->t_fbyte_in = ticks;
10444                 if (tp->t_fbyte_in == 0)
10445                         tp->t_fbyte_in = 1;
10446                 if (tp->t_fbyte_out && tp->t_fbyte_in)
10447                         tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10448         }
10449         /*
10450          * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10451          */
10452         tp->snd_wl1 = th->th_seq;
10453         /*
10454          * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10455          */
10456         tp->rcv_up = tp->rcv_nxt;
10457         KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10458         KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10459 #ifdef TCPDEBUG
10460         if (so->so_options & SO_DEBUG)
10461                 tcp_trace(TA_INPUT, ostate, tp,
10462                     (void *)tcp_saveipgen, &tcp_savetcp, 0);
10463 #endif
10464         newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10465
10466         /* Add data to socket buffer. */
10467         SOCKBUF_LOCK(&so->so_rcv);
10468         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10469                 m_freem(m);
10470         } else {
10471                 /*
10472                  * Set new socket buffer size. Give up when limit is
10473                  * reached.
10474                  */
10475                 if (newsize)
10476                         if (!sbreserve_locked(&so->so_rcv,
10477                             newsize, so, NULL))
10478                                 so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10479                 m_adj(m, drop_hdrlen);  /* delayed header drop */
10480 #ifdef NETFLIX_SB_LIMITS
10481                 appended =
10482 #endif
10483                         sbappendstream_locked(&so->so_rcv, m, 0);
10484                 ctf_calc_rwin(so, tp);
10485         }
10486         rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10487         tp->t_flags |= TF_WAKESOR;
10488 #ifdef NETFLIX_SB_LIMITS
10489         if (so->so_rcv.sb_shlim && mcnt != appended)
10490                 counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10491 #endif
10492         rack_handle_delayed_ack(tp, rack, tlen, 0);
10493         if (tp->snd_una == tp->snd_max)
10494                 sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10495         tcp_handle_wakeup(tp, so);
10496         return (1);
10497 }
10498
10499 /*
10500  * This subfunction is used to try to highly optimize the
10501  * fast path. We again allow window updates that are
10502  * in sequence to remain in the fast-path. We also add
10503  * in the __predict's to attempt to help the compiler.
10504  * Note that if we return a 0, then we can *not* process
10505  * it and the caller should push the packet into the
10506  * slow-path.
10507  */
10508 static int
10509 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10510     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10511     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10512 {
10513         int32_t acked;
10514         int32_t nsegs;
10515 #ifdef TCPDEBUG
10516         /*
10517          * The size of tcp_saveipgen must be the size of the max ip header,
10518          * now IPv6.
10519          */
10520         u_char tcp_saveipgen[IP6_HDR_LEN];
10521         struct tcphdr tcp_savetcp;
10522         short ostate = 0;
10523 #endif
10524         int32_t under_pacing = 0;
10525         struct tcp_rack *rack;
10526
10527         if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10528                 /* Old ack, behind (or duplicate to) the last one rcv'd */
10529                 return (0);
10530         }
10531         if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10532                 /* Above what we have sent? */
10533                 return (0);
10534         }
10535         if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10536                 /* We are retransmitting */
10537                 return (0);
10538         }
10539         if (__predict_false(tiwin == 0)) {
10540                 /* zero window */
10541                 return (0);
10542         }
10543         if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10544                 /* We need a SYN or a FIN, unlikely.. */
10545                 return (0);
10546         }
10547         if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10548                 /* Timestamp is behind .. old ack with seq wrap? */
10549                 return (0);
10550         }
10551         if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10552                 /* Still recovering */
10553                 return (0);
10554         }
10555         rack = (struct tcp_rack *)tp->t_fb_ptr;
10556         if (rack->r_ctl.rc_sacked) {
10557                 /* We have sack holes on our scoreboard */
10558                 return (0);
10559         }
10560         /* Ok if we reach here, we can process a fast-ack */
10561         if (rack->gp_ready &&
10562             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10563                 under_pacing = 1;
10564         }
10565         nsegs = max(1, m->m_pkthdr.lro_nsegs);
10566         rack_log_ack(tp, to, th, 0, 0);
10567         /* Did the window get updated? */
10568         if (tiwin != tp->snd_wnd) {
10569                 tp->snd_wnd = tiwin;
10570                 rack_validate_fo_sendwin_up(tp, rack);
10571                 tp->snd_wl1 = th->th_seq;
10572                 if (tp->snd_wnd > tp->max_sndwnd)
10573                         tp->max_sndwnd = tp->snd_wnd;
10574         }
10575         /* Do we exit persists? */
10576         if ((rack->rc_in_persist != 0) &&
10577             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10578                                rack->r_ctl.rc_pace_min_segs))) {
10579                 rack_exit_persist(tp, rack, cts);
10580         }
10581         /* Do we enter persists? */
10582         if ((rack->rc_in_persist == 0) &&
10583             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10584             TCPS_HAVEESTABLISHED(tp->t_state) &&
10585             (tp->snd_max == tp->snd_una) &&
10586             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10587             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10588                 /*
10589                  * Here the rwnd is less than
10590                  * the pacing size, we are established,
10591                  * nothing is outstanding, and there is
10592                  * data to send. Enter persists.
10593                  */
10594                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10595         }
10596         /*
10597          * If last ACK falls within this segment's sequence numbers, record
10598          * the timestamp. NOTE that the test is modified according to the
10599          * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10600          */
10601         if ((to->to_flags & TOF_TS) != 0 &&
10602             SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10603                 tp->ts_recent_age = tcp_ts_getticks();
10604                 tp->ts_recent = to->to_tsval;
10605         }
10606         /*
10607          * This is a pure ack for outstanding data.
10608          */
10609         KMOD_TCPSTAT_INC(tcps_predack);
10610
10611         /*
10612          * "bad retransmit" recovery.
10613          */
10614         if ((tp->t_flags & TF_PREVVALID) &&
10615             ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10616                 tp->t_flags &= ~TF_PREVVALID;
10617                 if (tp->t_rxtshift == 1 &&
10618                     (int)(ticks - tp->t_badrxtwin) < 0)
10619                         rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10620         }
10621         /*
10622          * Recalculate the transmit timer / rtt.
10623          *
10624          * Some boxes send broken timestamp replies during the SYN+ACK
10625          * phase, ignore timestamps of 0 or we could calculate a huge RTT
10626          * and blow up the retransmit timer.
10627          */
10628         acked = BYTES_THIS_ACK(tp, th);
10629
10630 #ifdef TCP_HHOOK
10631         /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
10632         hhook_run_tcp_est_in(tp, th, to);
10633 #endif
10634         KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10635         KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10636         if (acked) {
10637                 struct mbuf *mfree;
10638
10639                 rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
10640                 SOCKBUF_LOCK(&so->so_snd);
10641                 mfree = sbcut_locked(&so->so_snd, acked);
10642                 tp->snd_una = th->th_ack;
10643                 /* Note we want to hold the sb lock through the sendmap adjust */
10644                 rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10645                 /* Wake up the socket if we have room to write more */
10646                 rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10647                 sowwakeup_locked(so);
10648                 m_freem(mfree);
10649                 tp->t_rxtshift = 0;
10650                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10651                               rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10652                 rack->rc_tlp_in_progress = 0;
10653                 rack->r_ctl.rc_tlp_cnt_out = 0;
10654                 /*
10655                  * If it is the RXT timer we want to
10656                  * stop it, so we can restart a TLP.
10657                  */
10658                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10659                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10660 #ifdef NETFLIX_HTTP_LOGGING
10661                 tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10662 #endif
10663         }
10664         /*
10665          * Let the congestion control algorithm update congestion control
10666          * related information. This typically means increasing the
10667          * congestion window.
10668          */
10669         if (tp->snd_wnd < ctf_outstanding(tp)) {
10670                 /* The peer collapsed the window */
10671                 rack_collapsed_window(rack);
10672         } else if (rack->rc_has_collapsed)
10673                 rack_un_collapse_window(rack);
10674
10675         /*
10676          * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
10677          */
10678         tp->snd_wl2 = th->th_ack;
10679         tp->t_dupacks = 0;
10680         m_freem(m);
10681         /* ND6_HINT(tp);         *//* Some progress has been made. */
10682
10683         /*
10684          * If all outstanding data are acked, stop retransmit timer,
10685          * otherwise restart timer using current (possibly backed-off)
10686          * value. If process is waiting for space, wakeup/selwakeup/signal.
10687          * If data are ready to send, let tcp_output decide between more
10688          * output or persist.
10689          */
10690 #ifdef TCPDEBUG
10691         if (so->so_options & SO_DEBUG)
10692                 tcp_trace(TA_INPUT, ostate, tp,
10693                     (void *)tcp_saveipgen,
10694                     &tcp_savetcp, 0);
10695 #endif
10696         if (under_pacing &&
10697             (rack->use_fixed_rate == 0) &&
10698             (rack->in_probe_rtt == 0) &&
10699             rack->rc_gp_dyn_mul &&
10700             rack->rc_always_pace) {
10701                 /* Check if we are dragging bottom */
10702                 rack_check_bottom_drag(tp, rack, so, acked);
10703         }
10704         if (tp->snd_una == tp->snd_max) {
10705                 tp->t_flags &= ~TF_PREVVALID;
10706                 rack->r_ctl.retran_during_recovery = 0;
10707                 rack->r_ctl.dsack_byte_cnt = 0;
10708                 rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10709                 if (rack->r_ctl.rc_went_idle_time == 0)
10710                         rack->r_ctl.rc_went_idle_time = 1;
10711                 rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10712                 if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10713                         tp->t_acktime = 0;
10714                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10715         }
10716         if (acked && rack->r_fast_output)
10717                 rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
10718         if (sbavail(&so->so_snd)) {
10719                 rack->r_wanted_output = 1;
10720         }
10721         return (1);
10722 }
10723
10724 /*
10725  * Return value of 1, the TCB is unlocked and most
10726  * likely gone, return value of 0, the TCP is still
10727  * locked.
10728  */
10729 static int
10730 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
10731     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10732     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10733 {
10734         int32_t ret_val = 0;
10735         int32_t todrop;
10736         int32_t ourfinisacked = 0;
10737         struct tcp_rack *rack;
10738
10739         ctf_calc_rwin(so, tp);
10740         /*
10741          * If the state is SYN_SENT: if seg contains an ACK, but not for our
10742          * SYN, drop the input. if seg contains a RST, then drop the
10743          * connection. if seg does not contain SYN, then drop it. Otherwise
10744          * this is an acceptable SYN segment initialize tp->rcv_nxt and
10745          * tp->irs if seg contains ack then advance tp->snd_una if seg
10746          * contains an ECE and ECN support is enabled, the stream is ECN
10747          * capable. if SYN has been acked change to ESTABLISHED else
10748          * SYN_RCVD state arrange for segment to be acked (eventually)
10749          * continue processing rest of data/controls.
10750          */
10751         if ((thflags & TH_ACK) &&
10752             (SEQ_LEQ(th->th_ack, tp->iss) ||
10753             SEQ_GT(th->th_ack, tp->snd_max))) {
10754                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10755                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10756                 return (1);
10757         }
10758         if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
10759                 TCP_PROBE5(connect__refused, NULL, tp,
10760                     mtod(m, const char *), tp, th);
10761                 tp = tcp_drop(tp, ECONNREFUSED);
10762                 ctf_do_drop(m, tp);
10763                 return (1);
10764         }
10765         if (thflags & TH_RST) {
10766                 ctf_do_drop(m, tp);
10767                 return (1);
10768         }
10769         if (!(thflags & TH_SYN)) {
10770                 ctf_do_drop(m, tp);
10771                 return (1);
10772         }
10773         tp->irs = th->th_seq;
10774         tcp_rcvseqinit(tp);
10775         rack = (struct tcp_rack *)tp->t_fb_ptr;
10776         if (thflags & TH_ACK) {
10777                 int tfo_partial = 0;
10778
10779                 KMOD_TCPSTAT_INC(tcps_connects);
10780                 soisconnected(so);
10781 #ifdef MAC
10782                 mac_socketpeer_set_from_mbuf(m, so);
10783 #endif
10784                 /* Do window scaling on this connection? */
10785                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10786                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10787                         tp->rcv_scale = tp->request_r_scale;
10788                 }
10789                 tp->rcv_adv += min(tp->rcv_wnd,
10790                     TCP_MAXWIN << tp->rcv_scale);
10791                 /*
10792                  * If not all the data that was sent in the TFO SYN
10793                  * has been acked, resend the remainder right away.
10794                  */
10795                 if (IS_FASTOPEN(tp->t_flags) &&
10796                     (tp->snd_una != tp->snd_max)) {
10797                         tp->snd_nxt = th->th_ack;
10798                         tfo_partial = 1;
10799                 }
10800                 /*
10801                  * If there's data, delay ACK; if there's also a FIN ACKNOW
10802                  * will be turned on later.
10803                  */
10804                 if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
10805                         rack_timer_cancel(tp, rack,
10806                                           rack->r_ctl.rc_rcvtime, __LINE__);
10807                         tp->t_flags |= TF_DELACK;
10808                 } else {
10809                         rack->r_wanted_output = 1;
10810                         tp->t_flags |= TF_ACKNOW;
10811                         rack->rc_dack_toggle = 0;
10812                 }
10813                 if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
10814                     (V_tcp_do_ecn == 1)) {
10815                         tp->t_flags2 |= TF2_ECN_PERMIT;
10816                         KMOD_TCPSTAT_INC(tcps_ecn_shs);
10817                 }
10818                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
10819                         /*
10820                          * We advance snd_una for the
10821                          * fast open case. If th_ack is
10822                          * acknowledging data beyond
10823                          * snd_una we can't just call
10824                          * ack-processing since the
10825                          * data stream in our send-map
10826                          * will start at snd_una + 1 (one
10827                          * beyond the SYN). If its just
10828                          * equal we don't need to do that
10829                          * and there is no send_map.
10830                          */
10831                         tp->snd_una++;
10832                 }
10833                 /*
10834                  * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
10835                  * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
10836                  */
10837                 tp->t_starttime = ticks;
10838                 if (tp->t_flags & TF_NEEDFIN) {
10839                         tcp_state_change(tp, TCPS_FIN_WAIT_1);
10840                         tp->t_flags &= ~TF_NEEDFIN;
10841                         thflags &= ~TH_SYN;
10842                 } else {
10843                         tcp_state_change(tp, TCPS_ESTABLISHED);
10844                         TCP_PROBE5(connect__established, NULL, tp,
10845                             mtod(m, const char *), tp, th);
10846                         rack_cc_conn_init(tp);
10847                 }
10848         } else {
10849                 /*
10850                  * Received initial SYN in SYN-SENT[*] state => simultaneous
10851                  * open.  If segment contains CC option and there is a
10852                  * cached CC, apply TAO test. If it succeeds, connection is *
10853                  * half-synchronized. Otherwise, do 3-way handshake:
10854                  * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
10855                  * there was no CC option, clear cached CC value.
10856                  */
10857                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
10858                 tcp_state_change(tp, TCPS_SYN_RECEIVED);
10859         }
10860         INP_WLOCK_ASSERT(tp->t_inpcb);
10861         /*
10862          * Advance th->th_seq to correspond to first data byte. If data,
10863          * trim to stay within window, dropping FIN if necessary.
10864          */
10865         th->th_seq++;
10866         if (tlen > tp->rcv_wnd) {
10867                 todrop = tlen - tp->rcv_wnd;
10868                 m_adj(m, -todrop);
10869                 tlen = tp->rcv_wnd;
10870                 thflags &= ~TH_FIN;
10871                 KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
10872                 KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
10873         }
10874         tp->snd_wl1 = th->th_seq - 1;
10875         tp->rcv_up = th->th_seq;
10876         /*
10877          * Client side of transaction: already sent SYN and data. If the
10878          * remote host used T/TCP to validate the SYN, our data will be
10879          * ACK'd; if so, enter normal data segment processing in the middle
10880          * of step 5, ack processing. Otherwise, goto step 6.
10881          */
10882         if (thflags & TH_ACK) {
10883                 /* For syn-sent we need to possibly update the rtt */
10884                 if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
10885                         uint32_t t, mcts;
10886
10887                         mcts = tcp_ts_getticks();
10888                         t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
10889                         if (!tp->t_rttlow || tp->t_rttlow > t)
10890                                 tp->t_rttlow = t;
10891                         rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
10892                         tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
10893                         tcp_rack_xmit_timer_commit(rack, tp);
10894                 }
10895                 if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
10896                         return (ret_val);
10897                 /* We may have changed to FIN_WAIT_1 above */
10898                 if (tp->t_state == TCPS_FIN_WAIT_1) {
10899                         /*
10900                          * In FIN_WAIT_1 STATE in addition to the processing
10901                          * for the ESTABLISHED state if our FIN is now
10902                          * acknowledged then enter FIN_WAIT_2.
10903                          */
10904                         if (ourfinisacked) {
10905                                 /*
10906                                  * If we can't receive any more data, then
10907                                  * closing user can proceed. Starting the
10908                                  * timer is contrary to the specification,
10909                                  * but if we don't get a FIN we'll hang
10910                                  * forever.
10911                                  *
10912                                  * XXXjl: we should release the tp also, and
10913                                  * use a compressed state.
10914                                  */
10915                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10916                                         soisdisconnected(so);
10917                                         tcp_timer_activate(tp, TT_2MSL,
10918                                             (tcp_fast_finwait2_recycle ?
10919                                             tcp_finwait2_timeout :
10920                                             TP_MAXIDLE(tp)));
10921                                 }
10922                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
10923                         }
10924                 }
10925         }
10926         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
10927            tiwin, thflags, nxt_pkt));
10928 }
10929
10930 /*
10931  * Return value of 1, the TCB is unlocked and most
10932  * likely gone, return value of 0, the TCP is still
10933  * locked.
10934  */
10935 static int
10936 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
10937     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10938     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
10939 {
10940         struct tcp_rack *rack;
10941         int32_t ret_val = 0;
10942         int32_t ourfinisacked = 0;
10943
10944         ctf_calc_rwin(so, tp);
10945         if ((thflags & TH_ACK) &&
10946             (SEQ_LEQ(th->th_ack, tp->snd_una) ||
10947             SEQ_GT(th->th_ack, tp->snd_max))) {
10948                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10949                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10950                 return (1);
10951         }
10952         rack = (struct tcp_rack *)tp->t_fb_ptr;
10953         if (IS_FASTOPEN(tp->t_flags)) {
10954                 /*
10955                  * When a TFO connection is in SYN_RECEIVED, the
10956                  * only valid packets are the initial SYN, a
10957                  * retransmit/copy of the initial SYN (possibly with
10958                  * a subset of the original data), a valid ACK, a
10959                  * FIN, or a RST.
10960                  */
10961                 if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
10962                         tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10963                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
10964                         return (1);
10965                 } else if (thflags & TH_SYN) {
10966                         /* non-initial SYN is ignored */
10967                         if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
10968                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
10969                             (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
10970                                 ctf_do_drop(m, NULL);
10971                                 return (0);
10972                         }
10973                 } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
10974                         ctf_do_drop(m, NULL);
10975                         return (0);
10976                 }
10977         }
10978         if ((thflags & TH_RST) ||
10979             (tp->t_fin_is_rst && (thflags & TH_FIN)))
10980                 return (ctf_process_rst(m, th, so, tp));
10981         /*
10982          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
10983          * it's less than ts_recent, drop it.
10984          */
10985         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
10986             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
10987                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
10988                         return (ret_val);
10989         }
10990         /*
10991          * In the SYN-RECEIVED state, validate that the packet belongs to
10992          * this connection before trimming the data to fit the receive
10993          * window.  Check the sequence number versus IRS since we know the
10994          * sequence numbers haven't wrapped.  This is a partial fix for the
10995          * "LAND" DoS attack.
10996          */
10997         if (SEQ_LT(th->th_seq, tp->irs)) {
10998                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
10999                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11000                 return (1);
11001         }
11002         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11003                               &rack->r_ctl.challenge_ack_ts,
11004                               &rack->r_ctl.challenge_ack_cnt)) {
11005                 return (ret_val);
11006         }
11007         /*
11008          * If last ACK falls within this segment's sequence numbers, record
11009          * its timestamp. NOTE: 1) That the test incorporates suggestions
11010          * from the latest proposal of the tcplw@cray.com list (Braden
11011          * 1993/04/26). 2) That updating only on newer timestamps interferes
11012          * with our earlier PAWS tests, so this check should be solely
11013          * predicated on the sequence space of this segment. 3) That we
11014          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11015          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11016          * SEG.Len, This modified check allows us to overcome RFC1323's
11017          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11018          * p.869. In such cases, we can still calculate the RTT correctly
11019          * when RCV.NXT == Last.ACK.Sent.
11020          */
11021         if ((to->to_flags & TOF_TS) != 0 &&
11022             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11023             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11024             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11025                 tp->ts_recent_age = tcp_ts_getticks();
11026                 tp->ts_recent = to->to_tsval;
11027         }
11028         tp->snd_wnd = tiwin;
11029         rack_validate_fo_sendwin_up(tp, rack);
11030         /*
11031          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11032          * is on (half-synchronized state), then queue data for later
11033          * processing; else drop segment and return.
11034          */
11035         if ((thflags & TH_ACK) == 0) {
11036                 if (IS_FASTOPEN(tp->t_flags)) {
11037                         rack_cc_conn_init(tp);
11038                 }
11039                 return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11040                     tiwin, thflags, nxt_pkt));
11041         }
11042         KMOD_TCPSTAT_INC(tcps_connects);
11043         soisconnected(so);
11044         /* Do window scaling? */
11045         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11046             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11047                 tp->rcv_scale = tp->request_r_scale;
11048         }
11049         /*
11050          * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11051          * FIN-WAIT-1
11052          */
11053         tp->t_starttime = ticks;
11054         if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11055                 tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11056                 tp->t_tfo_pending = NULL;
11057         }
11058         if (tp->t_flags & TF_NEEDFIN) {
11059                 tcp_state_change(tp, TCPS_FIN_WAIT_1);
11060                 tp->t_flags &= ~TF_NEEDFIN;
11061         } else {
11062                 tcp_state_change(tp, TCPS_ESTABLISHED);
11063                 TCP_PROBE5(accept__established, NULL, tp,
11064                     mtod(m, const char *), tp, th);
11065                 /*
11066                  * TFO connections call cc_conn_init() during SYN
11067                  * processing.  Calling it again here for such connections
11068                  * is not harmless as it would undo the snd_cwnd reduction
11069                  * that occurs when a TFO SYN|ACK is retransmitted.
11070                  */
11071                 if (!IS_FASTOPEN(tp->t_flags))
11072                         rack_cc_conn_init(tp);
11073         }
11074         /*
11075          * Account for the ACK of our SYN prior to
11076          * regular ACK processing below, except for
11077          * simultaneous SYN, which is handled later.
11078          */
11079         if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11080                 tp->snd_una++;
11081         /*
11082          * If segment contains data or ACK, will call tcp_reass() later; if
11083          * not, do so now to pass queued data to user.
11084          */
11085         if (tlen == 0 && (thflags & TH_FIN) == 0) {
11086                 (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11087                     (struct mbuf *)0);
11088                 tcp_handle_wakeup(tp, so);
11089         }
11090         tp->snd_wl1 = th->th_seq - 1;
11091         /* For syn-recv we need to possibly update the rtt */
11092         if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11093                 uint32_t t, mcts;
11094
11095                 mcts = tcp_ts_getticks();
11096                 t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11097                 if (!tp->t_rttlow || tp->t_rttlow > t)
11098                         tp->t_rttlow = t;
11099                 rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11100                 tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11101                 tcp_rack_xmit_timer_commit(rack, tp);
11102         }
11103         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11104                 return (ret_val);
11105         }
11106         if (tp->t_state == TCPS_FIN_WAIT_1) {
11107                 /* We could have went to FIN_WAIT_1 (or EST) above */
11108                 /*
11109                  * In FIN_WAIT_1 STATE in addition to the processing for the
11110                  * ESTABLISHED state if our FIN is now acknowledged then
11111                  * enter FIN_WAIT_2.
11112                  */
11113                 if (ourfinisacked) {
11114                         /*
11115                          * If we can't receive any more data, then closing
11116                          * user can proceed. Starting the timer is contrary
11117                          * to the specification, but if we don't get a FIN
11118                          * we'll hang forever.
11119                          *
11120                          * XXXjl: we should release the tp also, and use a
11121                          * compressed state.
11122                          */
11123                         if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11124                                 soisdisconnected(so);
11125                                 tcp_timer_activate(tp, TT_2MSL,
11126                                     (tcp_fast_finwait2_recycle ?
11127                                     tcp_finwait2_timeout :
11128                                     TP_MAXIDLE(tp)));
11129                         }
11130                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
11131                 }
11132         }
11133         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11134             tiwin, thflags, nxt_pkt));
11135 }
11136
11137 /*
11138  * Return value of 1, the TCB is unlocked and most
11139  * likely gone, return value of 0, the TCP is still
11140  * locked.
11141  */
11142 static int
11143 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11144     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11145     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11146 {
11147         int32_t ret_val = 0;
11148         struct tcp_rack *rack;
11149
11150         /*
11151          * Header prediction: check for the two common cases of a
11152          * uni-directional data xfer.  If the packet has no control flags,
11153          * is in-sequence, the window didn't change and we're not
11154          * retransmitting, it's a candidate.  If the length is zero and the
11155          * ack moved forward, we're the sender side of the xfer.  Just free
11156          * the data acked & wake any higher level process that was blocked
11157          * waiting for space.  If the length is non-zero and the ack didn't
11158          * move, we're the receiver side.  If we're getting packets in-order
11159          * (the reassembly queue is empty), add the data toc The socket
11160          * buffer and note that we need a delayed ack. Make sure that the
11161          * hidden state-flags are also off. Since we check for
11162          * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11163          */
11164         rack = (struct tcp_rack *)tp->t_fb_ptr;
11165         if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11166             __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11167             __predict_true(SEGQ_EMPTY(tp)) &&
11168             __predict_true(th->th_seq == tp->rcv_nxt)) {
11169                 if (tlen == 0) {
11170                         if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11171                             tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11172                                 return (0);
11173                         }
11174                 } else {
11175                         if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11176                             tiwin, nxt_pkt, iptos)) {
11177                                 return (0);
11178                         }
11179                 }
11180         }
11181         ctf_calc_rwin(so, tp);
11182
11183         if ((thflags & TH_RST) ||
11184             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11185                 return (ctf_process_rst(m, th, so, tp));
11186
11187         /*
11188          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11189          * synchronized state.
11190          */
11191         if (thflags & TH_SYN) {
11192                 ctf_challenge_ack(m, th, tp, &ret_val);
11193                 return (ret_val);
11194         }
11195         /*
11196          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11197          * it's less than ts_recent, drop it.
11198          */
11199         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11200             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11201                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11202                         return (ret_val);
11203         }
11204         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11205                               &rack->r_ctl.challenge_ack_ts,
11206                               &rack->r_ctl.challenge_ack_cnt)) {
11207                 return (ret_val);
11208         }
11209         /*
11210          * If last ACK falls within this segment's sequence numbers, record
11211          * its timestamp. NOTE: 1) That the test incorporates suggestions
11212          * from the latest proposal of the tcplw@cray.com list (Braden
11213          * 1993/04/26). 2) That updating only on newer timestamps interferes
11214          * with our earlier PAWS tests, so this check should be solely
11215          * predicated on the sequence space of this segment. 3) That we
11216          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11217          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11218          * SEG.Len, This modified check allows us to overcome RFC1323's
11219          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11220          * p.869. In such cases, we can still calculate the RTT correctly
11221          * when RCV.NXT == Last.ACK.Sent.
11222          */
11223         if ((to->to_flags & TOF_TS) != 0 &&
11224             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11225             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11226             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11227                 tp->ts_recent_age = tcp_ts_getticks();
11228                 tp->ts_recent = to->to_tsval;
11229         }
11230         /*
11231          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11232          * is on (half-synchronized state), then queue data for later
11233          * processing; else drop segment and return.
11234          */
11235         if ((thflags & TH_ACK) == 0) {
11236                 if (tp->t_flags & TF_NEEDSYN) {
11237                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11238                             tiwin, thflags, nxt_pkt));
11239
11240                 } else if (tp->t_flags & TF_ACKNOW) {
11241                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11242                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11243                         return (ret_val);
11244                 } else {
11245                         ctf_do_drop(m, NULL);
11246                         return (0);
11247                 }
11248         }
11249         /*
11250          * Ack processing.
11251          */
11252         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11253                 return (ret_val);
11254         }
11255         if (sbavail(&so->so_snd)) {
11256                 if (ctf_progress_timeout_check(tp, true)) {
11257                         rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11258                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11259                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11260                         return (1);
11261                 }
11262         }
11263         /* State changes only happen in rack_process_data() */
11264         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11265             tiwin, thflags, nxt_pkt));
11266 }
11267
11268 /*
11269  * Return value of 1, the TCB is unlocked and most
11270  * likely gone, return value of 0, the TCP is still
11271  * locked.
11272  */
11273 static int
11274 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11275     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11276     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11277 {
11278         int32_t ret_val = 0;
11279         struct tcp_rack *rack;
11280
11281         rack = (struct tcp_rack *)tp->t_fb_ptr;
11282         ctf_calc_rwin(so, tp);
11283         if ((thflags & TH_RST) ||
11284             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11285                 return (ctf_process_rst(m, th, so, tp));
11286         /*
11287          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11288          * synchronized state.
11289          */
11290         if (thflags & TH_SYN) {
11291                 ctf_challenge_ack(m, th, tp, &ret_val);
11292                 return (ret_val);
11293         }
11294         /*
11295          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11296          * it's less than ts_recent, drop it.
11297          */
11298         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11299             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11300                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11301                         return (ret_val);
11302         }
11303         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11304                               &rack->r_ctl.challenge_ack_ts,
11305                               &rack->r_ctl.challenge_ack_cnt)) {
11306                 return (ret_val);
11307         }
11308         /*
11309          * If last ACK falls within this segment's sequence numbers, record
11310          * its timestamp. NOTE: 1) That the test incorporates suggestions
11311          * from the latest proposal of the tcplw@cray.com list (Braden
11312          * 1993/04/26). 2) That updating only on newer timestamps interferes
11313          * with our earlier PAWS tests, so this check should be solely
11314          * predicated on the sequence space of this segment. 3) That we
11315          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11316          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11317          * SEG.Len, This modified check allows us to overcome RFC1323's
11318          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11319          * p.869. In such cases, we can still calculate the RTT correctly
11320          * when RCV.NXT == Last.ACK.Sent.
11321          */
11322         if ((to->to_flags & TOF_TS) != 0 &&
11323             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11324             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11325             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11326                 tp->ts_recent_age = tcp_ts_getticks();
11327                 tp->ts_recent = to->to_tsval;
11328         }
11329         /*
11330          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11331          * is on (half-synchronized state), then queue data for later
11332          * processing; else drop segment and return.
11333          */
11334         if ((thflags & TH_ACK) == 0) {
11335                 if (tp->t_flags & TF_NEEDSYN) {
11336                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11337                             tiwin, thflags, nxt_pkt));
11338
11339                 } else if (tp->t_flags & TF_ACKNOW) {
11340                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11341                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11342                         return (ret_val);
11343                 } else {
11344                         ctf_do_drop(m, NULL);
11345                         return (0);
11346                 }
11347         }
11348         /*
11349          * Ack processing.
11350          */
11351         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11352                 return (ret_val);
11353         }
11354         if (sbavail(&so->so_snd)) {
11355                 if (ctf_progress_timeout_check(tp, true)) {
11356                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11357                                                 tp, tick, PROGRESS_DROP, __LINE__);
11358                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11359                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11360                         return (1);
11361                 }
11362         }
11363         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11364             tiwin, thflags, nxt_pkt));
11365 }
11366
11367 static int
11368 rack_check_data_after_close(struct mbuf *m,
11369     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11370 {
11371         struct tcp_rack *rack;
11372
11373         rack = (struct tcp_rack *)tp->t_fb_ptr;
11374         if (rack->rc_allow_data_af_clo == 0) {
11375         close_now:
11376                 tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11377                 /* tcp_close will kill the inp pre-log the Reset */
11378                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11379                 tp = tcp_close(tp);
11380                 KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11381                 ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11382                 return (1);
11383         }
11384         if (sbavail(&so->so_snd) == 0)
11385                 goto close_now;
11386         /* Ok we allow data that is ignored and a followup reset */
11387         tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11388         tp->rcv_nxt = th->th_seq + *tlen;
11389         tp->t_flags2 |= TF2_DROP_AF_DATA;
11390         rack->r_wanted_output = 1;
11391         *tlen = 0;
11392         return (0);
11393 }
11394
11395 /*
11396  * Return value of 1, the TCB is unlocked and most
11397  * likely gone, return value of 0, the TCP is still
11398  * locked.
11399  */
11400 static int
11401 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11402     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11403     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11404 {
11405         int32_t ret_val = 0;
11406         int32_t ourfinisacked = 0;
11407         struct tcp_rack *rack;
11408
11409         rack = (struct tcp_rack *)tp->t_fb_ptr;
11410         ctf_calc_rwin(so, tp);
11411
11412         if ((thflags & TH_RST) ||
11413             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11414                 return (ctf_process_rst(m, th, so, tp));
11415         /*
11416          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11417          * synchronized state.
11418          */
11419         if (thflags & TH_SYN) {
11420                 ctf_challenge_ack(m, th, tp, &ret_val);
11421                 return (ret_val);
11422         }
11423         /*
11424          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11425          * it's less than ts_recent, drop it.
11426          */
11427         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11428             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11429                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11430                         return (ret_val);
11431         }
11432         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11433                               &rack->r_ctl.challenge_ack_ts,
11434                               &rack->r_ctl.challenge_ack_cnt)) {
11435                 return (ret_val);
11436         }
11437         /*
11438          * If new data are received on a connection after the user processes
11439          * are gone, then RST the other end.
11440          */
11441         if ((so->so_state & SS_NOFDREF) && tlen) {
11442                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11443                         return (1);
11444         }
11445         /*
11446          * If last ACK falls within this segment's sequence numbers, record
11447          * its timestamp. NOTE: 1) That the test incorporates suggestions
11448          * from the latest proposal of the tcplw@cray.com list (Braden
11449          * 1993/04/26). 2) That updating only on newer timestamps interferes
11450          * with our earlier PAWS tests, so this check should be solely
11451          * predicated on the sequence space of this segment. 3) That we
11452          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11453          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11454          * SEG.Len, This modified check allows us to overcome RFC1323's
11455          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11456          * p.869. In such cases, we can still calculate the RTT correctly
11457          * when RCV.NXT == Last.ACK.Sent.
11458          */
11459         if ((to->to_flags & TOF_TS) != 0 &&
11460             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11461             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11462             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11463                 tp->ts_recent_age = tcp_ts_getticks();
11464                 tp->ts_recent = to->to_tsval;
11465         }
11466         /*
11467          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11468          * is on (half-synchronized state), then queue data for later
11469          * processing; else drop segment and return.
11470          */
11471         if ((thflags & TH_ACK) == 0) {
11472                 if (tp->t_flags & TF_NEEDSYN) {
11473                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11474                             tiwin, thflags, nxt_pkt));
11475                 } else if (tp->t_flags & TF_ACKNOW) {
11476                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11477                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11478                         return (ret_val);
11479                 } else {
11480                         ctf_do_drop(m, NULL);
11481                         return (0);
11482                 }
11483         }
11484         /*
11485          * Ack processing.
11486          */
11487         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11488                 return (ret_val);
11489         }
11490         if (ourfinisacked) {
11491                 /*
11492                  * If we can't receive any more data, then closing user can
11493                  * proceed. Starting the timer is contrary to the
11494                  * specification, but if we don't get a FIN we'll hang
11495                  * forever.
11496                  *
11497                  * XXXjl: we should release the tp also, and use a
11498                  * compressed state.
11499                  */
11500                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11501                         soisdisconnected(so);
11502                         tcp_timer_activate(tp, TT_2MSL,
11503                             (tcp_fast_finwait2_recycle ?
11504                             tcp_finwait2_timeout :
11505                             TP_MAXIDLE(tp)));
11506                 }
11507                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
11508         }
11509         if (sbavail(&so->so_snd)) {
11510                 if (ctf_progress_timeout_check(tp, true)) {
11511                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11512                                                 tp, tick, PROGRESS_DROP, __LINE__);
11513                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11514                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11515                         return (1);
11516                 }
11517         }
11518         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11519             tiwin, thflags, nxt_pkt));
11520 }
11521
11522 /*
11523  * Return value of 1, the TCB is unlocked and most
11524  * likely gone, return value of 0, the TCP is still
11525  * locked.
11526  */
11527 static int
11528 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11529     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11530     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11531 {
11532         int32_t ret_val = 0;
11533         int32_t ourfinisacked = 0;
11534         struct tcp_rack *rack;
11535
11536         rack = (struct tcp_rack *)tp->t_fb_ptr;
11537         ctf_calc_rwin(so, tp);
11538
11539         if ((thflags & TH_RST) ||
11540             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11541                 return (ctf_process_rst(m, th, so, tp));
11542         /*
11543          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11544          * synchronized state.
11545          */
11546         if (thflags & TH_SYN) {
11547                 ctf_challenge_ack(m, th, tp, &ret_val);
11548                 return (ret_val);
11549         }
11550         /*
11551          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11552          * it's less than ts_recent, drop it.
11553          */
11554         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11555             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11556                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11557                         return (ret_val);
11558         }
11559         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11560                               &rack->r_ctl.challenge_ack_ts,
11561                               &rack->r_ctl.challenge_ack_cnt)) {
11562                 return (ret_val);
11563         }
11564         /*
11565          * If new data are received on a connection after the user processes
11566          * are gone, then RST the other end.
11567          */
11568         if ((so->so_state & SS_NOFDREF) && tlen) {
11569                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11570                         return (1);
11571         }
11572         /*
11573          * If last ACK falls within this segment's sequence numbers, record
11574          * its timestamp. NOTE: 1) That the test incorporates suggestions
11575          * from the latest proposal of the tcplw@cray.com list (Braden
11576          * 1993/04/26). 2) That updating only on newer timestamps interferes
11577          * with our earlier PAWS tests, so this check should be solely
11578          * predicated on the sequence space of this segment. 3) That we
11579          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11580          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11581          * SEG.Len, This modified check allows us to overcome RFC1323's
11582          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11583          * p.869. In such cases, we can still calculate the RTT correctly
11584          * when RCV.NXT == Last.ACK.Sent.
11585          */
11586         if ((to->to_flags & TOF_TS) != 0 &&
11587             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11588             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11589             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11590                 tp->ts_recent_age = tcp_ts_getticks();
11591                 tp->ts_recent = to->to_tsval;
11592         }
11593         /*
11594          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11595          * is on (half-synchronized state), then queue data for later
11596          * processing; else drop segment and return.
11597          */
11598         if ((thflags & TH_ACK) == 0) {
11599                 if (tp->t_flags & TF_NEEDSYN) {
11600                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11601                             tiwin, thflags, nxt_pkt));
11602                 } else if (tp->t_flags & TF_ACKNOW) {
11603                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11604                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11605                         return (ret_val);
11606                 } else {
11607                         ctf_do_drop(m, NULL);
11608                         return (0);
11609                 }
11610         }
11611         /*
11612          * Ack processing.
11613          */
11614         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11615                 return (ret_val);
11616         }
11617         if (ourfinisacked) {
11618                 tcp_twstart(tp);
11619                 m_freem(m);
11620                 return (1);
11621         }
11622         if (sbavail(&so->so_snd)) {
11623                 if (ctf_progress_timeout_check(tp, true)) {
11624                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11625                                                 tp, tick, PROGRESS_DROP, __LINE__);
11626                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11627                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11628                         return (1);
11629                 }
11630         }
11631         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11632             tiwin, thflags, nxt_pkt));
11633 }
11634
11635 /*
11636  * Return value of 1, the TCB is unlocked and most
11637  * likely gone, return value of 0, the TCP is still
11638  * locked.
11639  */
11640 static int
11641 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11642     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11643     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11644 {
11645         int32_t ret_val = 0;
11646         int32_t ourfinisacked = 0;
11647         struct tcp_rack *rack;
11648
11649         rack = (struct tcp_rack *)tp->t_fb_ptr;
11650         ctf_calc_rwin(so, tp);
11651
11652         if ((thflags & TH_RST) ||
11653             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11654                 return (ctf_process_rst(m, th, so, tp));
11655         /*
11656          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11657          * synchronized state.
11658          */
11659         if (thflags & TH_SYN) {
11660                 ctf_challenge_ack(m, th, tp, &ret_val);
11661                 return (ret_val);
11662         }
11663         /*
11664          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11665          * it's less than ts_recent, drop it.
11666          */
11667         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11668             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11669                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11670                         return (ret_val);
11671         }
11672         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11673                               &rack->r_ctl.challenge_ack_ts,
11674                               &rack->r_ctl.challenge_ack_cnt)) {
11675                 return (ret_val);
11676         }
11677         /*
11678          * If new data are received on a connection after the user processes
11679          * are gone, then RST the other end.
11680          */
11681         if ((so->so_state & SS_NOFDREF) && tlen) {
11682                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11683                         return (1);
11684         }
11685         /*
11686          * If last ACK falls within this segment's sequence numbers, record
11687          * its timestamp. NOTE: 1) That the test incorporates suggestions
11688          * from the latest proposal of the tcplw@cray.com list (Braden
11689          * 1993/04/26). 2) That updating only on newer timestamps interferes
11690          * with our earlier PAWS tests, so this check should be solely
11691          * predicated on the sequence space of this segment. 3) That we
11692          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11693          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11694          * SEG.Len, This modified check allows us to overcome RFC1323's
11695          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11696          * p.869. In such cases, we can still calculate the RTT correctly
11697          * when RCV.NXT == Last.ACK.Sent.
11698          */
11699         if ((to->to_flags & TOF_TS) != 0 &&
11700             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11701             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11702             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11703                 tp->ts_recent_age = tcp_ts_getticks();
11704                 tp->ts_recent = to->to_tsval;
11705         }
11706         /*
11707          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11708          * is on (half-synchronized state), then queue data for later
11709          * processing; else drop segment and return.
11710          */
11711         if ((thflags & TH_ACK) == 0) {
11712                 if (tp->t_flags & TF_NEEDSYN) {
11713                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11714                             tiwin, thflags, nxt_pkt));
11715                 } else if (tp->t_flags & TF_ACKNOW) {
11716                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11717                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11718                         return (ret_val);
11719                 } else {
11720                         ctf_do_drop(m, NULL);
11721                         return (0);
11722                 }
11723         }
11724         /*
11725          * case TCPS_LAST_ACK: Ack processing.
11726          */
11727         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11728                 return (ret_val);
11729         }
11730         if (ourfinisacked) {
11731                 tp = tcp_close(tp);
11732                 ctf_do_drop(m, tp);
11733                 return (1);
11734         }
11735         if (sbavail(&so->so_snd)) {
11736                 if (ctf_progress_timeout_check(tp, true)) {
11737                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11738                                                 tp, tick, PROGRESS_DROP, __LINE__);
11739                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11740                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11741                         return (1);
11742                 }
11743         }
11744         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11745             tiwin, thflags, nxt_pkt));
11746 }
11747
11748 /*
11749  * Return value of 1, the TCB is unlocked and most
11750  * likely gone, return value of 0, the TCP is still
11751  * locked.
11752  */
11753 static int
11754 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
11755     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11756     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11757 {
11758         int32_t ret_val = 0;
11759         int32_t ourfinisacked = 0;
11760         struct tcp_rack *rack;
11761
11762         rack = (struct tcp_rack *)tp->t_fb_ptr;
11763         ctf_calc_rwin(so, tp);
11764
11765         /* Reset receive buffer auto scaling when not in bulk receive mode. */
11766         if ((thflags & TH_RST) ||
11767             (tp->t_fin_is_rst && (thflags & TH_FIN)))
11768                 return (ctf_process_rst(m, th, so, tp));
11769         /*
11770          * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11771          * synchronized state.
11772          */
11773         if (thflags & TH_SYN) {
11774                 ctf_challenge_ack(m, th, tp, &ret_val);
11775                 return (ret_val);
11776         }
11777         /*
11778          * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11779          * it's less than ts_recent, drop it.
11780          */
11781         if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11782             TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11783                 if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11784                         return (ret_val);
11785         }
11786         if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11787                               &rack->r_ctl.challenge_ack_ts,
11788                               &rack->r_ctl.challenge_ack_cnt)) {
11789                 return (ret_val);
11790         }
11791         /*
11792          * If new data are received on a connection after the user processes
11793          * are gone, then RST the other end.
11794          */
11795         if ((so->so_state & SS_NOFDREF) &&
11796             tlen) {
11797                 if (rack_check_data_after_close(m, tp, &tlen, th, so))
11798                         return (1);
11799         }
11800         /*
11801          * If last ACK falls within this segment's sequence numbers, record
11802          * its timestamp. NOTE: 1) That the test incorporates suggestions
11803          * from the latest proposal of the tcplw@cray.com list (Braden
11804          * 1993/04/26). 2) That updating only on newer timestamps interferes
11805          * with our earlier PAWS tests, so this check should be solely
11806          * predicated on the sequence space of this segment. 3) That we
11807          * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11808          * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11809          * SEG.Len, This modified check allows us to overcome RFC1323's
11810          * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11811          * p.869. In such cases, we can still calculate the RTT correctly
11812          * when RCV.NXT == Last.ACK.Sent.
11813          */
11814         if ((to->to_flags & TOF_TS) != 0 &&
11815             SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11816             SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11817             ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11818                 tp->ts_recent_age = tcp_ts_getticks();
11819                 tp->ts_recent = to->to_tsval;
11820         }
11821         /*
11822          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11823          * is on (half-synchronized state), then queue data for later
11824          * processing; else drop segment and return.
11825          */
11826         if ((thflags & TH_ACK) == 0) {
11827                 if (tp->t_flags & TF_NEEDSYN) {
11828                         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11829                             tiwin, thflags, nxt_pkt));
11830                 } else if (tp->t_flags & TF_ACKNOW) {
11831                         ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11832                         ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11833                         return (ret_val);
11834                 } else {
11835                         ctf_do_drop(m, NULL);
11836                         return (0);
11837                 }
11838         }
11839         /*
11840          * Ack processing.
11841          */
11842         if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11843                 return (ret_val);
11844         }
11845         if (sbavail(&so->so_snd)) {
11846                 if (ctf_progress_timeout_check(tp, true)) {
11847                         rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11848                                                 tp, tick, PROGRESS_DROP, __LINE__);
11849                         tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11850                         ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11851                         return (1);
11852                 }
11853         }
11854         return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11855             tiwin, thflags, nxt_pkt));
11856 }
11857
11858 static void inline
11859 rack_clear_rate_sample(struct tcp_rack *rack)
11860 {
11861         rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
11862         rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
11863         rack->r_ctl.rack_rs.rs_rtt_tot = 0;
11864 }
11865
11866 static void
11867 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
11868 {
11869         uint64_t bw_est, rate_wanted;
11870         int chged = 0;
11871         uint32_t user_max, orig_min, orig_max;
11872
11873         orig_min = rack->r_ctl.rc_pace_min_segs;
11874         orig_max = rack->r_ctl.rc_pace_max_segs;
11875         user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
11876         if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
11877                 chged = 1;
11878         rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
11879         if (rack->use_fixed_rate || rack->rc_force_max_seg) {
11880                 if (user_max != rack->r_ctl.rc_pace_max_segs)
11881                         chged = 1;
11882         }
11883         if (rack->rc_force_max_seg) {
11884                 rack->r_ctl.rc_pace_max_segs = user_max;
11885         } else if (rack->use_fixed_rate) {
11886                 bw_est = rack_get_bw(rack);
11887                 if ((rack->r_ctl.crte == NULL) ||
11888                     (bw_est != rack->r_ctl.crte->rate)) {
11889                         rack->r_ctl.rc_pace_max_segs = user_max;
11890                 } else {
11891                         /* We are pacing right at the hardware rate */
11892                         uint32_t segsiz;
11893
11894                         segsiz = min(ctf_fixed_maxseg(tp),
11895                                      rack->r_ctl.rc_pace_min_segs);
11896                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
11897                                                            tp, bw_est, segsiz, 0,
11898                                                            rack->r_ctl.crte, NULL);
11899                 }
11900         } else if (rack->rc_always_pace) {
11901                 if (rack->r_ctl.gp_bw ||
11902 #ifdef NETFLIX_PEAKRATE
11903                     rack->rc_tp->t_maxpeakrate ||
11904 #endif
11905                     rack->r_ctl.init_rate) {
11906                         /* We have a rate of some sort set */
11907                         uint32_t  orig;
11908
11909                         bw_est = rack_get_bw(rack);
11910                         orig = rack->r_ctl.rc_pace_max_segs;
11911                         if (fill_override)
11912                                 rate_wanted = *fill_override;
11913                         else
11914                                 rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
11915                         if (rate_wanted) {
11916                                 /* We have something */
11917                                 rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
11918                                                                                    rate_wanted,
11919                                                                                    ctf_fixed_maxseg(rack->rc_tp));
11920                         } else
11921                                 rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
11922                         if (orig != rack->r_ctl.rc_pace_max_segs)
11923                                 chged = 1;
11924                 } else if ((rack->r_ctl.gp_bw == 0) &&
11925                            (rack->r_ctl.rc_pace_max_segs == 0)) {
11926                         /*
11927                          * If we have nothing limit us to bursting
11928                          * out IW sized pieces.
11929                          */
11930                         chged = 1;
11931                         rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
11932                 }
11933         }
11934         if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
11935                 chged = 1;
11936                 rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
11937         }
11938         if (chged)
11939                 rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
11940 }
11941
11942
11943 static void
11944 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
11945 {
11946 #ifdef INET6
11947         struct ip6_hdr *ip6 = NULL;
11948 #endif
11949 #ifdef INET
11950         struct ip *ip = NULL;
11951 #endif
11952         struct udphdr *udp = NULL;
11953
11954         /* Ok lets fill in the fast block, it can only be used with no IP options! */
11955 #ifdef INET6
11956         if (rack->r_is_v6) {
11957                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
11958                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
11959                 if (tp->t_port) {
11960                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11961                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
11962                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11963                         udp->uh_dport = tp->t_port;
11964                         rack->r_ctl.fsb.udp = udp;
11965                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11966                 } else
11967                 {
11968                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
11969                         rack->r_ctl.fsb.udp = NULL;
11970                 }
11971                 tcpip_fillheaders(rack->rc_inp,
11972                                   tp->t_port,
11973                                   ip6, rack->r_ctl.fsb.th);
11974         } else
11975 #endif                          /* INET6 */
11976         {
11977                 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
11978                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
11979                 if (tp->t_port) {
11980                         rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
11981                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
11982                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
11983                         udp->uh_dport = tp->t_port;
11984                         rack->r_ctl.fsb.udp = udp;
11985                         rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
11986                 } else
11987                 {
11988                         rack->r_ctl.fsb.udp = NULL;
11989                         rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
11990                 }
11991                 tcpip_fillheaders(rack->rc_inp,
11992                                   tp->t_port,
11993                                   ip, rack->r_ctl.fsb.th);
11994         }
11995         rack->r_fsb_inited = 1;
11996 }
11997
11998 static int
11999 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12000 {
12001         /*
12002          * Allocate the larger of spaces V6 if available else just
12003          * V4 and include udphdr (overbook)
12004          */
12005 #ifdef INET6
12006         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12007 #else
12008         rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12009 #endif
12010         rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12011                                             M_TCPFSB, M_NOWAIT|M_ZERO);
12012         if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12013                 return (ENOMEM);
12014         }
12015         rack->r_fsb_inited = 0;
12016         return (0);
12017 }
12018
12019 static int
12020 rack_init(struct tcpcb *tp)
12021 {
12022         struct tcp_rack *rack = NULL;
12023         struct rack_sendmap *insret;
12024         uint32_t iwin, snt, us_cts;
12025         int err;
12026
12027         tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12028         if (tp->t_fb_ptr == NULL) {
12029                 /*
12030                  * We need to allocate memory but cant. The INP and INP_INFO
12031                  * locks and they are recusive (happens during setup. So a
12032                  * scheme to drop the locks fails :(
12033                  *
12034                  */
12035                 return (ENOMEM);
12036         }
12037         memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12038
12039         rack = (struct tcp_rack *)tp->t_fb_ptr;
12040         RB_INIT(&rack->r_ctl.rc_mtree);
12041         TAILQ_INIT(&rack->r_ctl.rc_free);
12042         TAILQ_INIT(&rack->r_ctl.rc_tmap);
12043         rack->rc_tp = tp;
12044         rack->rc_inp = tp->t_inpcb;
12045         /* Set the flag */
12046         rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12047         /* Probably not needed but lets be sure */
12048         rack_clear_rate_sample(rack);
12049         /*
12050          * Save off the default values, socket options will poke
12051          * at these if pacing is not on or we have not yet
12052          * reached where pacing is on (gp_ready/fixed enabled).
12053          * When they get set into the CC module (when gp_ready
12054          * is enabled or we enable fixed) then we will set these
12055          * values into the CC and place in here the old values
12056          * so we have a restoral. Then we will set the flag
12057          * rc_pacing_cc_set. That way whenever we turn off pacing
12058          * or switch off this stack, we will know to go restore
12059          * the saved values.
12060          */
12061         rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12062         rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12063         /* We want abe like behavior as well */
12064         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
12065         rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12066         rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12067         rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12068         if (use_rack_rr)
12069                 rack->use_rack_rr = 1;
12070         if (V_tcp_delack_enabled)
12071                 tp->t_delayed_ack = 1;
12072         else
12073                 tp->t_delayed_ack = 0;
12074 #ifdef TCP_ACCOUNTING
12075         if (rack_tcp_accounting) {
12076                 tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12077         }
12078 #endif
12079         if (rack_enable_shared_cwnd)
12080                 rack->rack_enable_scwnd = 1;
12081         rack->rc_user_set_max_segs = rack_hptsi_segments;
12082         rack->rc_force_max_seg = 0;
12083         if (rack_use_imac_dack)
12084                 rack->rc_dack_mode = 1;
12085         TAILQ_INIT(&rack->r_ctl.opt_list);
12086         rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12087         rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12088         rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12089         rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12090         rack->r_ctl.rc_highest_us_rtt = 0;
12091         rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12092         rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12093         if (rack_use_cmp_acks)
12094                 rack->r_use_cmp_ack = 1;
12095         if (rack_disable_prr)
12096                 rack->rack_no_prr = 1;
12097         if (rack_gp_no_rec_chg)
12098                 rack->rc_gp_no_rec_chg = 1;
12099         if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12100                 rack->rc_always_pace = 1;
12101                 if (rack->use_fixed_rate || rack->gp_ready)
12102                         rack_set_cc_pacing(rack);
12103         } else
12104                 rack->rc_always_pace = 0;
12105         if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12106                 rack->r_mbuf_queue = 1;
12107         else
12108                 rack->r_mbuf_queue = 0;
12109         if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12110                 tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12111         else
12112                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12113         rack_set_pace_segments(tp, rack, __LINE__, NULL);
12114         if (rack_limits_scwnd)
12115                 rack->r_limit_scw = 1;
12116         else
12117                 rack->r_limit_scw = 0;
12118         rack->rc_labc = V_tcp_abc_l_var;
12119         rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12120         rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12121         rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12122         rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12123         rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12124         rack->r_ctl.rc_min_to = rack_min_to;
12125         microuptime(&rack->r_ctl.act_rcv_time);
12126         rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12127         rack->r_running_late = 0;
12128         rack->r_running_early = 0;
12129         rack->rc_init_win = rack_default_init_window;
12130         rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12131         if (rack_hw_up_only)
12132                 rack->r_up_only = 1;
12133         if (rack_do_dyn_mul) {
12134                 /* When dynamic adjustment is on CA needs to start at 100% */
12135                 rack->rc_gp_dyn_mul = 1;
12136                 if (rack_do_dyn_mul >= 100)
12137                         rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12138         } else
12139                 rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12140         rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12141         rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12142         rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12143         setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12144                                 rack_probertt_filter_life);
12145         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12146         rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12147         rack->r_ctl.rc_time_of_last_probertt = us_cts;
12148         rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12149         rack->r_ctl.rc_time_probertt_starts = 0;
12150         /* We require at least one measurement, even if the sysctl is 0 */
12151         if (rack_req_measurements)
12152                 rack->r_ctl.req_measurements = rack_req_measurements;
12153         else
12154                 rack->r_ctl.req_measurements = 1;
12155         if (rack_enable_hw_pacing)
12156                 rack->rack_hdw_pace_ena = 1;
12157         if (rack_hw_rate_caps)
12158                 rack->r_rack_hw_rate_caps = 1;
12159         /* Do we force on detection? */
12160 #ifdef NETFLIX_EXP_DETECTION
12161         if (tcp_force_detection)
12162                 rack->do_detection = 1;
12163         else
12164 #endif
12165                 rack->do_detection = 0;
12166         if (rack_non_rxt_use_cr)
12167                 rack->rack_rec_nonrxt_use_cr = 1;
12168         err = rack_init_fsb(tp, rack);
12169         if (err) {
12170                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12171                 tp->t_fb_ptr = NULL;
12172                 return (err);
12173         }
12174         if (tp->snd_una != tp->snd_max) {
12175                 /* Create a send map for the current outstanding data */
12176                 struct rack_sendmap *rsm;
12177
12178                 rsm = rack_alloc(rack);
12179                 if (rsm == NULL) {
12180                         uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12181                         tp->t_fb_ptr = NULL;
12182                         return (ENOMEM);
12183                 }
12184                 rsm->r_no_rtt_allowed = 1;
12185                 rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12186                 rsm->r_rtr_cnt = 1;
12187                 rsm->r_rtr_bytes = 0;
12188                 if (tp->t_flags & TF_SENTFIN) {
12189                         rsm->r_end = tp->snd_max - 1;
12190                         rsm->r_flags |= RACK_HAS_FIN;
12191                 } else {
12192                         rsm->r_end = tp->snd_max;
12193                 }
12194                 if (tp->snd_una == tp->iss) {
12195                         /* The data space is one beyond snd_una */
12196                         rsm->r_flags |= RACK_HAS_SYN;
12197                         rsm->r_start = tp->iss;
12198                         rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12199                 } else
12200                         rsm->r_start = tp->snd_una;
12201                 rsm->r_dupack = 0;
12202                 if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12203                         rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12204                         if (rsm->m)
12205                                 rsm->orig_m_len = rsm->m->m_len;
12206                         else
12207                                 rsm->orig_m_len = 0;
12208                 } else {
12209                         /*
12210                          * This can happen if we have a stand-alone FIN or
12211                          *  SYN.
12212                          */
12213                         rsm->m = NULL;
12214                         rsm->orig_m_len = 0;
12215                         rsm->soff = 0;
12216                 }
12217                 insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12218 #ifdef INVARIANTS
12219                 if (insret != NULL) {
12220                         panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12221                               insret, rack, rsm);
12222                 }
12223 #endif
12224                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12225                 rsm->r_in_tmap = 1;
12226         }
12227         /*
12228          * Timers in Rack are kept in microseconds so lets
12229          * convert any initial incoming variables
12230          * from ticks into usecs. Note that we
12231          * also change the values of t_srtt and t_rttvar, if
12232          * they are non-zero. They are kept with a 5
12233          * bit decimal so we have to carefully convert
12234          * these to get the full precision.
12235          */
12236         rack_convert_rtts(tp);
12237         tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12238         if (rack_def_profile)
12239                 rack_set_profile(rack, rack_def_profile);
12240         /* Cancel the GP measurement in progress */
12241         tp->t_flags &= ~TF_GPUTINPROG;
12242         if (SEQ_GT(tp->snd_max, tp->iss))
12243                 snt = tp->snd_max - tp->iss;
12244         else
12245                 snt = 0;
12246         iwin = rc_init_window(rack);
12247         if (snt < iwin) {
12248                 /* We are not past the initial window
12249                  * so we need to make sure cwnd is
12250                  * correct.
12251                  */
12252                 if (tp->snd_cwnd < iwin)
12253                         tp->snd_cwnd = iwin;
12254                 /*
12255                  * If we are within the initial window
12256                  * we want ssthresh to be unlimited. Setting
12257                  * it to the rwnd (which the default stack does
12258                  * and older racks) is not really a good idea
12259                  * since we want to be in SS and grow both the
12260                  * cwnd and the rwnd (via dynamic rwnd growth). If
12261                  * we set it to the rwnd then as the peer grows its
12262                  * rwnd we will be stuck in CA and never hit SS.
12263                  *
12264                  * Its far better to raise it up high (this takes the
12265                  * risk that there as been a loss already, probably
12266                  * we should have an indicator in all stacks of loss
12267                  * but we don't), but considering the normal use this
12268                  * is a risk worth taking. The consequences of not
12269                  * hitting SS are far worse than going one more time
12270                  * into it early on (before we have sent even a IW).
12271                  * It is highly unlikely that we will have had a loss
12272                  * before getting the IW out.
12273                  */
12274                 tp->snd_ssthresh = 0xffffffff;
12275         }
12276         rack_stop_all_timers(tp);
12277         /* Lets setup the fsb block */
12278         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12279         rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12280                              __LINE__, RACK_RTTS_INIT);
12281         return (0);
12282 }
12283
12284 static int
12285 rack_handoff_ok(struct tcpcb *tp)
12286 {
12287         if ((tp->t_state == TCPS_CLOSED) ||
12288             (tp->t_state == TCPS_LISTEN)) {
12289                 /* Sure no problem though it may not stick */
12290                 return (0);
12291         }
12292         if ((tp->t_state == TCPS_SYN_SENT) ||
12293             (tp->t_state == TCPS_SYN_RECEIVED)) {
12294                 /*
12295                  * We really don't know if you support sack,
12296                  * you have to get to ESTAB or beyond to tell.
12297                  */
12298                 return (EAGAIN);
12299         }
12300         if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12301                 /*
12302                  * Rack will only send a FIN after all data is acknowledged.
12303                  * So in this case we have more data outstanding. We can't
12304                  * switch stacks until either all data and only the FIN
12305                  * is left (in which case rack_init() now knows how
12306                  * to deal with that) <or> all is acknowledged and we
12307                  * are only left with incoming data, though why you
12308                  * would want to switch to rack after all data is acknowledged
12309                  * I have no idea (rrs)!
12310                  */
12311                 return (EAGAIN);
12312         }
12313         if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12314                 return (0);
12315         }
12316         /*
12317          * If we reach here we don't do SACK on this connection so we can
12318          * never do rack.
12319          */
12320         return (EINVAL);
12321 }
12322
12323
12324 static void
12325 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12326 {
12327         int ack_cmp = 0;
12328
12329         if (tp->t_fb_ptr) {
12330                 struct tcp_rack *rack;
12331                 struct rack_sendmap *rsm, *nrsm, *rm;
12332
12333                 rack = (struct tcp_rack *)tp->t_fb_ptr;
12334                 if (tp->t_in_pkt) {
12335                         /*
12336                          * Since we are switching we need to process any
12337                          * inbound packets in case a compressed ack is
12338                          * in queue or the new stack does not support
12339                          * mbuf queuing. These packets in theory should
12340                          * have been handled by the old stack anyway.
12341                          */
12342                         if ((rack->rc_inp->inp_flags & (INP_DROPPED|INP_TIMEWAIT)) ||
12343                             (rack->rc_inp->inp_flags2 & INP_FREED)) {
12344                                 /* Kill all the packets */
12345                                 struct mbuf *save, *m;
12346
12347                                 m = tp->t_in_pkt;
12348                                 tp->t_in_pkt = NULL;
12349                                 tp->t_tail_pkt = NULL;
12350                                 while (m) {
12351                                         save = m->m_nextpkt;
12352                                         m->m_nextpkt = NULL;
12353                                         m_freem(m);
12354                                         m = save;
12355                                 }
12356                         } else {
12357                                 /* Process all the packets */
12358                                 ctf_do_queued_segments(rack->rc_inp->inp_socket, rack->rc_tp, 0);
12359                         }
12360                         if ((tp->t_inpcb) &&
12361                             (tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP))
12362                                 ack_cmp = 1;
12363                         if (ack_cmp) {
12364                                 /* Total if we used large or small (if ack-cmp was used). */
12365                                 if (rack->rc_inp->inp_flags2 & INP_MBUF_L_ACKS)
12366                                         counter_u64_add(rack_large_ackcmp, 1);
12367                                 else
12368                                         counter_u64_add(rack_small_ackcmp, 1);
12369                         }
12370                 }
12371                 tp->t_flags &= ~TF_FORCEDATA;
12372 #ifdef NETFLIX_SHARED_CWND
12373                 if (rack->r_ctl.rc_scw) {
12374                         uint32_t limit;
12375
12376                         if (rack->r_limit_scw)
12377                                 limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12378                         else
12379                                 limit = 0;
12380                         tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12381                                                   rack->r_ctl.rc_scw_index,
12382                                                   limit);
12383                         rack->r_ctl.rc_scw = NULL;
12384                 }
12385 #endif
12386                 if (rack->r_ctl.fsb.tcp_ip_hdr) {
12387                         free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12388                         rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12389                         rack->r_ctl.fsb.th = NULL;
12390                 }
12391                 /* Convert back to ticks, with  */
12392                 if (tp->t_srtt > 1) {
12393                         uint32_t val, frac;
12394
12395                         val = USEC_2_TICKS(tp->t_srtt);
12396                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12397                         tp->t_srtt = val << TCP_RTT_SHIFT;
12398                         /*
12399                          * frac is the fractional part here is left
12400                          * over from converting to hz and shifting.
12401                          * We need to convert this to the 5 bit
12402                          * remainder.
12403                          */
12404                         if (frac) {
12405                                 if (hz == 1000) {
12406                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12407                                 } else {
12408                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12409                                 }
12410                                 tp->t_srtt += frac;
12411                         }
12412                 }
12413                 if (tp->t_rttvar) {
12414                         uint32_t val, frac;
12415
12416                         val = USEC_2_TICKS(tp->t_rttvar);
12417                         frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12418                         tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12419                         /*
12420                          * frac is the fractional part here is left
12421                          * over from converting to hz and shifting.
12422                          * We need to convert this to the 5 bit
12423                          * remainder.
12424                          */
12425                         if (frac) {
12426                                 if (hz == 1000) {
12427                                         frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12428                                 } else {
12429                                         frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12430                                 }
12431                                 tp->t_rttvar += frac;
12432                         }
12433                 }
12434                 tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12435                 tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12436                 if (rack->rc_always_pace) {
12437                         tcp_decrement_paced_conn();
12438                         rack_undo_cc_pacing(rack);
12439                         rack->rc_always_pace = 0;
12440                 }
12441                 /* Clean up any options if they were not applied */
12442                 while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12443                         struct deferred_opt_list *dol;
12444
12445                         dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12446                         TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12447                         free(dol, M_TCPDO);
12448                 }
12449                 /* rack does not use force data but other stacks may clear it */
12450                 if (rack->r_ctl.crte != NULL) {
12451                         tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12452                         rack->rack_hdrw_pacing = 0;
12453                         rack->r_ctl.crte = NULL;
12454                 }
12455 #ifdef TCP_BLACKBOX
12456                 tcp_log_flowend(tp);
12457 #endif
12458                 RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12459                         rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12460 #ifdef INVARIANTS
12461                         if (rm != rsm) {
12462                                 panic("At fini, rack:%p rsm:%p rm:%p",
12463                                       rack, rsm, rm);
12464                         }
12465 #endif
12466                         uma_zfree(rack_zone, rsm);
12467                 }
12468                 rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12469                 while (rsm) {
12470                         TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12471                         uma_zfree(rack_zone, rsm);
12472                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12473                 }
12474                 rack->rc_free_cnt = 0;
12475                 uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12476                 tp->t_fb_ptr = NULL;
12477         }
12478         if (tp->t_inpcb) {
12479                 tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12480                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12481                 tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12482                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
12483                 /* Cancel the GP measurement in progress */
12484                 tp->t_flags &= ~TF_GPUTINPROG;
12485                 tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
12486         }
12487         /* Make sure snd_nxt is correctly set */
12488         tp->snd_nxt = tp->snd_max;
12489 }
12490
12491 static void
12492 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12493 {
12494         switch (tp->t_state) {
12495         case TCPS_SYN_SENT:
12496                 rack->r_state = TCPS_SYN_SENT;
12497                 rack->r_substate = rack_do_syn_sent;
12498                 break;
12499         case TCPS_SYN_RECEIVED:
12500                 rack->r_state = TCPS_SYN_RECEIVED;
12501                 rack->r_substate = rack_do_syn_recv;
12502                 break;
12503         case TCPS_ESTABLISHED:
12504                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12505                 rack->r_state = TCPS_ESTABLISHED;
12506                 rack->r_substate = rack_do_established;
12507                 break;
12508         case TCPS_CLOSE_WAIT:
12509                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12510                 rack->r_state = TCPS_CLOSE_WAIT;
12511                 rack->r_substate = rack_do_close_wait;
12512                 break;
12513         case TCPS_FIN_WAIT_1:
12514                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12515                 rack->r_state = TCPS_FIN_WAIT_1;
12516                 rack->r_substate = rack_do_fin_wait_1;
12517                 break;
12518         case TCPS_CLOSING:
12519                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12520                 rack->r_state = TCPS_CLOSING;
12521                 rack->r_substate = rack_do_closing;
12522                 break;
12523         case TCPS_LAST_ACK:
12524                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12525                 rack->r_state = TCPS_LAST_ACK;
12526                 rack->r_substate = rack_do_lastack;
12527                 break;
12528         case TCPS_FIN_WAIT_2:
12529                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
12530                 rack->r_state = TCPS_FIN_WAIT_2;
12531                 rack->r_substate = rack_do_fin_wait_2;
12532                 break;
12533         case TCPS_LISTEN:
12534         case TCPS_CLOSED:
12535         case TCPS_TIME_WAIT:
12536         default:
12537                 break;
12538         };
12539         if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12540                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12541
12542 }
12543
12544 static void
12545 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12546 {
12547         /*
12548          * We received an ack, and then did not
12549          * call send or were bounced out due to the
12550          * hpts was running. Now a timer is up as well, is
12551          * it the right timer?
12552          */
12553         struct rack_sendmap *rsm;
12554         int tmr_up;
12555
12556         tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12557         if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12558                 return;
12559         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12560         if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12561             (tmr_up == PACE_TMR_RXT)) {
12562                 /* Should be an RXT */
12563                 return;
12564         }
12565         if (rsm == NULL) {
12566                 /* Nothing outstanding? */
12567                 if (tp->t_flags & TF_DELACK) {
12568                         if (tmr_up == PACE_TMR_DELACK)
12569                                 /* We are supposed to have delayed ack up and we do */
12570                                 return;
12571                 } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12572                         /*
12573                          * if we hit enobufs then we would expect the possiblity
12574                          * of nothing outstanding and the RXT up (and the hptsi timer).
12575                          */
12576                         return;
12577                 } else if (((V_tcp_always_keepalive ||
12578                              rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12579                             (tp->t_state <= TCPS_CLOSING)) &&
12580                            (tmr_up == PACE_TMR_KEEP) &&
12581                            (tp->snd_max == tp->snd_una)) {
12582                         /* We should have keep alive up and we do */
12583                         return;
12584                 }
12585         }
12586         if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12587                    ((tmr_up == PACE_TMR_TLP) ||
12588                     (tmr_up == PACE_TMR_RACK) ||
12589                     (tmr_up == PACE_TMR_RXT))) {
12590                 /*
12591                  * Either a Rack, TLP or RXT is fine if  we
12592                  * have outstanding data.
12593                  */
12594                 return;
12595         } else if (tmr_up == PACE_TMR_DELACK) {
12596                 /*
12597                  * If the delayed ack was going to go off
12598                  * before the rtx/tlp/rack timer were going to
12599                  * expire, then that would be the timer in control.
12600                  * Note we don't check the time here trusting the
12601                  * code is correct.
12602                  */
12603                 return;
12604         }
12605         /*
12606          * Ok the timer originally started is not what we want now.
12607          * We will force the hpts to be stopped if any, and restart
12608          * with the slot set to what was in the saved slot.
12609          */
12610         if (rack->rc_inp->inp_in_hpts) {
12611                 if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12612                         uint32_t us_cts;
12613
12614                         us_cts = tcp_get_usecs(NULL);
12615                         if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12616                                 rack->r_early = 1;
12617                                 rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12618                         }
12619                         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12620                 }
12621                 tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
12622         }
12623         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12624         rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12625 }
12626
12627
12628 static void
12629 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)
12630 {
12631         tp->snd_wnd = tiwin;
12632         rack_validate_fo_sendwin_up(tp, rack);
12633         tp->snd_wl1 = seq;
12634         tp->snd_wl2 = ack;
12635         if (tp->snd_wnd > tp->max_sndwnd)
12636                 tp->max_sndwnd = tp->snd_wnd;
12637         if (tp->snd_wnd < (tp->snd_max - high_seq)) {
12638                 /* The peer collapsed the window */
12639                 rack_collapsed_window(rack);
12640         } else if (rack->rc_has_collapsed)
12641                 rack_un_collapse_window(rack);
12642         /* Do we exit persists? */
12643         if ((rack->rc_in_persist != 0) &&
12644             (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12645                                 rack->r_ctl.rc_pace_min_segs))) {
12646                 rack_exit_persist(tp, rack, cts);
12647         }
12648         /* Do we enter persists? */
12649         if ((rack->rc_in_persist == 0) &&
12650             (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12651             TCPS_HAVEESTABLISHED(tp->t_state) &&
12652             (tp->snd_max == tp->snd_una) &&
12653             sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
12654             (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
12655                 /*
12656                  * Here the rwnd is less than
12657                  * the pacing size, we are established,
12658                  * nothing is outstanding, and there is
12659                  * data to send. Enter persists.
12660                  */
12661                 rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12662         }
12663 }
12664
12665 static void
12666 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
12667 {
12668
12669         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
12670                 union tcp_log_stackspecific log;
12671                 struct timeval ltv;
12672                 char tcp_hdr_buf[60];
12673                 struct tcphdr *th;
12674                 struct timespec ts;
12675                 uint32_t orig_snd_una;
12676                 uint8_t xx = 0;
12677
12678 #ifdef NETFLIX_HTTP_LOGGING
12679                 struct http_sendfile_track *http_req;
12680
12681                 if (SEQ_GT(ae->ack, tp->snd_una)) {
12682                         http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
12683                 } else {
12684                         http_req = tcp_http_find_req_for_seq(tp, ae->ack);
12685                 }
12686 #endif
12687                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
12688                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
12689                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
12690                 if (rack->rack_no_prr == 0)
12691                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
12692                 else
12693                         log.u_bbr.flex1 = 0;
12694                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
12695                 log.u_bbr.use_lt_bw <<= 1;
12696                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
12697                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
12698                 log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
12699                 log.u_bbr.pkts_out = tp->t_maxseg;
12700                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
12701                 log.u_bbr.flex7 = 1;
12702                 log.u_bbr.lost = ae->flags;
12703                 log.u_bbr.cwnd_gain = ackval;
12704                 log.u_bbr.pacing_gain = 0x2;
12705                 if (ae->flags & TSTMP_HDWR) {
12706                         /* Record the hardware timestamp if present */
12707                         log.u_bbr.flex3 = M_TSTMP;
12708                         ts.tv_sec = ae->timestamp / 1000000000;
12709                         ts.tv_nsec = ae->timestamp % 1000000000;
12710                         ltv.tv_sec = ts.tv_sec;
12711                         ltv.tv_usec = ts.tv_nsec / 1000;
12712                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
12713                 } else if (ae->flags & TSTMP_LRO) {
12714                         /* Record the LRO the arrival timestamp */
12715                         log.u_bbr.flex3 = M_TSTMP_LRO;
12716                         ts.tv_sec = ae->timestamp / 1000000000;
12717                         ts.tv_nsec = ae->timestamp % 1000000000;
12718                         ltv.tv_sec = ts.tv_sec;
12719                         ltv.tv_usec = ts.tv_nsec / 1000;
12720                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
12721                 }
12722                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
12723                 /* Log the rcv time */
12724                 log.u_bbr.delRate = ae->timestamp;
12725 #ifdef NETFLIX_HTTP_LOGGING
12726                 log.u_bbr.applimited = tp->t_http_closed;
12727                 log.u_bbr.applimited <<= 8;
12728                 log.u_bbr.applimited |= tp->t_http_open;
12729                 log.u_bbr.applimited <<= 8;
12730                 log.u_bbr.applimited |= tp->t_http_req;
12731                 if (http_req) {
12732                         /* Copy out any client req info */
12733                         /* seconds */
12734                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
12735                         /* useconds */
12736                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
12737                         log.u_bbr.rttProp = http_req->timestamp;
12738                         log.u_bbr.cur_del_rate = http_req->start;
12739                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
12740                                 log.u_bbr.flex8 |= 1;
12741                         } else {
12742                                 log.u_bbr.flex8 |= 2;
12743                                 log.u_bbr.bw_inuse = http_req->end;
12744                         }
12745                         log.u_bbr.flex6 = http_req->start_seq;
12746                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
12747                                 log.u_bbr.flex8 |= 4;
12748                                 log.u_bbr.epoch = http_req->end_seq;
12749                         }
12750                 }
12751 #endif
12752                 memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
12753                 th = (struct tcphdr *)tcp_hdr_buf;
12754                 th->th_seq = ae->seq;
12755                 th->th_ack = ae->ack;
12756                 th->th_win = ae->win;
12757                 /* Now fill in the ports */
12758                 th->th_sport = tp->t_inpcb->inp_fport;
12759                 th->th_dport = tp->t_inpcb->inp_lport;
12760                 th->th_flags = ae->flags & 0xff;
12761                 /* Now do we have a timestamp option? */
12762                 if (ae->flags & HAS_TSTMP) {
12763                         u_char *cp;
12764                         uint32_t val;
12765
12766                         th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
12767                         cp = (u_char *)(th + 1);
12768                         *cp = TCPOPT_NOP;
12769                         cp++;
12770                         *cp = TCPOPT_NOP;
12771                         cp++;
12772                         *cp = TCPOPT_TIMESTAMP;
12773                         cp++;
12774                         *cp = TCPOLEN_TIMESTAMP;
12775                         cp++;
12776                         val = htonl(ae->ts_value);
12777                         bcopy((char *)&val,
12778                               (char *)cp, sizeof(uint32_t));
12779                         val = htonl(ae->ts_echo);
12780                         bcopy((char *)&val,
12781                               (char *)(cp + 4), sizeof(uint32_t));
12782                 } else
12783                         th->th_off = (sizeof(struct tcphdr) >> 2);
12784
12785                 /*
12786                  * For sane logging we need to play a little trick.
12787                  * If the ack were fully processed we would have moved
12788                  * snd_una to high_seq, but since compressed acks are
12789                  * processed in two phases, at this point (logging) snd_una
12790                  * won't be advanced. So we would see multiple acks showing
12791                  * the advancement. We can prevent that by "pretending" that
12792                  * snd_una was advanced and then un-advancing it so that the
12793                  * logging code has the right value for tlb_snd_una.
12794                  */
12795                 if (tp->snd_una != high_seq) {
12796                         orig_snd_una = tp->snd_una;
12797                         tp->snd_una = high_seq;
12798                         xx = 1;
12799                 } else
12800                         xx = 0;
12801                 TCP_LOG_EVENTP(tp, th,
12802                                &tp->t_inpcb->inp_socket->so_rcv,
12803                                &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
12804                                0, &log, true, &ltv);
12805                 if (xx) {
12806                         tp->snd_una = orig_snd_una;
12807                 }
12808         }
12809
12810 }
12811
12812 static int
12813 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
12814 {
12815         /*
12816          * Handle a "special" compressed ack mbuf. Each incoming
12817          * ack has only four possible dispositions:
12818          *
12819          * A) It moves the cum-ack forward
12820          * B) It is behind the cum-ack.
12821          * C) It is a window-update ack.
12822          * D) It is a dup-ack.
12823          *
12824          * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
12825          * in the incoming mbuf. We also need to still pay attention
12826          * to nxt_pkt since there may be another packet after this
12827          * one.
12828          */
12829 #ifdef TCP_ACCOUNTING
12830         uint64_t ts_val;
12831         uint64_t rdstc;
12832 #endif
12833         int segsiz;
12834         struct timespec ts;
12835         struct tcp_rack *rack;
12836         struct tcp_ackent *ae;
12837         uint32_t tiwin, us_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
12838         int cnt, i, did_out, ourfinisacked = 0;
12839         int win_up_req = 0;
12840         struct tcpopt to_holder, *to = NULL;
12841         int nsegs = 0;
12842         int under_pacing = 1;
12843         int recovery = 0;
12844         int idx;
12845 #ifdef TCP_ACCOUNTING
12846         sched_pin();
12847 #endif
12848         rack = (struct tcp_rack *)tp->t_fb_ptr;
12849         if (rack->gp_ready &&
12850             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
12851                 under_pacing = 0;
12852         else
12853                 under_pacing = 1;
12854
12855         if (rack->r_state != tp->t_state)
12856                 rack_set_state(tp, rack);
12857         to = &to_holder;
12858         to->to_flags = 0;
12859         KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
12860                 ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
12861         cnt = m->m_len / sizeof(struct tcp_ackent);
12862         idx = cnt / 5;
12863         if (idx >= MAX_NUM_OF_CNTS)
12864                 idx = MAX_NUM_OF_CNTS - 1;
12865         counter_u64_add(rack_proc_comp_ack[idx], 1);
12866         counter_u64_add(rack_multi_single_eq, cnt);
12867         high_seq = tp->snd_una;
12868         the_win = tp->snd_wnd;
12869         win_seq = tp->snd_wl1;
12870         win_upd_ack = tp->snd_wl2;
12871         cts = us_cts = tcp_tv_to_usectick(tv);
12872         segsiz = ctf_fixed_maxseg(tp);
12873         if ((rack->rc_gp_dyn_mul) &&
12874             (rack->use_fixed_rate == 0) &&
12875             (rack->rc_always_pace)) {
12876                 /* Check in on probertt */
12877                 rack_check_probe_rtt(rack, us_cts);
12878         }
12879         for (i = 0; i < cnt; i++) {
12880 #ifdef TCP_ACCOUNTING
12881                 ts_val = get_cyclecount();
12882 #endif
12883                 rack_clear_rate_sample(rack);
12884                 ae = ((mtod(m, struct tcp_ackent *)) + i);
12885                 /* Setup the window */
12886                 tiwin = ae->win << tp->snd_scale;
12887                 /* figure out the type of ack */
12888                 if (SEQ_LT(ae->ack, high_seq)) {
12889                         /* Case B*/
12890                         ae->ack_val_set = ACK_BEHIND;
12891                 } else if (SEQ_GT(ae->ack, high_seq)) {
12892                         /* Case A */
12893                         ae->ack_val_set = ACK_CUMACK;
12894                 } else if (tiwin == the_win) {
12895                         /* Case D */
12896                         ae->ack_val_set = ACK_DUPACK;
12897                 } else {
12898                         /* Case C */
12899                         ae->ack_val_set = ACK_RWND;
12900                 }
12901                 rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
12902                 /* Validate timestamp */
12903                 if (ae->flags & HAS_TSTMP) {
12904                         /* Setup for a timestamp */
12905                         to->to_flags = TOF_TS;
12906                         ae->ts_echo -= tp->ts_offset;
12907                         to->to_tsecr = ae->ts_echo;
12908                         to->to_tsval = ae->ts_value;
12909                         /*
12910                          * If echoed timestamp is later than the current time, fall back to
12911                          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
12912                          * were used when this connection was established.
12913                          */
12914                         if (TSTMP_GT(ae->ts_echo, cts))
12915                                 ae->ts_echo = 0;
12916                         if (tp->ts_recent &&
12917                             TSTMP_LT(ae->ts_value, tp->ts_recent)) {
12918                                 if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
12919 #ifdef TCP_ACCOUNTING
12920                                         rdstc = get_cyclecount();
12921                                         if (rdstc > ts_val) {
12922                                                 counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
12923                                                                 (rdstc - ts_val));
12924                                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12925                                                         tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
12926                                                 }
12927                                         }
12928 #endif
12929                                         continue;
12930                                 }
12931                         }
12932                         if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
12933                             SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
12934                                 tp->ts_recent_age = tcp_ts_getticks();
12935                                 tp->ts_recent = ae->ts_value;
12936                         }
12937                 } else {
12938                         /* Setup for a no options */
12939                         to->to_flags = 0;
12940                 }
12941                 /* Update the rcv time and perform idle reduction possibly */
12942                 if  (tp->t_idle_reduce &&
12943                      (tp->snd_max == tp->snd_una) &&
12944                      ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
12945                         counter_u64_add(rack_input_idle_reduces, 1);
12946                         rack_cc_after_idle(rack, tp);
12947                 }
12948                 tp->t_rcvtime = ticks;
12949                 /* Now what about ECN? */
12950                 if (tp->t_flags2 & TF2_ECN_PERMIT) {
12951                         if (ae->flags & TH_CWR) {
12952                                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
12953                                 tp->t_flags |= TF_ACKNOW;
12954                         }
12955                         switch (ae->codepoint & IPTOS_ECN_MASK) {
12956                         case IPTOS_ECN_CE:
12957                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
12958                                 KMOD_TCPSTAT_INC(tcps_ecn_ce);
12959                                 break;
12960                         case IPTOS_ECN_ECT0:
12961                                 KMOD_TCPSTAT_INC(tcps_ecn_ect0);
12962                                 break;
12963                         case IPTOS_ECN_ECT1:
12964                                 KMOD_TCPSTAT_INC(tcps_ecn_ect1);
12965                                 break;
12966                         }
12967
12968                         /* Process a packet differently from RFC3168. */
12969                         cc_ecnpkt_handler_flags(tp, ae->flags, ae->codepoint);
12970                         /* Congestion experienced. */
12971                         if (ae->flags & TH_ECE) {
12972                                 rack_cong_signal(tp,  CC_ECN, ae->ack);
12973                         }
12974                 }
12975 #ifdef TCP_ACCOUNTING
12976                 /* Count for the specific type of ack in */
12977                 counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
12978                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
12979                         tp->tcp_cnt_counters[ae->ack_val_set]++;
12980                 }
12981 #endif
12982                 /*
12983                  * Note how we could move up these in the determination
12984                  * above, but we don't so that way the timestamp checks (and ECN)
12985                  * is done first before we do any processing on the ACK.
12986                  * The non-compressed path through the code has this
12987                  * weakness (noted by @jtl) that it actually does some
12988                  * processing before verifying the timestamp information.
12989                  * We don't take that path here which is why we set
12990                  * the ack_val_set first, do the timestamp and ecn
12991                  * processing, and then look at what we have setup.
12992                  */
12993                 if (ae->ack_val_set == ACK_BEHIND) {
12994                         /*
12995                          * Case B flag reordering, if window is not closed
12996                          * or it could be a keep-alive or persists
12997                          */
12998                         if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
12999                                 counter_u64_add(rack_reorder_seen, 1);
13000                                 rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13001                         }
13002                 } else if (ae->ack_val_set == ACK_DUPACK) {
13003                         /* Case D */
13004
13005                         rack_strike_dupack(rack);
13006                 } else if (ae->ack_val_set == ACK_RWND) {
13007                         /* Case C */
13008
13009                         win_up_req = 1;
13010                         win_upd_ack = ae->ack;
13011                         win_seq = ae->seq;
13012                         the_win = tiwin;
13013                 } else {
13014                         /* Case A */
13015
13016                         if (SEQ_GT(ae->ack, tp->snd_max)) {
13017                                 /*
13018                                  * We just send an ack since the incoming
13019                                  * ack is beyond the largest seq we sent.
13020                                  */
13021                                 if ((tp->t_flags & TF_ACKNOW) == 0) {
13022                                         ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13023                                         if (tp->t_flags && TF_ACKNOW)
13024                                                 rack->r_wanted_output = 1;
13025                                 }
13026                         } else {
13027                                 nsegs++;
13028                                 /* If the window changed setup to update */
13029                                 if (tiwin != tp->snd_wnd) {
13030                                         win_up_req = 1;
13031                                         win_upd_ack = ae->ack;
13032                                         win_seq = ae->seq;
13033                                         the_win = tiwin;
13034                                 }
13035 #ifdef TCP_ACCOUNTING
13036                                 /* Account for the acks */
13037                                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13038                                         tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13039                                 }
13040                                 counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13041                                                 (((ae->ack - high_seq) + segsiz - 1) / segsiz));
13042 #endif
13043                                 high_seq = ae->ack;
13044                                 /* Setup our act_rcv_time */
13045                                 if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13046                                         ts.tv_sec = ae->timestamp / 1000000000;
13047                                         ts.tv_nsec = ae->timestamp % 1000000000;
13048                                         rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13049                                         rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13050                                 } else {
13051                                         rack->r_ctl.act_rcv_time = *tv;
13052                                 }
13053                                 rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13054                         }
13055                 }
13056                 /* And lets be sure to commit the rtt measurements for this ack */
13057                 tcp_rack_xmit_timer_commit(rack, tp);
13058 #ifdef TCP_ACCOUNTING
13059                 rdstc = get_cyclecount();
13060                 if (rdstc > ts_val) {
13061                         counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13062                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13063                                 tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13064                                 if (ae->ack_val_set == ACK_CUMACK)
13065                                         tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13066                         }
13067                 }
13068 #endif
13069         }
13070 #ifdef TCP_ACCOUNTING
13071         ts_val = get_cyclecount();
13072 #endif
13073         acked_amount = acked = (high_seq - tp->snd_una);
13074         if (win_up_req) {
13075                 rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13076         }
13077         if (acked) {
13078                 if (rack->sack_attack_disable == 0)
13079                         rack_do_decay(rack);
13080                 if (acked >= segsiz) {
13081                         /*
13082                          * You only get credit for
13083                          * MSS and greater (and you get extra
13084                          * credit for larger cum-ack moves).
13085                          */
13086                         int ac;
13087
13088                         ac = acked / segsiz;
13089                         rack->r_ctl.ack_count += ac;
13090                         counter_u64_add(rack_ack_total, ac);
13091                 }
13092                 if (rack->r_ctl.ack_count > 0xfff00000) {
13093                         /*
13094                          * reduce the number to keep us under
13095                          * a uint32_t.
13096                          */
13097                         rack->r_ctl.ack_count /= 2;
13098                         rack->r_ctl.sack_count /= 2;
13099                 }
13100                 if (tp->t_flags & TF_NEEDSYN) {
13101                         /*
13102                          * T/TCP: Connection was half-synchronized, and our SYN has
13103                          * been ACK'd (so connection is now fully synchronized).  Go
13104                          * to non-starred state, increment snd_una for ACK of SYN,
13105                          * and check if we can do window scaling.
13106                          */
13107                         tp->t_flags &= ~TF_NEEDSYN;
13108                         tp->snd_una++;
13109                         acked_amount = acked = (high_seq - tp->snd_una);
13110                 }
13111                 if (acked > sbavail(&so->so_snd))
13112                         acked_amount = sbavail(&so->so_snd);
13113 #ifdef NETFLIX_EXP_DETECTION
13114                 /*
13115                  * We only care on a cum-ack move if we are in a sack-disabled
13116                  * state. We have already added in to the ack_count, and we never
13117                  * would disable on a cum-ack move, so we only care to do the
13118                  * detection if it may "undo" it, i.e. we were in disabled already.
13119                  */
13120                 if (rack->sack_attack_disable)
13121                         rack_do_detection(tp, rack, acked_amount, segsiz);
13122 #endif
13123                 if (IN_FASTRECOVERY(tp->t_flags) &&
13124                     (rack->rack_no_prr == 0))
13125                         rack_update_prr(tp, rack, acked_amount, high_seq);
13126                 if (IN_RECOVERY(tp->t_flags)) {
13127                         if (SEQ_LT(high_seq, tp->snd_recover) &&
13128                             (SEQ_LT(high_seq, tp->snd_max))) {
13129                                 tcp_rack_partialack(tp);
13130                         } else {
13131                                 rack_post_recovery(tp, high_seq);
13132                                 recovery = 1;
13133                         }
13134                 }
13135                 /* Handle the rack-log-ack part (sendmap) */
13136                 if ((sbused(&so->so_snd) == 0) &&
13137                     (acked > acked_amount) &&
13138                     (tp->t_state >= TCPS_FIN_WAIT_1) &&
13139                     (tp->t_flags & TF_SENTFIN)) {
13140                         /*
13141                          * We must be sure our fin
13142                          * was sent and acked (we can be
13143                          * in FIN_WAIT_1 without having
13144                          * sent the fin).
13145                          */
13146                         ourfinisacked = 1;
13147                         /*
13148                          * Lets make sure snd_una is updated
13149                          * since most likely acked_amount = 0 (it
13150                          * should be).
13151                          */
13152                         tp->snd_una = high_seq;
13153                 }
13154                 /* Did we make a RTO error? */
13155                 if ((tp->t_flags & TF_PREVVALID) &&
13156                     ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13157                         tp->t_flags &= ~TF_PREVVALID;
13158                         if (tp->t_rxtshift == 1 &&
13159                             (int)(ticks - tp->t_badrxtwin) < 0)
13160                                 rack_cong_signal(tp, CC_RTO_ERR, high_seq);
13161                 }
13162                 /* Handle the data in the socket buffer */
13163                 KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13164                 KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13165                 if (acked_amount > 0) {
13166                         struct mbuf *mfree;
13167
13168                         rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13169                         SOCKBUF_LOCK(&so->so_snd);
13170                         mfree = sbcut_locked(&so->so_snd, acked);
13171                         tp->snd_una = high_seq;
13172                         /* Note we want to hold the sb lock through the sendmap adjust */
13173                         rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13174                         /* Wake up the socket if we have room to write more */
13175                         rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13176                         sowwakeup_locked(so);
13177                         m_freem(mfree);
13178                 }
13179                 /* update progress */
13180                 tp->t_acktime = ticks;
13181                 rack_log_progress_event(rack, tp, tp->t_acktime,
13182                                         PROGRESS_UPDATE, __LINE__);
13183                 /* Clear out shifts and such */
13184                 tp->t_rxtshift = 0;
13185                 RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13186                                    rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13187                 rack->rc_tlp_in_progress = 0;
13188                 rack->r_ctl.rc_tlp_cnt_out = 0;
13189                 /* Send recover and snd_nxt must be dragged along */
13190                 if (SEQ_GT(tp->snd_una, tp->snd_recover))
13191                         tp->snd_recover = tp->snd_una;
13192                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13193                         tp->snd_nxt = tp->snd_una;
13194                 /*
13195                  * If the RXT timer is running we want to
13196                  * stop it, so we can restart a TLP (or new RXT).
13197                  */
13198                 if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13199                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13200 #ifdef NETFLIX_HTTP_LOGGING
13201                 tcp_http_check_for_comp(rack->rc_tp, high_seq);
13202 #endif
13203                 tp->snd_wl2 = high_seq;
13204                 tp->t_dupacks = 0;
13205                 if (under_pacing &&
13206                     (rack->use_fixed_rate == 0) &&
13207                     (rack->in_probe_rtt == 0) &&
13208                     rack->rc_gp_dyn_mul &&
13209                     rack->rc_always_pace) {
13210                         /* Check if we are dragging bottom */
13211                         rack_check_bottom_drag(tp, rack, so, acked);
13212                 }
13213                 if (tp->snd_una == tp->snd_max) {
13214                         tp->t_flags &= ~TF_PREVVALID;
13215                         rack->r_ctl.retran_during_recovery = 0;
13216                         rack->r_ctl.dsack_byte_cnt = 0;
13217                         rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13218                         if (rack->r_ctl.rc_went_idle_time == 0)
13219                                 rack->r_ctl.rc_went_idle_time = 1;
13220                         rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13221                         if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13222                                 tp->t_acktime = 0;
13223                         /* Set so we might enter persists... */
13224                         rack->r_wanted_output = 1;
13225                         rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13226                         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13227                         if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13228                             (sbavail(&so->so_snd) == 0) &&
13229                             (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13230                                 /*
13231                                  * The socket was gone and the
13232                                  * peer sent data (not now in the past), time to
13233                                  * reset him.
13234                                  */
13235                                 rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13236                                 /* tcp_close will kill the inp pre-log the Reset */
13237                                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13238 #ifdef TCP_ACCOUNTING
13239                                 rdstc = get_cyclecount();
13240                                 if (rdstc > ts_val) {
13241                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13242                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13243                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13244                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13245                                         }
13246                                 }
13247 #endif
13248                                 m_freem(m);
13249                                 tp = tcp_close(tp);
13250                                 if (tp == NULL) {
13251 #ifdef TCP_ACCOUNTING
13252                                         sched_unpin();
13253 #endif
13254                                         return (1);
13255                                 }
13256                                 /*
13257                                  * We would normally do drop-with-reset which would
13258                                  * send back a reset. We can't since we don't have
13259                                  * all the needed bits. Instead lets arrange for
13260                                  * a call to tcp_output(). That way since we
13261                                  * are in the closed state we will generate a reset.
13262                                  *
13263                                  * Note if tcp_accounting is on we don't unpin since
13264                                  * we do that after the goto label.
13265                                  */
13266                                 goto send_out_a_rst;
13267                         }
13268                         if ((sbused(&so->so_snd) == 0) &&
13269                             (tp->t_state >= TCPS_FIN_WAIT_1) &&
13270                             (tp->t_flags & TF_SENTFIN)) {
13271                                 /*
13272                                  * If we can't receive any more data, then closing user can
13273                                  * proceed. Starting the timer is contrary to the
13274                                  * specification, but if we don't get a FIN we'll hang
13275                                  * forever.
13276                                  *
13277                                  */
13278                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13279                                         soisdisconnected(so);
13280                                         tcp_timer_activate(tp, TT_2MSL,
13281                                                            (tcp_fast_finwait2_recycle ?
13282                                                             tcp_finwait2_timeout :
13283                                                             TP_MAXIDLE(tp)));
13284                                 }
13285                                 if (ourfinisacked == 0) {
13286                                         /*
13287                                          * We don't change to fin-wait-2 if we have our fin acked
13288                                          * which means we are probably in TCPS_CLOSING.
13289                                          */
13290                                         tcp_state_change(tp, TCPS_FIN_WAIT_2);
13291                                 }
13292                         }
13293                 }
13294                 /* Wake up the socket if we have room to write more */
13295                 if (sbavail(&so->so_snd)) {
13296                         rack->r_wanted_output = 1;
13297                         if (ctf_progress_timeout_check(tp, true)) {
13298                                 rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13299                                                         tp, tick, PROGRESS_DROP, __LINE__);
13300                                 tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
13301                                 /*
13302                                  * We cheat here and don't send a RST, we should send one
13303                                  * when the pacer drops the connection.
13304                                  */
13305 #ifdef TCP_ACCOUNTING
13306                                 rdstc = get_cyclecount();
13307                                 if (rdstc > ts_val) {
13308                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13309                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13310                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13311                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13312                                         }
13313                                 }
13314                                 sched_unpin();
13315 #endif
13316                                 INP_WUNLOCK(rack->rc_inp);
13317                                 m_freem(m);
13318                                 return (1);
13319                         }
13320                 }
13321                 if (ourfinisacked) {
13322                         switch(tp->t_state) {
13323                         case TCPS_CLOSING:
13324 #ifdef TCP_ACCOUNTING
13325                                 rdstc = get_cyclecount();
13326                                 if (rdstc > ts_val) {
13327                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13328                                                         (rdstc - ts_val));
13329                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13330                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13331                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13332                                         }
13333                                 }
13334                                 sched_unpin();
13335 #endif
13336                                 tcp_twstart(tp);
13337                                 m_freem(m);
13338                                 return (1);
13339                                 break;
13340                         case TCPS_LAST_ACK:
13341 #ifdef TCP_ACCOUNTING
13342                                 rdstc = get_cyclecount();
13343                                 if (rdstc > ts_val) {
13344                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13345                                                         (rdstc - ts_val));
13346                                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13347                                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13348                                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13349                                         }
13350                                 }
13351                                 sched_unpin();
13352 #endif
13353                                 tp = tcp_close(tp);
13354                                 ctf_do_drop(m, tp);
13355                                 return (1);
13356                                 break;
13357                         case TCPS_FIN_WAIT_1:
13358 #ifdef TCP_ACCOUNTING
13359                                 rdstc = get_cyclecount();
13360                                 if (rdstc > ts_val) {
13361                                         counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13362                                                         (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                                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13370                                         soisdisconnected(so);
13371                                         tcp_timer_activate(tp, TT_2MSL,
13372                                                            (tcp_fast_finwait2_recycle ?
13373                                                             tcp_finwait2_timeout :
13374                                                             TP_MAXIDLE(tp)));
13375                                 }
13376                                 tcp_state_change(tp, TCPS_FIN_WAIT_2);
13377                                 break;
13378                         default:
13379                                 break;
13380                         }
13381                 }
13382                 if (rack->r_fast_output) {
13383                         /*
13384                          * We re doing fast output.. can we expand that?
13385                          */
13386                         rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13387                 }
13388 #ifdef TCP_ACCOUNTING
13389                 rdstc = get_cyclecount();
13390                 if (rdstc > ts_val) {
13391                         counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13392                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13393                                 tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13394                                 tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13395                         }
13396                 }
13397
13398         } else if (win_up_req) {
13399                 rdstc = get_cyclecount();
13400                 if (rdstc > ts_val) {
13401                         counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13402                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13403                                 tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13404                         }
13405                 }
13406 #endif
13407         }
13408         /* Now is there a next packet, if so we are done */
13409         m_freem(m);
13410         did_out = 0;
13411         if (nxt_pkt) {
13412 #ifdef TCP_ACCOUNTING
13413                 sched_unpin();
13414 #endif
13415                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13416                 return (0);
13417         }
13418         rack_handle_might_revert(tp, rack);
13419         ctf_calc_rwin(so, tp);
13420         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13421         send_out_a_rst:
13422                 (void)tp->t_fb->tfb_tcp_output(tp);
13423                 did_out = 1;
13424         }
13425         rack_free_trim(rack);
13426 #ifdef TCP_ACCOUNTING
13427         sched_unpin();
13428 #endif
13429         rack_timer_audit(tp, rack, &so->so_snd);
13430         rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13431         return (0);
13432 }
13433
13434
13435 static int
13436 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13437     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13438     int32_t nxt_pkt, struct timeval *tv)
13439 {
13440 #ifdef TCP_ACCOUNTING
13441         uint64_t ts_val;
13442 #endif
13443         int32_t thflags, retval, did_out = 0;
13444         int32_t way_out = 0;
13445         uint32_t cts;
13446         uint32_t tiwin;
13447         struct timespec ts;
13448         struct tcpopt to;
13449         struct tcp_rack *rack;
13450         struct rack_sendmap *rsm;
13451         int32_t prev_state = 0;
13452 #ifdef TCP_ACCOUNTING
13453         int ack_val_set = 0xf;
13454 #endif
13455         uint32_t us_cts;
13456         /*
13457          * tv passed from common code is from either M_TSTMP_LRO or
13458          * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13459          */
13460         if (m->m_flags & M_ACKCMP) {
13461                 return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13462         }
13463         if (m->m_flags & M_ACKCMP) {
13464                 panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13465         }
13466         counter_u64_add(rack_proc_non_comp_ack, 1);
13467         thflags = th->th_flags;
13468 #ifdef TCP_ACCOUNTING
13469         sched_pin();
13470         if (thflags & TH_ACK)
13471                 ts_val = get_cyclecount();
13472 #endif
13473         cts = tcp_tv_to_usectick(tv);
13474         rack = (struct tcp_rack *)tp->t_fb_ptr;
13475
13476         if ((m->m_flags & M_TSTMP) ||
13477             (m->m_flags & M_TSTMP_LRO)) {
13478                 mbuf_tstmp2timespec(m, &ts);
13479                 rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13480                 rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13481         } else
13482                 rack->r_ctl.act_rcv_time = *tv;
13483         kern_prefetch(rack, &prev_state);
13484         prev_state = 0;
13485         /*
13486          * Unscale the window into a 32-bit value. For the SYN_SENT state
13487          * the scale is zero.
13488          */
13489         tiwin = th->th_win << tp->snd_scale;
13490         /*
13491          * Parse options on any incoming segment.
13492          */
13493         memset(&to, 0, sizeof(to));
13494         tcp_dooptions(&to, (u_char *)(th + 1),
13495             (th->th_off << 2) - sizeof(struct tcphdr),
13496             (thflags & TH_SYN) ? TO_SYN : 0);
13497 #ifdef TCP_ACCOUNTING
13498         if (thflags & TH_ACK) {
13499                 /*
13500                  * We have a tradeoff here. We can either do what we are
13501                  * doing i.e. pinning to this CPU and then doing the accounting
13502                  * <or> we could do a critical enter, setup the rdtsc and cpu
13503                  * as in below, and then validate we are on the same CPU on
13504                  * exit. I have choosen to not do the critical enter since
13505                  * that often will gain you a context switch, and instead lock
13506                  * us (line above this if) to the same CPU with sched_pin(). This
13507                  * means we may be context switched out for a higher priority
13508                  * interupt but we won't be moved to another CPU.
13509                  *
13510                  * If this occurs (which it won't very often since we most likely
13511                  * are running this code in interupt context and only a higher
13512                  * priority will bump us ... clock?) we will falsely add in
13513                  * to the time the interupt processing time plus the ack processing
13514                  * time. This is ok since its a rare event.
13515                  */
13516                 ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
13517                                                     ctf_fixed_maxseg(tp));
13518         }
13519 #endif
13520         NET_EPOCH_ASSERT();
13521         INP_WLOCK_ASSERT(tp->t_inpcb);
13522         KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
13523             __func__));
13524         KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
13525             __func__));
13526         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13527                 union tcp_log_stackspecific log;
13528                 struct timeval ltv;
13529 #ifdef NETFLIX_HTTP_LOGGING
13530                 struct http_sendfile_track *http_req;
13531
13532                 if (SEQ_GT(th->th_ack, tp->snd_una)) {
13533                         http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
13534                 } else {
13535                         http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
13536                 }
13537 #endif
13538                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13539                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13540                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13541                 if (rack->rack_no_prr == 0)
13542                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13543                 else
13544                         log.u_bbr.flex1 = 0;
13545                 log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13546                 log.u_bbr.use_lt_bw <<= 1;
13547                 log.u_bbr.use_lt_bw |= rack->r_might_revert;
13548                 log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13549                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
13550                 log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
13551                 log.u_bbr.flex3 = m->m_flags;
13552                 log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13553                 log.u_bbr.lost = thflags;
13554                 log.u_bbr.pacing_gain = 0x1;
13555 #ifdef TCP_ACCOUNTING
13556                 log.u_bbr.cwnd_gain = ack_val_set;
13557 #endif
13558                 log.u_bbr.flex7 = 2;
13559                 if (m->m_flags & M_TSTMP) {
13560                         /* Record the hardware timestamp if present */
13561                         mbuf_tstmp2timespec(m, &ts);
13562                         ltv.tv_sec = ts.tv_sec;
13563                         ltv.tv_usec = ts.tv_nsec / 1000;
13564                         log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13565                 } else if (m->m_flags & M_TSTMP_LRO) {
13566                         /* Record the LRO the arrival timestamp */
13567                         mbuf_tstmp2timespec(m, &ts);
13568                         ltv.tv_sec = ts.tv_sec;
13569                         ltv.tv_usec = ts.tv_nsec / 1000;
13570                         log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13571                 }
13572                 log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13573                 /* Log the rcv time */
13574                 log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
13575 #ifdef NETFLIX_HTTP_LOGGING
13576                 log.u_bbr.applimited = tp->t_http_closed;
13577                 log.u_bbr.applimited <<= 8;
13578                 log.u_bbr.applimited |= tp->t_http_open;
13579                 log.u_bbr.applimited <<= 8;
13580                 log.u_bbr.applimited |= tp->t_http_req;
13581                 if (http_req) {
13582                         /* Copy out any client req info */
13583                         /* seconds */
13584                         log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13585                         /* useconds */
13586                         log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13587                         log.u_bbr.rttProp = http_req->timestamp;
13588                         log.u_bbr.cur_del_rate = http_req->start;
13589                         if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13590                                 log.u_bbr.flex8 |= 1;
13591                         } else {
13592                                 log.u_bbr.flex8 |= 2;
13593                                 log.u_bbr.bw_inuse = http_req->end;
13594                         }
13595                         log.u_bbr.flex6 = http_req->start_seq;
13596                         if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13597                                 log.u_bbr.flex8 |= 4;
13598                                 log.u_bbr.epoch = http_req->end_seq;
13599                         }
13600                 }
13601 #endif
13602                 TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
13603                     tlen, &log, true, &ltv);
13604         }
13605         if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
13606                 way_out = 4;
13607                 retval = 0;
13608                 goto done_with_input;
13609         }
13610         /*
13611          * If a segment with the ACK-bit set arrives in the SYN-SENT state
13612          * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
13613          */
13614         if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
13615             (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
13616                 tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13617                 ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13618 #ifdef TCP_ACCOUNTING
13619                 sched_unpin();
13620 #endif
13621                 return (1);
13622         }
13623
13624         /*
13625          * Parse options on any incoming segment.
13626          */
13627         tcp_dooptions(&to, (u_char *)(th + 1),
13628             (th->th_off << 2) - sizeof(struct tcphdr),
13629             (thflags & TH_SYN) ? TO_SYN : 0);
13630
13631         /*
13632          * If timestamps were negotiated during SYN/ACK and a
13633          * segment without a timestamp is received, silently drop
13634          * the segment, unless it is a RST segment or missing timestamps are
13635          * tolerated.
13636          * See section 3.2 of RFC 7323.
13637          */
13638         if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
13639             ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
13640                 way_out = 5;
13641                 retval = 0;
13642                 goto done_with_input;
13643         }
13644
13645         /*
13646          * Segment received on connection. Reset idle time and keep-alive
13647          * timer. XXX: This should be done after segment validation to
13648          * ignore broken/spoofed segs.
13649          */
13650         if  (tp->t_idle_reduce &&
13651              (tp->snd_max == tp->snd_una) &&
13652              ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13653                 counter_u64_add(rack_input_idle_reduces, 1);
13654                 rack_cc_after_idle(rack, tp);
13655         }
13656         tp->t_rcvtime = ticks;
13657 #ifdef STATS
13658         stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
13659 #endif
13660         if (tiwin > rack->r_ctl.rc_high_rwnd)
13661                 rack->r_ctl.rc_high_rwnd = tiwin;
13662         /*
13663          * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
13664          * this to occur after we've validated the segment.
13665          */
13666         if (tp->t_flags2 & TF2_ECN_PERMIT) {
13667                 if (thflags & TH_CWR) {
13668                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13669                         tp->t_flags |= TF_ACKNOW;
13670                 }
13671                 switch (iptos & IPTOS_ECN_MASK) {
13672                 case IPTOS_ECN_CE:
13673                         tp->t_flags2 |= TF2_ECN_SND_ECE;
13674                         KMOD_TCPSTAT_INC(tcps_ecn_ce);
13675                         break;
13676                 case IPTOS_ECN_ECT0:
13677                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13678                         break;
13679                 case IPTOS_ECN_ECT1:
13680                         KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13681                         break;
13682                 }
13683
13684                 /* Process a packet differently from RFC3168. */
13685                 cc_ecnpkt_handler(tp, th, iptos);
13686
13687                 /* Congestion experienced. */
13688                 if (thflags & TH_ECE) {
13689                         rack_cong_signal(tp, CC_ECN, th->th_ack);
13690                 }
13691         }
13692
13693         /*
13694          * If echoed timestamp is later than the current time, fall back to
13695          * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13696          * were used when this connection was established.
13697          */
13698         if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
13699                 to.to_tsecr -= tp->ts_offset;
13700                 if (TSTMP_GT(to.to_tsecr, cts))
13701                         to.to_tsecr = 0;
13702         }
13703
13704         /*
13705          * If its the first time in we need to take care of options and
13706          * verify we can do SACK for rack!
13707          */
13708         if (rack->r_state == 0) {
13709                 /* Should be init'd by rack_init() */
13710                 KASSERT(rack->rc_inp != NULL,
13711                     ("%s: rack->rc_inp unexpectedly NULL", __func__));
13712                 if (rack->rc_inp == NULL) {
13713                         rack->rc_inp = tp->t_inpcb;
13714                 }
13715
13716                 /*
13717                  * Process options only when we get SYN/ACK back. The SYN
13718                  * case for incoming connections is handled in tcp_syncache.
13719                  * According to RFC1323 the window field in a SYN (i.e., a
13720                  * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
13721                  * this is traditional behavior, may need to be cleaned up.
13722                  */
13723                 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
13724                         /* Handle parallel SYN for ECN */
13725                         if (!(thflags & TH_ACK) &&
13726                             ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
13727                             ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
13728                                 tp->t_flags2 |= TF2_ECN_PERMIT;
13729                                 tp->t_flags2 |= TF2_ECN_SND_ECE;
13730                                 TCPSTAT_INC(tcps_ecn_shs);
13731                         }
13732                         if ((to.to_flags & TOF_SCALE) &&
13733                             (tp->t_flags & TF_REQ_SCALE)) {
13734                                 tp->t_flags |= TF_RCVD_SCALE;
13735                                 tp->snd_scale = to.to_wscale;
13736                         } else
13737                                 tp->t_flags &= ~TF_REQ_SCALE;
13738                         /*
13739                          * Initial send window.  It will be updated with the
13740                          * next incoming segment to the scaled value.
13741                          */
13742                         tp->snd_wnd = th->th_win;
13743                         rack_validate_fo_sendwin_up(tp, rack);
13744                         if ((to.to_flags & TOF_TS) &&
13745                             (tp->t_flags & TF_REQ_TSTMP)) {
13746                                 tp->t_flags |= TF_RCVD_TSTMP;
13747                                 tp->ts_recent = to.to_tsval;
13748                                 tp->ts_recent_age = cts;
13749                         } else
13750                                 tp->t_flags &= ~TF_REQ_TSTMP;
13751                         if (to.to_flags & TOF_MSS) {
13752                                 tcp_mss(tp, to.to_mss);
13753                         }
13754                         if ((tp->t_flags & TF_SACK_PERMIT) &&
13755                             (to.to_flags & TOF_SACKPERM) == 0)
13756                                 tp->t_flags &= ~TF_SACK_PERMIT;
13757                         if (IS_FASTOPEN(tp->t_flags)) {
13758                                 if (to.to_flags & TOF_FASTOPEN) {
13759                                         uint16_t mss;
13760
13761                                         if (to.to_flags & TOF_MSS)
13762                                                 mss = to.to_mss;
13763                                         else
13764                                                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
13765                                                         mss = TCP6_MSS;
13766                                                 else
13767                                                         mss = TCP_MSS;
13768                                         tcp_fastopen_update_cache(tp, mss,
13769                                             to.to_tfo_len, to.to_tfo_cookie);
13770                                 } else
13771                                         tcp_fastopen_disable_path(tp);
13772                         }
13773                 }
13774                 /*
13775                  * At this point we are at the initial call. Here we decide
13776                  * if we are doing RACK or not. We do this by seeing if
13777                  * TF_SACK_PERMIT is set and the sack-not-required is clear.
13778                  * The code now does do dup-ack counting so if you don't
13779                  * switch back you won't get rack & TLP, but you will still
13780                  * get this stack.
13781                  */
13782
13783                 if ((rack_sack_not_required == 0) &&
13784                     ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
13785                         tcp_switch_back_to_default(tp);
13786                         (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
13787                             tlen, iptos);
13788 #ifdef TCP_ACCOUNTING
13789                         sched_unpin();
13790 #endif
13791                         return (1);
13792                 }
13793                 tcp_set_hpts(tp->t_inpcb);
13794                 sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
13795         }
13796         if (thflags & TH_FIN)
13797                 tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
13798         us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13799         if ((rack->rc_gp_dyn_mul) &&
13800             (rack->use_fixed_rate == 0) &&
13801             (rack->rc_always_pace)) {
13802                 /* Check in on probertt */
13803                 rack_check_probe_rtt(rack, us_cts);
13804         }
13805         if (rack->forced_ack) {
13806                 uint32_t us_rtt;
13807
13808                 /*
13809                  * A persist or keep-alive was forced out, update our
13810                  * min rtt time. Note we do not worry about lost
13811                  * retransmissions since KEEP-ALIVES and persists
13812                  * are usually way long on times of sending (though
13813                  * if we were really paranoid or worried we could
13814                  * at least use timestamps if available to validate).
13815                  */
13816                 rack->forced_ack = 0;
13817                 us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13818                 if (us_rtt == 0)
13819                         us_rtt = 1;
13820                 rack_log_rtt_upd(tp, rack, us_rtt, 0, NULL, 3);
13821                 rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13822         }
13823         /*
13824          * This is the one exception case where we set the rack state
13825          * always. All other times (timers etc) we must have a rack-state
13826          * set (so we assure we have done the checks above for SACK).
13827          */
13828         rack->r_ctl.rc_rcvtime = cts;
13829         if (rack->r_state != tp->t_state)
13830                 rack_set_state(tp, rack);
13831         if (SEQ_GT(th->th_ack, tp->snd_una) &&
13832             (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
13833                 kern_prefetch(rsm, &prev_state);
13834         prev_state = rack->r_state;
13835         rack_clear_rate_sample(rack);
13836         retval = (*rack->r_substate) (m, th, so,
13837             tp, &to, drop_hdrlen,
13838             tlen, tiwin, thflags, nxt_pkt, iptos);
13839 #ifdef INVARIANTS
13840         if ((retval == 0) &&
13841             (tp->t_inpcb == NULL)) {
13842                 panic("retval:%d tp:%p t_inpcb:NULL state:%d",
13843                     retval, tp, prev_state);
13844         }
13845 #endif
13846         if (retval == 0) {
13847                 /*
13848                  * If retval is 1 the tcb is unlocked and most likely the tp
13849                  * is gone.
13850                  */
13851                 INP_WLOCK_ASSERT(tp->t_inpcb);
13852                 if ((rack->rc_gp_dyn_mul) &&
13853                     (rack->rc_always_pace) &&
13854                     (rack->use_fixed_rate == 0) &&
13855                     rack->in_probe_rtt &&
13856                     (rack->r_ctl.rc_time_probertt_starts == 0)) {
13857                         /*
13858                          * If we are going for target, lets recheck before
13859                          * we output.
13860                          */
13861                         rack_check_probe_rtt(rack, us_cts);
13862                 }
13863                 if (rack->set_pacing_done_a_iw == 0) {
13864                         /* How much has been acked? */
13865                         if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
13866                                 /* We have enough to set in the pacing segment size */
13867                                 rack->set_pacing_done_a_iw = 1;
13868                                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
13869                         }
13870                 }
13871                 tcp_rack_xmit_timer_commit(rack, tp);
13872 #ifdef TCP_ACCOUNTING
13873                 /*
13874                  * If we set the ack_val_se to what ack processing we are doing
13875                  * we also want to track how many cycles we burned. Note
13876                  * the bits after tcp_output we let be "free". This is because
13877                  * we are also tracking the tcp_output times as well. Note the
13878                  * use of 0xf here since we only have 11 counter (0 - 0xa) and
13879                  * 0xf cannot be returned and is what we initialize it too to
13880                  * indicate we are not doing the tabulations.
13881                  */
13882                 if (ack_val_set != 0xf) {
13883                         uint64_t crtsc;
13884
13885                         crtsc = get_cyclecount();
13886                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13887                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13888                                 tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
13889                         }
13890                 }
13891 #endif
13892                 if (nxt_pkt == 0) {
13893                         if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13894 do_output_now:
13895                                 did_out = 1;
13896                                 (void)tp->t_fb->tfb_tcp_output(tp);
13897                         }
13898                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
13899                         rack_free_trim(rack);
13900                 }
13901                 if ((nxt_pkt == 0) &&
13902                     ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
13903                     (SEQ_GT(tp->snd_max, tp->snd_una) ||
13904                      (tp->t_flags & TF_DELACK) ||
13905                      ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13906                       (tp->t_state <= TCPS_CLOSING)))) {
13907                         /* We could not send (probably in the hpts but stopped the timer earlier)? */
13908                         if ((tp->snd_max == tp->snd_una) &&
13909                             ((tp->t_flags & TF_DELACK) == 0) &&
13910                             (rack->rc_inp->inp_in_hpts) &&
13911                             (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13912                                 /* keep alive not needed if we are hptsi output yet */
13913                                 ;
13914                         } else {
13915                                 int late = 0;
13916                                 if (rack->rc_inp->inp_in_hpts) {
13917                                         if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13918                                                 us_cts = tcp_get_usecs(NULL);
13919                                                 if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13920                                                         rack->r_early = 1;
13921                                                         rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13922                                                 } else
13923                                                         late = 1;
13924                                                 rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13925                                         }
13926                                         tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
13927                                 }
13928                                 if (late && (did_out == 0)) {
13929                                         /*
13930                                          * We are late in the sending
13931                                          * and we did not call the output
13932                                          * (this probably should not happen).
13933                                          */
13934                                         goto do_output_now;
13935                                 }
13936                                 rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13937                         }
13938                         way_out = 1;
13939                 } else if (nxt_pkt == 0) {
13940                         /* Do we have the correct timer running? */
13941                         rack_timer_audit(tp, rack, &so->so_snd);
13942                         way_out = 2;
13943                 }
13944         done_with_input:
13945                 rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, m->m_pkthdr.lro_nsegs));
13946                 if (did_out)
13947                         rack->r_wanted_output = 0;
13948 #ifdef INVARIANTS
13949                 if (tp->t_inpcb == NULL) {
13950                         panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
13951                               did_out,
13952                               retval, tp, prev_state);
13953                 }
13954 #endif
13955 #ifdef TCP_ACCOUNTING
13956         } else {
13957                 /*
13958                  * Track the time (see above).
13959                  */
13960                 if (ack_val_set != 0xf) {
13961                         uint64_t crtsc;
13962
13963                         crtsc = get_cyclecount();
13964                         counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
13965                         /*
13966                          * Note we *DO NOT* increment the per-tcb counters since
13967                          * in the else the TP may be gone!!
13968                          */
13969                 }
13970 #endif
13971         }
13972 #ifdef TCP_ACCOUNTING
13973         sched_unpin();
13974 #endif
13975         return (retval);
13976 }
13977
13978 void
13979 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
13980     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
13981 {
13982         struct timeval tv;
13983
13984         /* First lets see if we have old packets */
13985         if (tp->t_in_pkt) {
13986                 if (ctf_do_queued_segments(so, tp, 1)) {
13987                         m_freem(m);
13988                         return;
13989                 }
13990         }
13991         if (m->m_flags & M_TSTMP_LRO) {
13992                 tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
13993                 tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
13994         } else {
13995                 /* Should not be should we kassert instead? */
13996                 tcp_get_usecs(&tv);
13997         }
13998         if (rack_do_segment_nounlock(m, th, so, tp,
13999                                      drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14000                 tcp_handle_wakeup(tp, so);
14001                 INP_WUNLOCK(tp->t_inpcb);
14002         }
14003 }
14004
14005 struct rack_sendmap *
14006 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14007 {
14008         struct rack_sendmap *rsm = NULL;
14009         int32_t idx;
14010         uint32_t srtt = 0, thresh = 0, ts_low = 0;
14011
14012         /* Return the next guy to be re-transmitted */
14013         if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14014                 return (NULL);
14015         }
14016         if (tp->t_flags & TF_SENTFIN) {
14017                 /* retran the end FIN? */
14018                 return (NULL);
14019         }
14020         /* ok lets look at this one */
14021         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14022         if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14023                 goto check_it;
14024         }
14025         rsm = rack_find_lowest_rsm(rack);
14026         if (rsm == NULL) {
14027                 return (NULL);
14028         }
14029 check_it:
14030         if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14031             (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14032                 /*
14033                  * No sack so we automatically do the 3 strikes and
14034                  * retransmit (no rack timer would be started).
14035                  */
14036
14037                 return (rsm);
14038         }
14039         if (rsm->r_flags & RACK_ACKED) {
14040                 return (NULL);
14041         }
14042         if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14043             (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14044                 /* Its not yet ready */
14045                 return (NULL);
14046         }
14047         srtt = rack_grab_rtt(tp, rack);
14048         idx = rsm->r_rtr_cnt - 1;
14049         ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14050         thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14051         if ((tsused == ts_low) ||
14052             (TSTMP_LT(tsused, ts_low))) {
14053                 /* No time since sending */
14054                 return (NULL);
14055         }
14056         if ((tsused - ts_low) < thresh) {
14057                 /* It has not been long enough yet */
14058                 return (NULL);
14059         }
14060         if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14061             ((rsm->r_flags & RACK_SACK_PASSED) &&
14062              (rack->sack_attack_disable == 0))) {
14063                 /*
14064                  * We have passed the dup-ack threshold <or>
14065                  * a SACK has indicated this is missing.
14066                  * Note that if you are a declared attacker
14067                  * it is only the dup-ack threshold that
14068                  * will cause retransmits.
14069                  */
14070                 /* log retransmit reason */
14071                 rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14072                 rack->r_fast_output = 0;
14073                 return (rsm);
14074         }
14075         return (NULL);
14076 }
14077
14078 static void
14079 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14080                            uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14081                            int line, struct rack_sendmap *rsm)
14082 {
14083         if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14084                 union tcp_log_stackspecific log;
14085                 struct timeval tv;
14086
14087                 memset(&log, 0, sizeof(log));
14088                 log.u_bbr.flex1 = slot;
14089                 log.u_bbr.flex2 = len;
14090                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14091                 log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14092                 log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14093                 log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14094                 log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14095                 log.u_bbr.use_lt_bw <<= 1;
14096                 log.u_bbr.use_lt_bw |= rack->r_late;
14097                 log.u_bbr.use_lt_bw <<= 1;
14098                 log.u_bbr.use_lt_bw |= rack->r_early;
14099                 log.u_bbr.use_lt_bw <<= 1;
14100                 log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14101                 log.u_bbr.use_lt_bw <<= 1;
14102                 log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14103                 log.u_bbr.use_lt_bw <<= 1;
14104                 log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14105                 log.u_bbr.use_lt_bw <<= 1;
14106                 log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14107                 log.u_bbr.use_lt_bw <<= 1;
14108                 log.u_bbr.use_lt_bw |= rack->gp_ready;
14109                 log.u_bbr.pkt_epoch = line;
14110                 log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14111                 log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14112                 log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14113                 log.u_bbr.bw_inuse = bw_est;
14114                 log.u_bbr.delRate = bw;
14115                 if (rack->r_ctl.gp_bw == 0)
14116                         log.u_bbr.cur_del_rate = 0;
14117                 else
14118                         log.u_bbr.cur_del_rate = rack_get_bw(rack);
14119                 log.u_bbr.rttProp = len_time;
14120                 log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14121                 log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14122                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14123                 if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14124                         /* We are in slow start */
14125                         log.u_bbr.flex7 = 1;
14126                 } else {
14127                         /* we are on congestion avoidance */
14128                         log.u_bbr.flex7 = 0;
14129                 }
14130                 log.u_bbr.flex8 = method;
14131                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14132                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14133                 log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14134                 log.u_bbr.cwnd_gain <<= 1;
14135                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14136                 log.u_bbr.cwnd_gain <<= 1;
14137                 log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14138                 TCP_LOG_EVENTP(rack->rc_tp, NULL,
14139                     &rack->rc_inp->inp_socket->so_rcv,
14140                     &rack->rc_inp->inp_socket->so_snd,
14141                     BBR_LOG_HPTSI_CALC, 0,
14142                     0, &log, false, &tv);
14143         }
14144 }
14145
14146 static uint32_t
14147 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14148 {
14149         uint32_t new_tso, user_max;
14150
14151         user_max = rack->rc_user_set_max_segs * mss;
14152         if (rack->rc_force_max_seg) {
14153                 return (user_max);
14154         }
14155         if (rack->use_fixed_rate &&
14156             ((rack->r_ctl.crte == NULL) ||
14157              (bw != rack->r_ctl.crte->rate))) {
14158                 /* Use the user mss since we are not exactly matched */
14159                 return (user_max);
14160         }
14161         new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14162         if (new_tso > user_max)
14163                 new_tso = user_max;
14164         return (new_tso);
14165 }
14166
14167 static int32_t
14168 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)
14169 {
14170         uint64_t lentim, fill_bw;
14171
14172         /* Lets first see if we are full, if so continue with normal rate */
14173         rack->r_via_fill_cw = 0;
14174         if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14175                 return (slot);
14176         if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14177                 return (slot);
14178         if (rack->r_ctl.rc_last_us_rtt == 0)
14179                 return (slot);
14180         if (rack->rc_pace_fill_if_rttin_range &&
14181             (rack->r_ctl.rc_last_us_rtt >=
14182              (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14183                 /* The rtt is huge, N * smallest, lets not fill */
14184                 return (slot);
14185         }
14186         /*
14187          * first lets calculate the b/w based on the last us-rtt
14188          * and the sndwnd.
14189          */
14190         fill_bw = rack->r_ctl.cwnd_to_use;
14191         /* Take the rwnd if its smaller */
14192         if (fill_bw > rack->rc_tp->snd_wnd)
14193                 fill_bw = rack->rc_tp->snd_wnd;
14194         if (rack->r_fill_less_agg) {
14195                 /*
14196                  * Now take away the inflight (this will reduce our
14197                  * aggressiveness and yeah, if we get that much out in 1RTT
14198                  * we will have had acks come back and still be behind).
14199                  */
14200                 fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14201         }
14202         /* Now lets make it into a b/w */
14203         fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14204         fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14205         /* We are below the min b/w */
14206         if (non_paced)
14207                 *rate_wanted = fill_bw;
14208         if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14209                 return (slot);
14210         if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14211                 fill_bw = rack->r_ctl.bw_rate_cap;
14212         rack->r_via_fill_cw = 1;
14213         if (rack->r_rack_hw_rate_caps &&
14214             (rack->r_ctl.crte != NULL)) {
14215                 uint64_t high_rate;
14216
14217                 high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14218                 if (fill_bw > high_rate) {
14219                         /* We are capping bw at the highest rate table entry */
14220                         if (*rate_wanted > high_rate) {
14221                                 /* The original rate was also capped */
14222                                 rack->r_via_fill_cw = 0;
14223                         }
14224                         rack_log_hdwr_pacing(rack,
14225                                              fill_bw, high_rate, __LINE__,
14226                                              0, 3);
14227                         fill_bw = high_rate;
14228                         if (capped)
14229                                 *capped = 1;
14230                 }
14231         } else if ((rack->r_ctl.crte == NULL) &&
14232                    (rack->rack_hdrw_pacing == 0) &&
14233                    (rack->rack_hdw_pace_ena) &&
14234                    rack->r_rack_hw_rate_caps &&
14235                    (rack->rack_attempt_hdwr_pace == 0) &&
14236                    (rack->rc_inp->inp_route.ro_nh != NULL) &&
14237                    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14238                 /*
14239                  * Ok we may have a first attempt that is greater than our top rate
14240                  * lets check.
14241                  */
14242                 uint64_t high_rate;
14243
14244                 high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14245                 if (high_rate) {
14246                         if (fill_bw > high_rate) {
14247                                 fill_bw = high_rate;
14248                                 if (capped)
14249                                         *capped = 1;
14250                         }
14251                 }
14252         }
14253         /*
14254          * Ok fill_bw holds our mythical b/w to fill the cwnd
14255          * in a rtt, what does that time wise equate too?
14256          */
14257         lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14258         lentim /= fill_bw;
14259         *rate_wanted = fill_bw;
14260         if (non_paced || (lentim < slot)) {
14261                 rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14262                                            0, lentim, 12, __LINE__, NULL);
14263                 return ((int32_t)lentim);
14264         } else
14265                 return (slot);
14266 }
14267
14268 static int32_t
14269 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14270 {
14271         struct rack_sendmap *lrsm;
14272         int32_t slot = 0;
14273         int can_start_hw_pacing = 1;
14274         int err;
14275
14276         if (rack->rc_always_pace == 0) {
14277                 /*
14278                  * We use the most optimistic possible cwnd/srtt for
14279                  * sending calculations. This will make our
14280                  * calculation anticipate getting more through
14281                  * quicker then possible. But thats ok we don't want
14282                  * the peer to have a gap in data sending.
14283                  */
14284                 uint32_t srtt, cwnd, tr_perms = 0;
14285                 int32_t reduce = 0;
14286
14287         old_method:
14288                 /*
14289                  * We keep no precise pacing with the old method
14290                  * instead we use the pacer to mitigate bursts.
14291                  */
14292                 if (rack->r_ctl.rc_rack_min_rtt)
14293                         srtt = rack->r_ctl.rc_rack_min_rtt;
14294                 else
14295                         srtt = max(tp->t_srtt, 1);
14296                 if (rack->r_ctl.rc_rack_largest_cwnd)
14297                         cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14298                 else
14299                         cwnd = rack->r_ctl.cwnd_to_use;
14300                 /* Inflate cwnd by 1000 so srtt of usecs is in ms */
14301                 tr_perms = (cwnd * 1000) / srtt;
14302                 if (tr_perms == 0) {
14303                         tr_perms = ctf_fixed_maxseg(tp);
14304                 }
14305                 /*
14306                  * Calculate how long this will take to drain, if
14307                  * the calculation comes out to zero, thats ok we
14308                  * will use send_a_lot to possibly spin around for
14309                  * more increasing tot_len_this_send to the point
14310                  * that its going to require a pace, or we hit the
14311                  * cwnd. Which in that case we are just waiting for
14312                  * a ACK.
14313                  */
14314                 slot = len / tr_perms;
14315                 /* Now do we reduce the time so we don't run dry? */
14316                 if (slot && rack_slot_reduction) {
14317                         reduce = (slot / rack_slot_reduction);
14318                         if (reduce < slot) {
14319                                 slot -= reduce;
14320                         } else
14321                                 slot = 0;
14322                 }
14323                 slot *= HPTS_USEC_IN_MSEC;
14324                 if (rsm == NULL) {
14325                         /*
14326                          * We always consider ourselves app limited with old style
14327                          * that are not retransmits. This could be the initial
14328                          * measurement, but thats ok its all setup and specially
14329                          * handled. If another send leaks out, then that too will
14330                          * be mark app-limited.
14331                          */
14332                         lrsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14333                         if (lrsm && ((lrsm->r_flags & RACK_APP_LIMITED) == 0)) {
14334                                 rack->r_ctl.rc_first_appl = lrsm;
14335                                 lrsm->r_flags |= RACK_APP_LIMITED;
14336                                 rack->r_ctl.rc_app_limited_cnt++;
14337                         }
14338                 }
14339                 if (rack->rc_pace_to_cwnd) {
14340                         uint64_t rate_wanted = 0;
14341
14342                         slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14343                         rack->rc_ack_can_sendout_data = 1;
14344                         rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL);
14345                 } else
14346                         rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL);
14347         } else {
14348                 uint64_t bw_est, res, lentim, rate_wanted;
14349                 uint32_t orig_val, srtt, segs, oh;
14350                 int capped = 0;
14351                 int prev_fill;
14352
14353                 if ((rack->r_rr_config == 1) && rsm) {
14354                         return (rack->r_ctl.rc_min_to);
14355                 }
14356                 if (rack->use_fixed_rate) {
14357                         rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14358                 } else if ((rack->r_ctl.init_rate == 0) &&
14359 #ifdef NETFLIX_PEAKRATE
14360                            (rack->rc_tp->t_maxpeakrate == 0) &&
14361 #endif
14362                            (rack->r_ctl.gp_bw == 0)) {
14363                         /* no way to yet do an estimate */
14364                         bw_est = rate_wanted = 0;
14365                 } else {
14366                         bw_est = rack_get_bw(rack);
14367                         rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14368                 }
14369                 if ((bw_est == 0) || (rate_wanted == 0) ||
14370                     ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14371                         /*
14372                          * No way yet to make a b/w estimate or
14373                          * our raise is set incorrectly.
14374                          */
14375                         goto old_method;
14376                 }
14377                 /* We need to account for all the overheads */
14378                 segs = (len + segsiz - 1) / segsiz;
14379                 /*
14380                  * We need the diff between 1514 bytes (e-mtu with e-hdr)
14381                  * and how much data we put in each packet. Yes this
14382                  * means we may be off if we are larger than 1500 bytes
14383                  * or smaller. But this just makes us more conservative.
14384                  */
14385                 if (rack_hw_rate_min &&
14386                     (bw_est < rack_hw_rate_min))
14387                         can_start_hw_pacing = 0;
14388                 if (ETHERNET_SEGMENT_SIZE > segsiz)
14389                         oh = ETHERNET_SEGMENT_SIZE - segsiz;
14390                 else
14391                         oh = 0;
14392                 segs *= oh;
14393                 lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14394                 res = lentim / rate_wanted;
14395                 slot = (uint32_t)res;
14396                 orig_val = rack->r_ctl.rc_pace_max_segs;
14397                 if (rack->r_ctl.crte == NULL) {
14398                         /*
14399                          * Only do this if we are not hardware pacing
14400                          * since if we are doing hw-pacing below we will
14401                          * set make a call after setting up or changing
14402                          * the rate.
14403                          */
14404                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14405                 } else if (rack->rc_inp->inp_snd_tag == NULL) {
14406                         /*
14407                          * We lost our rate somehow, this can happen
14408                          * if the interface changed underneath us.
14409                          */
14410                         tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14411                         rack->r_ctl.crte = NULL;
14412                         /* Lets re-allow attempting to setup pacing */
14413                         rack->rack_hdrw_pacing = 0;
14414                         rack->rack_attempt_hdwr_pace = 0;
14415                         rack_log_hdwr_pacing(rack,
14416                                              rate_wanted, bw_est, __LINE__,
14417                                              0, 6);
14418                 }
14419                 /* Did we change the TSO size, if so log it */
14420                 if (rack->r_ctl.rc_pace_max_segs != orig_val)
14421                         rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL);
14422                 prev_fill = rack->r_via_fill_cw;
14423                 if ((rack->rc_pace_to_cwnd) &&
14424                     (capped == 0) &&
14425                     (rack->use_fixed_rate == 0) &&
14426                     (rack->in_probe_rtt == 0) &&
14427                     (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14428                         /*
14429                          * We want to pace at our rate *or* faster to
14430                          * fill the cwnd to the max if its not full.
14431                          */
14432                         slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14433                 }
14434                 if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14435                     (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14436                         if ((rack->rack_hdw_pace_ena) &&
14437                             (can_start_hw_pacing > 0) &&
14438                             (rack->rack_hdrw_pacing == 0) &&
14439                             (rack->rack_attempt_hdwr_pace == 0)) {
14440                                 /*
14441                                  * Lets attempt to turn on hardware pacing
14442                                  * if we can.
14443                                  */
14444                                 rack->rack_attempt_hdwr_pace = 1;
14445                                 rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14446                                                                        rack->rc_inp->inp_route.ro_nh->nh_ifp,
14447                                                                        rate_wanted,
14448                                                                        RS_PACING_GEQ,
14449                                                                        &err, &rack->r_ctl.crte_prev_rate);
14450                                 if (rack->r_ctl.crte) {
14451                                         rack->rack_hdrw_pacing = 1;
14452                                         rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14453                                                                                                  0, rack->r_ctl.crte,
14454                                                                                                  NULL);
14455                                         rack_log_hdwr_pacing(rack,
14456                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14457                                                              err, 0);
14458                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14459                                 } else {
14460                                         counter_u64_add(rack_hw_pace_init_fail, 1);
14461                                 }
14462                         } else if (rack->rack_hdrw_pacing &&
14463                                    (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14464                                 /* Do we need to adjust our rate? */
14465                                 const struct tcp_hwrate_limit_table *nrte;
14466
14467                                 if (rack->r_up_only &&
14468                                     (rate_wanted < rack->r_ctl.crte->rate)) {
14469                                         /**
14470                                          * We have four possible states here
14471                                          * having to do with the previous time
14472                                          * and this time.
14473                                          *   previous  |  this-time
14474                                          * A)     0      |     0   -- fill_cw not in the picture
14475                                          * B)     1      |     0   -- we were doing a fill-cw but now are not
14476                                          * C)     1      |     1   -- all rates from fill_cw
14477                                          * D)     0      |     1   -- we were doing non-fill and now we are filling
14478                                          *
14479                                          * For case A, C and D we don't allow a drop. But for
14480                                          * case B where we now our on our steady rate we do
14481                                          * allow a drop.
14482                                          *
14483                                          */
14484                                         if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14485                                                 goto done_w_hdwr;
14486                                 }
14487                                 if ((rate_wanted > rack->r_ctl.crte->rate) ||
14488                                     (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14489                                         if (rack_hw_rate_to_low &&
14490                                             (bw_est < rack_hw_rate_to_low)) {
14491                                                 /*
14492                                                  * The pacing rate is too low for hardware, but
14493                                                  * do allow hardware pacing to be restarted.
14494                                                  */
14495                                                 rack_log_hdwr_pacing(rack,
14496                                                              bw_est, rack->r_ctl.crte->rate, __LINE__,
14497                                                              0, 5);
14498                                                 tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14499                                                 rack->r_ctl.crte = NULL;
14500                                                 rack->rack_attempt_hdwr_pace = 0;
14501                                                 rack->rack_hdrw_pacing = 0;
14502                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14503                                                 goto done_w_hdwr;
14504                                         }
14505                                         nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
14506                                                                    rack->rc_tp,
14507                                                                    rack->rc_inp->inp_route.ro_nh->nh_ifp,
14508                                                                    rate_wanted,
14509                                                                    RS_PACING_GEQ,
14510                                                                    &err, &rack->r_ctl.crte_prev_rate);
14511                                         if (nrte == NULL) {
14512                                                 /* Lost the rate */
14513                                                 rack->rack_hdrw_pacing = 0;
14514                                                 rack->r_ctl.crte = NULL;
14515                                                 rack_log_hdwr_pacing(rack,
14516                                                                      rate_wanted, 0, __LINE__,
14517                                                                      err, 1);
14518                                                 rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14519                                                 counter_u64_add(rack_hw_pace_lost, 1);
14520                                         } else if (nrte != rack->r_ctl.crte) {
14521                                                 rack->r_ctl.crte = nrte;
14522                                                 rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
14523                                                                                                          segsiz, 0,
14524                                                                                                          rack->r_ctl.crte,
14525                                                                                                          NULL);
14526                                                 rack_log_hdwr_pacing(rack,
14527                                                                      rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14528                                                                      err, 2);
14529                                                 rack->r_ctl.last_hw_bw_req = rate_wanted;
14530                                         }
14531                                 } else {
14532                                         /* We just need to adjust the segment size */
14533                                         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
14534                                         rack_log_hdwr_pacing(rack,
14535                                                              rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14536                                                              0, 4);
14537                                         rack->r_ctl.last_hw_bw_req = rate_wanted;
14538                                 }
14539                         }
14540                 }
14541                 if ((rack->r_ctl.crte != NULL) &&
14542                     (rack->r_ctl.crte->rate == rate_wanted)) {
14543                         /*
14544                          * We need to add a extra if the rates
14545                          * are exactly matched. The idea is
14546                          * we want the software to make sure the
14547                          * queue is empty before adding more, this
14548                          * gives us N MSS extra pace times where
14549                          * N is our sysctl
14550                          */
14551                         slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
14552                 }
14553 done_w_hdwr:
14554                 if (rack_limit_time_with_srtt &&
14555                     (rack->use_fixed_rate == 0) &&
14556 #ifdef NETFLIX_PEAKRATE
14557                     (rack->rc_tp->t_maxpeakrate == 0) &&
14558 #endif
14559                     (rack->rack_hdrw_pacing == 0)) {
14560                         /*
14561                          * Sanity check, we do not allow the pacing delay
14562                          * to be longer than the SRTT of the path. If it is
14563                          * a slow path, then adding a packet should increase
14564                          * the RTT and compensate for this i.e. the srtt will
14565                          * be greater so the allowed pacing time will be greater.
14566                          *
14567                          * Note this restriction is not for where a peak rate
14568                          * is set, we are doing fixed pacing or hardware pacing.
14569                          */
14570                         if (rack->rc_tp->t_srtt)
14571                                 srtt = rack->rc_tp->t_srtt;
14572                         else
14573                                 srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;    /* its in ms convert */
14574                         if (srtt < slot) {
14575                                 rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL);
14576                                 slot = srtt;
14577                         }
14578                 }
14579                 rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm);
14580         }
14581         if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
14582                 /*
14583                  * If this rate is seeing enobufs when it
14584                  * goes to send then either the nic is out
14585                  * of gas or we are mis-estimating the time
14586                  * somehow and not letting the queue empty
14587                  * completely. Lets add to the pacing time.
14588                  */
14589                 int hw_boost_delay;
14590
14591                 hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
14592                 if (hw_boost_delay > rack_enobuf_hw_max)
14593                         hw_boost_delay = rack_enobuf_hw_max;
14594                 else if (hw_boost_delay < rack_enobuf_hw_min)
14595                         hw_boost_delay = rack_enobuf_hw_min;
14596                 slot += hw_boost_delay;
14597         }
14598         if (slot)
14599                 counter_u64_add(rack_calc_nonzero, 1);
14600         else
14601                 counter_u64_add(rack_calc_zero, 1);
14602         return (slot);
14603 }
14604
14605 static void
14606 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
14607     tcp_seq startseq, uint32_t sb_offset)
14608 {
14609         struct rack_sendmap *my_rsm = NULL;
14610         struct rack_sendmap fe;
14611
14612         if (tp->t_state < TCPS_ESTABLISHED) {
14613                 /*
14614                  * We don't start any measurements if we are
14615                  * not at least established.
14616                  */
14617                 return;
14618         }
14619         tp->t_flags |= TF_GPUTINPROG;
14620         rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
14621         rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
14622         tp->gput_seq = startseq;
14623         rack->app_limited_needs_set = 0;
14624         if (rack->in_probe_rtt)
14625                 rack->measure_saw_probe_rtt = 1;
14626         else if ((rack->measure_saw_probe_rtt) &&
14627                  (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
14628                 rack->measure_saw_probe_rtt = 0;
14629         if (rack->rc_gp_filled)
14630                 tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14631         else {
14632                 /* Special case initial measurement */
14633                 struct timeval tv;
14634
14635                 tp->gput_ts = tcp_get_usecs(&tv);
14636                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14637         }
14638         /*
14639          * We take a guess out into the future,
14640          * if we have no measurement and no
14641          * initial rate, we measure the first
14642          * initial-windows worth of data to
14643          * speed up getting some GP measurement and
14644          * thus start pacing.
14645          */
14646         if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
14647                 rack->app_limited_needs_set = 1;
14648                 tp->gput_ack = startseq + max(rc_init_window(rack),
14649                                               (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
14650                 rack_log_pacing_delay_calc(rack,
14651                                            tp->gput_seq,
14652                                            tp->gput_ack,
14653                                            0,
14654                                            tp->gput_ts,
14655                                            rack->r_ctl.rc_app_limited_cnt,
14656                                            9,
14657                                            __LINE__, NULL);
14658                 return;
14659         }
14660         if (sb_offset) {
14661                 /*
14662                  * We are out somewhere in the sb
14663                  * can we use the already outstanding data?
14664                  */
14665
14666                 if (rack->r_ctl.rc_app_limited_cnt == 0) {
14667                         /*
14668                          * Yes first one is good and in this case
14669                          * the tp->gput_ts is correctly set based on
14670                          * the last ack that arrived (no need to
14671                          * set things up when an ack comes in).
14672                          */
14673                         my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
14674                         if ((my_rsm == NULL) ||
14675                             (my_rsm->r_rtr_cnt != 1)) {
14676                                 /* retransmission? */
14677                                 goto use_latest;
14678                         }
14679                 } else {
14680                         if (rack->r_ctl.rc_first_appl == NULL) {
14681                                 /*
14682                                  * If rc_first_appl is NULL
14683                                  * then the cnt should be 0.
14684                                  * This is probably an error, maybe
14685                                  * a KASSERT would be approprate.
14686                                  */
14687                                 goto use_latest;
14688                         }
14689                         /*
14690                          * If we have a marker pointer to the last one that is
14691                          * app limited we can use that, but we need to set
14692                          * things up so that when it gets ack'ed we record
14693                          * the ack time (if its not already acked).
14694                          */
14695                         rack->app_limited_needs_set = 1;
14696                         /*
14697                          * We want to get to the rsm that is either
14698                          * next with space i.e. over 1 MSS or the one
14699                          * after that (after the app-limited).
14700                          */
14701                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14702                                          rack->r_ctl.rc_first_appl);
14703                         if (my_rsm) {
14704                                 if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
14705                                         /* Have to use the next one */
14706                                         my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
14707                                                          my_rsm);
14708                                 else {
14709                                         /* Use after the first MSS of it is acked */
14710                                         tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
14711                                         goto start_set;
14712                                 }
14713                         }
14714                         if ((my_rsm == NULL) ||
14715                             (my_rsm->r_rtr_cnt != 1)) {
14716                                 /*
14717                                  * Either its a retransmit or
14718                                  * the last is the app-limited one.
14719                                  */
14720                                 goto use_latest;
14721                         }
14722                 }
14723                 tp->gput_seq = my_rsm->r_start;
14724 start_set:
14725                 if (my_rsm->r_flags & RACK_ACKED) {
14726                         /*
14727                          * This one has been acked use the arrival ack time
14728                          */
14729                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14730                         rack->app_limited_needs_set = 0;
14731                 }
14732                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14733                 tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
14734                 rack_log_pacing_delay_calc(rack,
14735                                            tp->gput_seq,
14736                                            tp->gput_ack,
14737                                            (uint64_t)my_rsm,
14738                                            tp->gput_ts,
14739                                            rack->r_ctl.rc_app_limited_cnt,
14740                                            9,
14741                                            __LINE__, NULL);
14742                 return;
14743         }
14744
14745 use_latest:
14746         /*
14747          * We don't know how long we may have been
14748          * idle or if this is the first-send. Lets
14749          * setup the flag so we will trim off
14750          * the first ack'd data so we get a true
14751          * measurement.
14752          */
14753         rack->app_limited_needs_set = 1;
14754         tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
14755         /* Find this guy so we can pull the send time */
14756         fe.r_start = startseq;
14757         my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
14758         if (my_rsm) {
14759                 rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
14760                 if (my_rsm->r_flags & RACK_ACKED) {
14761                         /*
14762                          * Unlikely since its probably what was
14763                          * just transmitted (but I am paranoid).
14764                          */
14765                         tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
14766                         rack->app_limited_needs_set = 0;
14767                 }
14768                 if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
14769                         /* This also is unlikely */
14770                         tp->gput_seq = my_rsm->r_start;
14771                 }
14772         } else {
14773                 /*
14774                  * TSNH unless we have some send-map limit,
14775                  * and even at that it should not be hitting
14776                  * that limit (we should have stopped sending).
14777                  */
14778                 struct timeval tv;
14779
14780                 microuptime(&tv);
14781                 rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
14782         }
14783         rack_log_pacing_delay_calc(rack,
14784                                    tp->gput_seq,
14785                                    tp->gput_ack,
14786                                    (uint64_t)my_rsm,
14787                                    tp->gput_ts,
14788                                    rack->r_ctl.rc_app_limited_cnt,
14789                                    9, __LINE__, NULL);
14790 }
14791
14792 static inline uint32_t
14793 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
14794     uint32_t avail, int32_t sb_offset)
14795 {
14796         uint32_t len;
14797         uint32_t sendwin;
14798
14799         if (tp->snd_wnd > cwnd_to_use)
14800                 sendwin = cwnd_to_use;
14801         else
14802                 sendwin = tp->snd_wnd;
14803         if (ctf_outstanding(tp) >= tp->snd_wnd) {
14804                 /* We never want to go over our peers rcv-window */
14805                 len = 0;
14806         } else {
14807                 uint32_t flight;
14808
14809                 flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
14810                 if (flight >= sendwin) {
14811                         /*
14812                          * We have in flight what we are allowed by cwnd (if
14813                          * it was rwnd blocking it would have hit above out
14814                          * >= tp->snd_wnd).
14815                          */
14816                         return (0);
14817                 }
14818                 len = sendwin - flight;
14819                 if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
14820                         /* We would send too much (beyond the rwnd) */
14821                         len = tp->snd_wnd - ctf_outstanding(tp);
14822                 }
14823                 if ((len + sb_offset) > avail) {
14824                         /*
14825                          * We don't have that much in the SB, how much is
14826                          * there?
14827                          */
14828                         len = avail - sb_offset;
14829                 }
14830         }
14831         return (len);
14832 }
14833
14834 static void
14835 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
14836              unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
14837              int rsm_is_null, int optlen, int line, uint16_t mode)
14838 {
14839         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14840                 union tcp_log_stackspecific log;
14841                 struct timeval tv;
14842
14843                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14844                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
14845                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
14846                 log.u_bbr.flex1 = error;
14847                 log.u_bbr.flex2 = flags;
14848                 log.u_bbr.flex3 = rsm_is_null;
14849                 log.u_bbr.flex4 = ipoptlen;
14850                 log.u_bbr.flex5 = tp->rcv_numsacks;
14851                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
14852                 log.u_bbr.flex7 = optlen;
14853                 log.u_bbr.flex8 = rack->r_fsb_inited;
14854                 log.u_bbr.applimited = rack->r_fast_output;
14855                 log.u_bbr.bw_inuse = rack_get_bw(rack);
14856                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
14857                 log.u_bbr.cwnd_gain = mode;
14858                 log.u_bbr.pkts_out = orig_len;
14859                 log.u_bbr.lt_epoch = len;
14860                 log.u_bbr.delivered = line;
14861                 log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14862                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14863                 tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
14864                                len, &log, false, NULL, NULL, 0, &tv);
14865         }
14866 }
14867
14868
14869 static struct mbuf *
14870 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
14871                    struct rack_fast_send_blk *fsb,
14872                    int32_t seglimit, int32_t segsize)
14873 {
14874 #ifdef KERN_TLS
14875         struct ktls_session *tls, *ntls;
14876         struct mbuf *start;
14877 #endif
14878         struct mbuf *m, *n, **np, *smb;
14879         struct mbuf *top;
14880         int32_t off, soff;
14881         int32_t len = *plen;
14882         int32_t fragsize;
14883         int32_t len_cp = 0;
14884         uint32_t mlen, frags;
14885
14886         soff = off = the_off;
14887         smb = m = the_m;
14888         np = &top;
14889         top = NULL;
14890 #ifdef KERN_TLS
14891         if (hw_tls && (m->m_flags & M_EXTPG))
14892                 tls = m->m_epg_tls;
14893         else
14894                 tls = NULL;
14895         start = m;
14896 #endif
14897         while (len > 0) {
14898                 if (m == NULL) {
14899                         *plen = len_cp;
14900                         break;
14901                 }
14902 #ifdef KERN_TLS
14903                 if (hw_tls) {
14904                         if (m->m_flags & M_EXTPG)
14905                                 ntls = m->m_epg_tls;
14906                         else
14907                                 ntls = NULL;
14908
14909                         /*
14910                          * Avoid mixing TLS records with handshake
14911                          * data or TLS records from different
14912                          * sessions.
14913                          */
14914                         if (tls != ntls) {
14915                                 MPASS(m != start);
14916                                 *plen = len_cp;
14917                                 break;
14918                         }
14919                 }
14920 #endif
14921                 mlen = min(len, m->m_len - off);
14922                 if (seglimit) {
14923                         /*
14924                          * For M_EXTPG mbufs, add 3 segments
14925                          * + 1 in case we are crossing page boundaries
14926                          * + 2 in case the TLS hdr/trailer are used
14927                          * It is cheaper to just add the segments
14928                          * than it is to take the cache miss to look
14929                          * at the mbuf ext_pgs state in detail.
14930                          */
14931                         if (m->m_flags & M_EXTPG) {
14932                                 fragsize = min(segsize, PAGE_SIZE);
14933                                 frags = 3;
14934                         } else {
14935                                 fragsize = segsize;
14936                                 frags = 0;
14937                         }
14938
14939                         /* Break if we really can't fit anymore. */
14940                         if ((frags + 1) >= seglimit) {
14941                                 *plen = len_cp;
14942                                 break;
14943                         }
14944
14945                         /*
14946                          * Reduce size if you can't copy the whole
14947                          * mbuf. If we can't copy the whole mbuf, also
14948                          * adjust len so the loop will end after this
14949                          * mbuf.
14950                          */
14951                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
14952                                 mlen = (seglimit - frags - 1) * fragsize;
14953                                 len = mlen;
14954                                 *plen = len_cp + len;
14955                         }
14956                         frags += howmany(mlen, fragsize);
14957                         if (frags == 0)
14958                                 frags++;
14959                         seglimit -= frags;
14960                         KASSERT(seglimit > 0,
14961                             ("%s: seglimit went too low", __func__));
14962                 }
14963                 n = m_get(M_NOWAIT, m->m_type);
14964                 *np = n;
14965                 if (n == NULL)
14966                         goto nospace;
14967                 n->m_len = mlen;
14968                 soff += mlen;
14969                 len_cp += n->m_len;
14970                 if (m->m_flags & (M_EXT|M_EXTPG)) {
14971                         n->m_data = m->m_data + off;
14972                         mb_dupcl(n, m);
14973                 } else {
14974                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
14975                             (u_int)n->m_len);
14976                 }
14977                 len -= n->m_len;
14978                 off = 0;
14979                 m = m->m_next;
14980                 np = &n->m_next;
14981                 if (len || (soff == smb->m_len)) {
14982                         /*
14983                          * We have more so we move forward  or
14984                          * we have consumed the entire mbuf and
14985                          * len has fell to 0.
14986                          */
14987                         soff = 0;
14988                         smb = m;
14989                 }
14990
14991         }
14992         if (fsb != NULL) {
14993                 fsb->m = smb;
14994                 fsb->off = soff;
14995                 if (smb) {
14996                         /*
14997                          * Save off the size of the mbuf. We do
14998                          * this so that we can recognize when it
14999                          * has been trimmed by sbcut() as acks
15000                          * come in.
15001                          */
15002                         fsb->o_m_len = smb->m_len;
15003                 } else {
15004                         /*
15005                          * This is the case where the next mbuf went to NULL. This
15006                          * means with this copy we have sent everything in the sb.
15007                          * In theory we could clear the fast_output flag, but lets
15008                          * not since its possible that we could get more added
15009                          * and acks that call the extend function which would let
15010                          * us send more.
15011                          */
15012                         fsb->o_m_len = 0;
15013                 }
15014         }
15015         return (top);
15016 nospace:
15017         if (top)
15018                 m_freem(top);
15019         return (NULL);
15020
15021 }
15022
15023 /*
15024  * This is a copy of m_copym(), taking the TSO segment size/limit
15025  * constraints into account, and advancing the sndptr as it goes.
15026  */
15027 static struct mbuf *
15028 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15029                 int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15030 {
15031         struct mbuf *m, *n;
15032         int32_t soff;
15033
15034         soff = rack->r_ctl.fsb.off;
15035         m = rack->r_ctl.fsb.m;
15036         if (rack->r_ctl.fsb.o_m_len != m->m_len) {
15037                 /*
15038                  * The mbuf had the front of it chopped off by an ack
15039                  * we need to adjust the soff/off by that difference.
15040                  */
15041                 uint32_t delta;
15042
15043                 delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15044                 soff -= delta;
15045         }
15046         KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15047         KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15048         KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15049                                  __FUNCTION__,
15050                                  rack, *plen, m, m->m_len));
15051         /* Save off the right location before we copy and advance */
15052         *s_soff = soff;
15053         *s_mb = rack->r_ctl.fsb.m;
15054         n = rack_fo_base_copym(m, soff, plen,
15055                                &rack->r_ctl.fsb,
15056                                seglimit, segsize);
15057         return (n);
15058 }
15059
15060 static int
15061 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15062                      uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len)
15063 {
15064         /*
15065          * Enter the fast retransmit path. We are given that a sched_pin is
15066          * in place (if accounting is compliled in) and the cycle count taken
15067          * at the entry is in the ts_val. The concept her is that the rsm
15068          * now holds the mbuf offsets and such so we can directly transmit
15069          * without a lot of overhead, the len field is already set for
15070          * us to prohibit us from sending too much (usually its 1MSS).
15071          */
15072         struct ip *ip = NULL;
15073         struct udphdr *udp = NULL;
15074         struct tcphdr *th = NULL;
15075         struct mbuf *m = NULL;
15076         struct inpcb *inp;
15077         uint8_t *cpto;
15078         struct tcp_log_buffer *lgb;
15079 #ifdef TCP_ACCOUNTING
15080         uint64_t crtsc;
15081         int cnt_thru = 1;
15082 #endif
15083         int doing_tlp = 0;
15084         struct tcpopt to;
15085         u_char opt[TCP_MAXOLEN];
15086         uint32_t hdrlen, optlen;
15087         int32_t slot, segsiz, max_val, tso = 0, error, flags, ulen = 0;
15088         uint32_t us_cts;
15089         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15090         uint32_t if_hw_tsomaxsegsize;
15091
15092 #ifdef INET6
15093         struct ip6_hdr *ip6 = NULL;
15094
15095         if (rack->r_is_v6) {
15096                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15097                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15098         } else
15099 #endif                          /* INET6 */
15100         {
15101                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15102                 hdrlen = sizeof(struct tcpiphdr);
15103         }
15104         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15105                 goto failed;
15106         }
15107         if (rsm->r_flags & RACK_TLP)
15108                 doing_tlp = 1;
15109         startseq = rsm->r_start;
15110         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15111         inp = rack->rc_inp;
15112         to.to_flags = 0;
15113         flags = tcp_outflags[tp->t_state];
15114         if (flags & (TH_SYN|TH_RST)) {
15115                 goto failed;
15116         }
15117         if (rsm->r_flags & RACK_HAS_FIN) {
15118                 /* We can't send a FIN here */
15119                 goto failed;
15120         }
15121         if (flags & TH_FIN) {
15122                 /* We never send a FIN */
15123                 flags &= ~TH_FIN;
15124         }
15125         if (tp->t_flags & TF_RCVD_TSTMP) {
15126                 to.to_tsval = ms_cts + tp->ts_offset;
15127                 to.to_tsecr = tp->ts_recent;
15128                 to.to_flags = TOF_TS;
15129         }
15130         optlen = tcp_addoptions(&to, opt);
15131         hdrlen += optlen;
15132         udp = rack->r_ctl.fsb.udp;
15133         if (udp)
15134                 hdrlen += sizeof(struct udphdr);
15135         if (rack->r_ctl.rc_pace_max_segs)
15136                 max_val = rack->r_ctl.rc_pace_max_segs;
15137         else if (rack->rc_user_set_max_segs)
15138                 max_val = rack->rc_user_set_max_segs * segsiz;
15139         else
15140                 max_val = len;
15141         if ((tp->t_flags & TF_TSO) &&
15142             V_tcp_do_tso &&
15143             (len > segsiz) &&
15144             (tp->t_port == 0))
15145                 tso = 1;
15146 #ifdef INET6
15147         if (MHLEN < hdrlen + max_linkhdr)
15148                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15149         else
15150 #endif
15151                 m = m_gethdr(M_NOWAIT, MT_DATA);
15152         if (m == NULL)
15153                 goto failed;
15154         m->m_data += max_linkhdr;
15155         m->m_len = hdrlen;
15156         th = rack->r_ctl.fsb.th;
15157         /* Establish the len to send */
15158         if (len > max_val)
15159                 len = max_val;
15160         if ((tso) && (len + optlen > tp->t_maxseg)) {
15161                 uint32_t if_hw_tsomax;
15162                 int32_t max_len;
15163
15164                 /* extract TSO information */
15165                 if_hw_tsomax = tp->t_tsomax;
15166                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15167                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15168                 /*
15169                  * Check if we should limit by maximum payload
15170                  * length:
15171                  */
15172                 if (if_hw_tsomax != 0) {
15173                         /* compute maximum TSO length */
15174                         max_len = (if_hw_tsomax - hdrlen -
15175                                    max_linkhdr);
15176                         if (max_len <= 0) {
15177                                 goto failed;
15178                         } else if (len > max_len) {
15179                                 len = max_len;
15180                         }
15181                 }
15182                 if (len <= segsiz) {
15183                         /*
15184                          * In case there are too many small fragments don't
15185                          * use TSO:
15186                          */
15187                         tso = 0;
15188                 }
15189         } else {
15190                 tso = 0;
15191         }
15192         if ((tso == 0) && (len > segsiz))
15193                 len = segsiz;
15194         us_cts = tcp_get_usecs(tv);
15195         if ((len == 0) ||
15196             (len <= MHLEN - hdrlen - max_linkhdr)) {
15197                 goto failed;
15198         }
15199         th->th_seq = htonl(rsm->r_start);
15200         th->th_ack = htonl(tp->rcv_nxt);
15201         /*
15202          * The PUSH bit should only be applied
15203          * if the full retransmission is made. If
15204          * we are sending less than this is the
15205          * left hand edge and should not have
15206          * the PUSH bit.
15207          */
15208         if ((rsm->r_flags & RACK_HAD_PUSH) &&
15209             (len == (rsm->r_end - rsm->r_start)))
15210                 flags |= TH_PUSH;
15211         th->th_flags = flags;
15212         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15213         if (th->th_win == 0) {
15214                 tp->t_sndzerowin++;
15215                 tp->t_flags |= TF_RXWIN0SENT;
15216         } else
15217                 tp->t_flags &= ~TF_RXWIN0SENT;
15218         if (rsm->r_flags & RACK_TLP) {
15219                 /*
15220                  * TLP should not count in retran count, but
15221                  * in its own bin
15222                  */
15223                 counter_u64_add(rack_tlp_retran, 1);
15224                 counter_u64_add(rack_tlp_retran_bytes, len);
15225         } else {
15226                 tp->t_sndrexmitpack++;
15227                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15228                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15229         }
15230 #ifdef STATS
15231         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15232                                  len);
15233 #endif
15234         if (rsm->m == NULL)
15235                 goto failed;
15236         if (rsm->orig_m_len != rsm->m->m_len) {
15237                 /* Fix up the orig_m_len and possibly the mbuf offset */
15238                 rack_adjust_orig_mlen(rsm);
15239         }
15240         m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize);
15241         if (len <= segsiz) {
15242                 /*
15243                  * Must have ran out of mbufs for the copy
15244                  * shorten it to no longer need tso. Lets
15245                  * not put on sendalot since we are low on
15246                  * mbufs.
15247                  */
15248                 tso = 0;
15249         }
15250         if ((m->m_next == NULL) || (len <= 0)){
15251                 goto failed;
15252         }
15253         if (udp) {
15254                 if (rack->r_is_v6)
15255                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15256                 else
15257                         ulen = hdrlen + len - sizeof(struct ip);
15258                 udp->uh_ulen = htons(ulen);
15259         }
15260         m->m_pkthdr.rcvif = (struct ifnet *)0;
15261         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15262 #ifdef INET6
15263         if (rack->r_is_v6) {
15264                 if (tp->t_port) {
15265                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15266                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15267                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15268                         th->th_sum = htons(0);
15269                         UDPSTAT_INC(udps_opackets);
15270                 } else {
15271                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15272                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15273                         th->th_sum = in6_cksum_pseudo(ip6,
15274                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15275                                                       0);
15276                 }
15277         }
15278 #endif
15279 #if defined(INET6) && defined(INET)
15280         else
15281 #endif
15282 #ifdef INET
15283         {
15284                 if (tp->t_port) {
15285                         m->m_pkthdr.csum_flags = CSUM_UDP;
15286                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15287                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15288                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15289                         th->th_sum = htons(0);
15290                         UDPSTAT_INC(udps_opackets);
15291                 } else {
15292                         m->m_pkthdr.csum_flags = CSUM_TCP;
15293                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15294                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15295                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15296                                                                         IPPROTO_TCP + len + optlen));
15297                 }
15298                 /* IP version must be set here for ipv4/ipv6 checking later */
15299                 KASSERT(ip->ip_v == IPVERSION,
15300                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15301         }
15302 #endif
15303         if (tso) {
15304                 KASSERT(len > tp->t_maxseg - optlen,
15305                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15306                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15307                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15308         }
15309 #ifdef INET6
15310         if (rack->r_is_v6) {
15311                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15312                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15313                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15314                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15315                 else
15316                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15317         }
15318 #endif
15319 #if defined(INET) && defined(INET6)
15320         else
15321 #endif
15322 #ifdef INET
15323         {
15324                 ip->ip_len = htons(m->m_pkthdr.len);
15325                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15326                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15327                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15328                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15329                                 ip->ip_off |= htons(IP_DF);
15330                         }
15331                 } else {
15332                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15333                 }
15334         }
15335 #endif
15336         /* Time to copy in our header */
15337         cpto = mtod(m, uint8_t *);
15338         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15339         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15340         if (optlen) {
15341                 bcopy(opt, th + 1, optlen);
15342                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15343         } else {
15344                 th->th_off = sizeof(struct tcphdr) >> 2;
15345         }
15346         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15347                 union tcp_log_stackspecific log;
15348
15349                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15350                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15351                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15352                 if (rack->rack_no_prr)
15353                         log.u_bbr.flex1 = 0;
15354                 else
15355                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15356                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15357                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15358                 log.u_bbr.flex4 = max_val;
15359                 log.u_bbr.flex5 = 0;
15360                 /* Save off the early/late values */
15361                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15362                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15363                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15364                 log.u_bbr.flex8 = 1;
15365                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15366                 log.u_bbr.flex7 = 55;
15367                 log.u_bbr.pkts_out = tp->t_maxseg;
15368                 log.u_bbr.timeStamp = cts;
15369                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15370                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15371                 log.u_bbr.delivered = 0;
15372                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15373                                      len, &log, false, NULL, NULL, 0, tv);
15374         } else
15375                 lgb = NULL;
15376 #ifdef INET6
15377         if (rack->r_is_v6) {
15378                 error = ip6_output(m, NULL,
15379                                    &inp->inp_route6,
15380                                    0, NULL, NULL, inp);
15381         }
15382 #endif
15383 #if defined(INET) && defined(INET6)
15384         else
15385 #endif
15386 #ifdef INET
15387         {
15388                 error = ip_output(m, NULL,
15389                                   &inp->inp_route,
15390                                   0, 0, inp);
15391         }
15392 #endif
15393         m = NULL;
15394         if (lgb) {
15395                 lgb->tlb_errno = error;
15396                 lgb = NULL;
15397         }
15398         if (error) {
15399                 goto failed;
15400         }
15401         rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15402                         rsm, RACK_SENT_FP, rsm->m, rsm->soff);
15403         if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15404                 rack->rc_tlp_in_progress = 1;
15405                 rack->r_ctl.rc_tlp_cnt_out++;
15406         }
15407         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15408         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15409         if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15410                 rack->r_ctl.retran_during_recovery += len;
15411         {
15412                 int idx;
15413
15414                 idx = (len / segsiz) + 3;
15415                 if (idx >= TCP_MSS_ACCT_ATIMER)
15416                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15417                 else
15418                         counter_u64_add(rack_out_size[idx], 1);
15419         }
15420         if (tp->t_rtttime == 0) {
15421                 tp->t_rtttime = ticks;
15422                 tp->t_rtseq = startseq;
15423                 KMOD_TCPSTAT_INC(tcps_segstimed);
15424         }
15425         counter_u64_add(rack_fto_rsm_send, 1);
15426         if (error && (error == ENOBUFS)) {
15427                 slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15428                 if (rack->rc_enobuf < 0x7f)
15429                         rack->rc_enobuf++;
15430                 if (slot < (10 * HPTS_USEC_IN_MSEC))
15431                         slot = 10 * HPTS_USEC_IN_MSEC;
15432         } else
15433                 slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
15434         if ((slot == 0) ||
15435             (rack->rc_always_pace == 0) ||
15436             (rack->r_rr_config == 1)) {
15437                 /*
15438                  * We have no pacing set or we
15439                  * are using old-style rack or
15440                  * we are overriden to use the old 1ms pacing.
15441                  */
15442                 slot = rack->r_ctl.rc_min_to;
15443         }
15444         rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
15445         if (rack->r_must_retran) {
15446                 rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
15447                 if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
15448                         /*
15449                          * We have retransmitted all we need.
15450                          */
15451                         rack->r_must_retran = 0;
15452                         rack->r_ctl.rc_out_at_rto = 0;
15453                 }
15454         }
15455 #ifdef TCP_ACCOUNTING
15456         crtsc = get_cyclecount();
15457         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15458                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15459         }
15460         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15461         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15462                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15463         }
15464         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15465         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15466                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
15467         }
15468         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
15469         sched_unpin();
15470 #endif
15471         return (0);
15472 failed:
15473         if (m)
15474                 m_free(m);
15475         return (-1);
15476 }
15477
15478 static void
15479 rack_sndbuf_autoscale(struct tcp_rack *rack)
15480 {
15481         /*
15482          * Automatic sizing of send socket buffer.  Often the send buffer
15483          * size is not optimally adjusted to the actual network conditions
15484          * at hand (delay bandwidth product).  Setting the buffer size too
15485          * small limits throughput on links with high bandwidth and high
15486          * delay (eg. trans-continental/oceanic links).  Setting the
15487          * buffer size too big consumes too much real kernel memory,
15488          * especially with many connections on busy servers.
15489          *
15490          * The criteria to step up the send buffer one notch are:
15491          *  1. receive window of remote host is larger than send buffer
15492          *     (with a fudge factor of 5/4th);
15493          *  2. send buffer is filled to 7/8th with data (so we actually
15494          *     have data to make use of it);
15495          *  3. send buffer fill has not hit maximal automatic size;
15496          *  4. our send window (slow start and cogestion controlled) is
15497          *     larger than sent but unacknowledged data in send buffer.
15498          *
15499          * Note that the rack version moves things much faster since
15500          * we want to avoid hitting cache lines in the rack_fast_output()
15501          * path so this is called much less often and thus moves
15502          * the SB forward by a percentage.
15503          */
15504         struct socket *so;
15505         struct tcpcb *tp;
15506         uint32_t sendwin, scaleup;
15507
15508         tp = rack->rc_tp;
15509         so = rack->rc_inp->inp_socket;
15510         sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
15511         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
15512                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
15513                     sbused(&so->so_snd) >=
15514                     (so->so_snd.sb_hiwat / 8 * 7) &&
15515                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
15516                     sendwin >= (sbused(&so->so_snd) -
15517                     (tp->snd_nxt - tp->snd_una))) {
15518                         if (rack_autosndbuf_inc)
15519                                 scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
15520                         else
15521                                 scaleup = V_tcp_autosndbuf_inc;
15522                         if (scaleup < V_tcp_autosndbuf_inc)
15523                                 scaleup = V_tcp_autosndbuf_inc;
15524                         scaleup += so->so_snd.sb_hiwat;
15525                         if (scaleup > V_tcp_autosndbuf_max)
15526                                 scaleup = V_tcp_autosndbuf_max;
15527                         if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
15528                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
15529                 }
15530         }
15531 }
15532
15533 static int
15534 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
15535                  uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
15536 {
15537         /*
15538          * Enter to do fast output. We are given that the sched_pin is
15539          * in place (if accounting is compiled in) and the cycle count taken
15540          * at entry is in place in ts_val. The idea here is that
15541          * we know how many more bytes needs to be sent (presumably either
15542          * during pacing or to fill the cwnd and that was greater than
15543          * the max-burst). We have how much to send and all the info we
15544          * need to just send.
15545          */
15546         struct ip *ip = NULL;
15547         struct udphdr *udp = NULL;
15548         struct tcphdr *th = NULL;
15549         struct mbuf *m, *s_mb;
15550         struct inpcb *inp;
15551         uint8_t *cpto;
15552         struct tcp_log_buffer *lgb;
15553 #ifdef TCP_ACCOUNTING
15554         uint64_t crtsc;
15555 #endif
15556         struct tcpopt to;
15557         u_char opt[TCP_MAXOLEN];
15558         uint32_t hdrlen, optlen;
15559         int cnt_thru = 1;
15560         int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, flags, ulen = 0;
15561         uint32_t us_cts, s_soff;
15562         uint32_t if_hw_tsomaxsegcount = 0, startseq;
15563         uint32_t if_hw_tsomaxsegsize;
15564         uint16_t add_flag = RACK_SENT_FP;
15565 #ifdef INET6
15566         struct ip6_hdr *ip6 = NULL;
15567
15568         if (rack->r_is_v6) {
15569                 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15570                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15571         } else
15572 #endif                          /* INET6 */
15573         {
15574                 ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15575                 hdrlen = sizeof(struct tcpiphdr);
15576         }
15577         if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15578                 m = NULL;
15579                 goto failed;
15580         }
15581         startseq = tp->snd_max;
15582         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15583         inp = rack->rc_inp;
15584         len = rack->r_ctl.fsb.left_to_send;
15585         to.to_flags = 0;
15586         flags = rack->r_ctl.fsb.tcp_flags;
15587         if (tp->t_flags & TF_RCVD_TSTMP) {
15588                 to.to_tsval = ms_cts + tp->ts_offset;
15589                 to.to_tsecr = tp->ts_recent;
15590                 to.to_flags = TOF_TS;
15591         }
15592         optlen = tcp_addoptions(&to, opt);
15593         hdrlen += optlen;
15594         udp = rack->r_ctl.fsb.udp;
15595         if (udp)
15596                 hdrlen += sizeof(struct udphdr);
15597         if (rack->r_ctl.rc_pace_max_segs)
15598                 max_val = rack->r_ctl.rc_pace_max_segs;
15599         else if (rack->rc_user_set_max_segs)
15600                 max_val = rack->rc_user_set_max_segs * segsiz;
15601         else
15602                 max_val = len;
15603         if ((tp->t_flags & TF_TSO) &&
15604             V_tcp_do_tso &&
15605             (len > segsiz) &&
15606             (tp->t_port == 0))
15607                 tso = 1;
15608 again:
15609 #ifdef INET6
15610         if (MHLEN < hdrlen + max_linkhdr)
15611                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15612         else
15613 #endif
15614                 m = m_gethdr(M_NOWAIT, MT_DATA);
15615         if (m == NULL)
15616                 goto failed;
15617         m->m_data += max_linkhdr;
15618         m->m_len = hdrlen;
15619         th = rack->r_ctl.fsb.th;
15620         /* Establish the len to send */
15621         if (len > max_val)
15622                 len = max_val;
15623         if ((tso) && (len + optlen > tp->t_maxseg)) {
15624                 uint32_t if_hw_tsomax;
15625                 int32_t max_len;
15626
15627                 /* extract TSO information */
15628                 if_hw_tsomax = tp->t_tsomax;
15629                 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15630                 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15631                 /*
15632                  * Check if we should limit by maximum payload
15633                  * length:
15634                  */
15635                 if (if_hw_tsomax != 0) {
15636                         /* compute maximum TSO length */
15637                         max_len = (if_hw_tsomax - hdrlen -
15638                                    max_linkhdr);
15639                         if (max_len <= 0) {
15640                                 goto failed;
15641                         } else if (len > max_len) {
15642                                 len = max_len;
15643                         }
15644                 }
15645                 if (len <= segsiz) {
15646                         /*
15647                          * In case there are too many small fragments don't
15648                          * use TSO:
15649                          */
15650                         tso = 0;
15651                 }
15652         } else {
15653                 tso = 0;
15654         }
15655         if ((tso == 0) && (len > segsiz))
15656                 len = segsiz;
15657         us_cts = tcp_get_usecs(tv);
15658         if ((len == 0) ||
15659             (len <= MHLEN - hdrlen - max_linkhdr)) {
15660                 goto failed;
15661         }
15662         sb_offset = tp->snd_max - tp->snd_una;
15663         th->th_seq = htonl(tp->snd_max);
15664         th->th_ack = htonl(tp->rcv_nxt);
15665         th->th_flags = flags;
15666         th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15667         if (th->th_win == 0) {
15668                 tp->t_sndzerowin++;
15669                 tp->t_flags |= TF_RXWIN0SENT;
15670         } else
15671                 tp->t_flags &= ~TF_RXWIN0SENT;
15672         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
15673         KMOD_TCPSTAT_INC(tcps_sndpack);
15674         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
15675 #ifdef STATS
15676         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
15677                                  len);
15678 #endif
15679         if (rack->r_ctl.fsb.m == NULL)
15680                 goto failed;
15681
15682         /* s_mb and s_soff are saved for rack_log_output */
15683         m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, &s_mb, &s_soff);
15684         if (len <= segsiz) {
15685                 /*
15686                  * Must have ran out of mbufs for the copy
15687                  * shorten it to no longer need tso. Lets
15688                  * not put on sendalot since we are low on
15689                  * mbufs.
15690                  */
15691                 tso = 0;
15692         }
15693         if (rack->r_ctl.fsb.rfo_apply_push &&
15694             (len == rack->r_ctl.fsb.left_to_send)) {
15695                 th->th_flags |= TH_PUSH;
15696                 add_flag |= RACK_HAD_PUSH;
15697         }
15698         if ((m->m_next == NULL) || (len <= 0)){
15699                 goto failed;
15700         }
15701         if (udp) {
15702                 if (rack->r_is_v6)
15703                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
15704                 else
15705                         ulen = hdrlen + len - sizeof(struct ip);
15706                 udp->uh_ulen = htons(ulen);
15707         }
15708         m->m_pkthdr.rcvif = (struct ifnet *)0;
15709         if (tp->t_state == TCPS_ESTABLISHED &&
15710             (tp->t_flags2 & TF2_ECN_PERMIT)) {
15711                 /*
15712                  * If the peer has ECN, mark data packets with ECN capable
15713                  * transmission (ECT). Ignore pure ack packets,
15714                  * retransmissions.
15715                  */
15716                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) {
15717 #ifdef INET6
15718                         if (rack->r_is_v6)
15719                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
15720                         else
15721 #endif
15722                                 ip->ip_tos |= IPTOS_ECN_ECT0;
15723                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
15724                         /*
15725                          * Reply with proper ECN notifications.
15726                          * Only set CWR on new data segments.
15727                          */
15728                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
15729                                 flags |= TH_CWR;
15730                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
15731                         }
15732                 }
15733                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
15734                         flags |= TH_ECE;
15735         }
15736         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
15737 #ifdef INET6
15738         if (rack->r_is_v6) {
15739                 if (tp->t_port) {
15740                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15741                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15742                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15743                         th->th_sum = htons(0);
15744                         UDPSTAT_INC(udps_opackets);
15745                 } else {
15746                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15747                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15748                         th->th_sum = in6_cksum_pseudo(ip6,
15749                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15750                                                       0);
15751                 }
15752         }
15753 #endif
15754 #if defined(INET6) && defined(INET)
15755         else
15756 #endif
15757 #ifdef INET
15758         {
15759                 if (tp->t_port) {
15760                         m->m_pkthdr.csum_flags = CSUM_UDP;
15761                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15762                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15763                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15764                         th->th_sum = htons(0);
15765                         UDPSTAT_INC(udps_opackets);
15766                 } else {
15767                         m->m_pkthdr.csum_flags = CSUM_TCP;
15768                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15769                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
15770                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15771                                                                         IPPROTO_TCP + len + optlen));
15772                 }
15773                 /* IP version must be set here for ipv4/ipv6 checking later */
15774                 KASSERT(ip->ip_v == IPVERSION,
15775                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
15776         }
15777 #endif
15778         if (tso) {
15779                 KASSERT(len > tp->t_maxseg - optlen,
15780                         ("%s: len <= tso_segsz tp:%p", __func__, tp));
15781                 m->m_pkthdr.csum_flags |= CSUM_TSO;
15782                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15783         }
15784 #ifdef INET6
15785         if (rack->r_is_v6) {
15786                 ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15787                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15788                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15789                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15790                 else
15791                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15792         }
15793 #endif
15794 #if defined(INET) && defined(INET6)
15795         else
15796 #endif
15797 #ifdef INET
15798         {
15799                 ip->ip_len = htons(m->m_pkthdr.len);
15800                 ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15801                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15802                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15803                         if (tp->t_port == 0 || len < V_tcp_minmss) {
15804                                 ip->ip_off |= htons(IP_DF);
15805                         }
15806                 } else {
15807                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15808                 }
15809         }
15810 #endif
15811         /* Time to copy in our header */
15812         cpto = mtod(m, uint8_t *);
15813         memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15814         th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15815         if (optlen) {
15816                 bcopy(opt, th + 1, optlen);
15817                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15818         } else {
15819                 th->th_off = sizeof(struct tcphdr) >> 2;
15820         }
15821         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15822                 union tcp_log_stackspecific log;
15823
15824                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15825                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15826                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15827                 if (rack->rack_no_prr)
15828                         log.u_bbr.flex1 = 0;
15829                 else
15830                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15831                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15832                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15833                 log.u_bbr.flex4 = max_val;
15834                 log.u_bbr.flex5 = 0;
15835                 /* Save off the early/late values */
15836                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15837                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15838                 log.u_bbr.bw_inuse = rack_get_bw(rack);
15839                 log.u_bbr.flex8 = 0;
15840                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15841                 log.u_bbr.flex7 = 44;
15842                 log.u_bbr.pkts_out = tp->t_maxseg;
15843                 log.u_bbr.timeStamp = cts;
15844                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15845                 log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15846                 log.u_bbr.delivered = 0;
15847                 lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15848                                      len, &log, false, NULL, NULL, 0, tv);
15849         } else
15850                 lgb = NULL;
15851 #ifdef INET6
15852         if (rack->r_is_v6) {
15853                 error = ip6_output(m, NULL,
15854                                    &inp->inp_route6,
15855                                    0, NULL, NULL, inp);
15856         }
15857 #endif
15858 #if defined(INET) && defined(INET6)
15859         else
15860 #endif
15861 #ifdef INET
15862         {
15863                 error = ip_output(m, NULL,
15864                                   &inp->inp_route,
15865                                   0, 0, inp);
15866         }
15867 #endif
15868         if (lgb) {
15869                 lgb->tlb_errno = error;
15870                 lgb = NULL;
15871         }
15872         if (error) {
15873                 *send_err = error;
15874                 m = NULL;
15875                 goto failed;
15876         }
15877         rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
15878                         NULL, add_flag, s_mb, s_soff);
15879         m = NULL;
15880         if (tp->snd_una == tp->snd_max) {
15881                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
15882                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
15883                 tp->t_acktime = ticks;
15884         }
15885         rack->forced_ack = 0;   /* If we send something zap the FA flag */
15886         tot_len += len;
15887         if ((tp->t_flags & TF_GPUTINPROG) == 0)
15888                 rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
15889         tp->snd_max += len;
15890         tp->snd_nxt = tp->snd_max;
15891         {
15892                 int idx;
15893
15894                 idx = (len / segsiz) + 3;
15895                 if (idx >= TCP_MSS_ACCT_ATIMER)
15896                         counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15897                 else
15898                         counter_u64_add(rack_out_size[idx], 1);
15899         }
15900         if (len <= rack->r_ctl.fsb.left_to_send)
15901                 rack->r_ctl.fsb.left_to_send -= len;
15902         else
15903                 rack->r_ctl.fsb.left_to_send = 0;
15904         if (rack->r_ctl.fsb.left_to_send < segsiz) {
15905                 rack->r_fast_output = 0;
15906                 rack->r_ctl.fsb.left_to_send = 0;
15907                 /* At the end of fast_output scale up the sb */
15908                 SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
15909                 rack_sndbuf_autoscale(rack);
15910                 SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
15911         }
15912         if (tp->t_rtttime == 0) {
15913                 tp->t_rtttime = ticks;
15914                 tp->t_rtseq = startseq;
15915                 KMOD_TCPSTAT_INC(tcps_segstimed);
15916         }
15917         if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
15918             (max_val > len) &&
15919             (tso == 0)) {
15920                 max_val -= len;
15921                 len = segsiz;
15922                 th = rack->r_ctl.fsb.th;
15923                 cnt_thru++;
15924                 goto again;
15925         }
15926         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15927         counter_u64_add(rack_fto_send, 1);
15928         slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
15929         rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
15930 #ifdef TCP_ACCOUNTING
15931         crtsc = get_cyclecount();
15932         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15933                 tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
15934         }
15935         counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
15936         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15937                 tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
15938         }
15939         counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
15940         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
15941                 tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
15942         }
15943         counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
15944         sched_unpin();
15945 #endif
15946         return (0);
15947 failed:
15948         if (m)
15949                 m_free(m);
15950         rack->r_fast_output = 0;
15951         return (-1);
15952 }
15953
15954 static int
15955 rack_output(struct tcpcb *tp)
15956 {
15957         struct socket *so;
15958         uint32_t recwin;
15959         uint32_t sb_offset, s_moff = 0;
15960         int32_t len, flags, error = 0;
15961         struct mbuf *m, *s_mb = NULL;
15962         struct mbuf *mb;
15963         uint32_t if_hw_tsomaxsegcount = 0;
15964         uint32_t if_hw_tsomaxsegsize;
15965         int32_t segsiz, minseg;
15966         long tot_len_this_send = 0;
15967 #ifdef INET
15968         struct ip *ip = NULL;
15969 #endif
15970 #ifdef TCPDEBUG
15971         struct ipovly *ipov = NULL;
15972 #endif
15973         struct udphdr *udp = NULL;
15974         struct tcp_rack *rack;
15975         struct tcphdr *th;
15976         uint8_t pass = 0;
15977         uint8_t mark = 0;
15978         uint8_t wanted_cookie = 0;
15979         u_char opt[TCP_MAXOLEN];
15980         unsigned ipoptlen, optlen, hdrlen, ulen=0;
15981         uint32_t rack_seq;
15982
15983 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
15984         unsigned ipsec_optlen = 0;
15985
15986 #endif
15987         int32_t idle, sendalot;
15988         int32_t sub_from_prr = 0;
15989         volatile int32_t sack_rxmit;
15990         struct rack_sendmap *rsm = NULL;
15991         int32_t tso, mtu;
15992         struct tcpopt to;
15993         int32_t slot = 0;
15994         int32_t sup_rack = 0;
15995         uint32_t cts, ms_cts, delayed, early;
15996         uint16_t add_flag = RACK_SENT_SP;
15997         uint8_t hpts_calling,  doing_tlp = 0;
15998         uint32_t cwnd_to_use, pace_max_seg;
15999         int32_t do_a_prefetch = 0;
16000         int32_t prefetch_rsm = 0;
16001         int32_t orig_len = 0;
16002         struct timeval tv;
16003         int32_t prefetch_so_done = 0;
16004         struct tcp_log_buffer *lgb;
16005         struct inpcb *inp;
16006         struct sockbuf *sb;
16007         uint64_t ts_val = 0;
16008 #ifdef TCP_ACCOUNTING
16009         uint64_t crtsc;
16010 #endif
16011 #ifdef INET6
16012         struct ip6_hdr *ip6 = NULL;
16013         int32_t isipv6;
16014 #endif
16015         uint8_t filled_all = 0;
16016         bool hw_tls = false;
16017
16018         /* setup and take the cache hits here */
16019         rack = (struct tcp_rack *)tp->t_fb_ptr;
16020 #ifdef TCP_ACCOUNTING
16021         sched_pin();
16022         ts_val = get_cyclecount();
16023 #endif
16024         hpts_calling = rack->rc_inp->inp_hpts_calls;
16025         NET_EPOCH_ASSERT();
16026         INP_WLOCK_ASSERT(rack->rc_inp);
16027 #ifdef TCP_OFFLOAD
16028         if (tp->t_flags & TF_TOE) {
16029 #ifdef TCP_ACCOUNTING
16030                 sched_unpin();
16031 #endif
16032                 return (tcp_offload_output(tp));
16033         }
16034 #endif
16035         /*
16036          * For TFO connections in SYN_RECEIVED, only allow the initial
16037          * SYN|ACK and those sent by the retransmit timer.
16038          */
16039         if (IS_FASTOPEN(tp->t_flags) &&
16040             (tp->t_state == TCPS_SYN_RECEIVED) &&
16041             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16042             (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16043 #ifdef TCP_ACCOUNTING
16044                 sched_unpin();
16045 #endif
16046                 return (0);
16047         }
16048 #ifdef INET6
16049         if (rack->r_state) {
16050                 /* Use the cache line loaded if possible */
16051                 isipv6 = rack->r_is_v6;
16052         } else {
16053                 isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16054         }
16055 #endif
16056         early = 0;
16057         cts = tcp_get_usecs(&tv);
16058         ms_cts = tcp_tv_to_mssectick(&tv);
16059         if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16060             rack->rc_inp->inp_in_hpts) {
16061                 /*
16062                  * We are on the hpts for some timer but not hptsi output.
16063                  * Remove from the hpts unconditionally.
16064                  */
16065                 rack_timer_cancel(tp, rack, cts, __LINE__);
16066         }
16067         /* Are we pacing and late? */
16068         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16069             TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16070                 /* We are delayed */
16071                 delayed = cts - rack->r_ctl.rc_last_output_to;
16072         } else {
16073                 delayed = 0;
16074         }
16075         /* Do the timers, which may override the pacer */
16076         if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16077                 if (rack_process_timers(tp, rack, cts, hpts_calling)) {
16078                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16079 #ifdef TCP_ACCOUNTING
16080                         sched_unpin();
16081 #endif
16082                         return (0);
16083                 }
16084         }
16085         if (rack->rc_in_persist) {
16086                 if (rack->rc_inp->inp_in_hpts == 0) {
16087                         /* Timer is not running */
16088                         rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16089                 }
16090 #ifdef TCP_ACCOUNTING
16091                 sched_unpin();
16092 #endif
16093                 return (0);
16094         }
16095         if ((rack->r_timer_override) ||
16096             (rack->rc_ack_can_sendout_data) ||
16097             (delayed) ||
16098             (tp->t_state < TCPS_ESTABLISHED)) {
16099                 rack->rc_ack_can_sendout_data = 0;
16100                 if (rack->rc_inp->inp_in_hpts)
16101                         tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
16102         } else if (rack->rc_inp->inp_in_hpts) {
16103                 /*
16104                  * On the hpts you can't pass even if ACKNOW is on, we will
16105                  * when the hpts fires.
16106                  */
16107 #ifdef TCP_ACCOUNTING
16108                 crtsc = get_cyclecount();
16109                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16110                         tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16111                 }
16112                 counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16113                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16114                         tp->tcp_cnt_counters[SND_BLOCKED]++;
16115                 }
16116                 counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16117                 sched_unpin();
16118 #endif
16119                 counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16120                 return (0);
16121         }
16122         rack->rc_inp->inp_hpts_calls = 0;
16123         /* Finish out both pacing early and late accounting */
16124         if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16125             TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16126                 early = rack->r_ctl.rc_last_output_to - cts;
16127         } else
16128                 early = 0;
16129         if (delayed) {
16130                 rack->r_ctl.rc_agg_delayed += delayed;
16131                 rack->r_late = 1;
16132         } else if (early) {
16133                 rack->r_ctl.rc_agg_early += early;
16134                 rack->r_early = 1;
16135         }
16136         /* Now that early/late accounting is done turn off the flag */
16137         rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16138         rack->r_wanted_output = 0;
16139         rack->r_timer_override = 0;
16140         if ((tp->t_state != rack->r_state) &&
16141             TCPS_HAVEESTABLISHED(tp->t_state)) {
16142                 rack_set_state(tp, rack);
16143         }
16144         if ((rack->r_fast_output) &&
16145             (tp->rcv_numsacks == 0)) {
16146                 int ret;
16147
16148                 error = 0;
16149                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16150                 if (ret >= 0)
16151                         return(ret);
16152                 else if (error) {
16153                         inp = rack->rc_inp;
16154                         so = inp->inp_socket;
16155                         sb = &so->so_snd;
16156                         goto nomore;
16157                 }
16158         }
16159         inp = rack->rc_inp;
16160         /*
16161          * For TFO connections in SYN_SENT or SYN_RECEIVED,
16162          * only allow the initial SYN or SYN|ACK and those sent
16163          * by the retransmit timer.
16164          */
16165         if (IS_FASTOPEN(tp->t_flags) &&
16166             ((tp->t_state == TCPS_SYN_RECEIVED) ||
16167              (tp->t_state == TCPS_SYN_SENT)) &&
16168             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16169             (tp->t_rxtshift == 0)) {              /* not a retransmit */
16170                 cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16171                 so = inp->inp_socket;
16172                 sb = &so->so_snd;
16173                 goto just_return_nolock;
16174         }
16175         /*
16176          * Determine length of data that should be transmitted, and flags
16177          * that will be used. If there is some data or critical controls
16178          * (SYN, RST) to send, then transmit; otherwise, investigate
16179          * further.
16180          */
16181         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16182         if (tp->t_idle_reduce) {
16183                 if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16184                         rack_cc_after_idle(rack, tp);
16185         }
16186         tp->t_flags &= ~TF_LASTIDLE;
16187         if (idle) {
16188                 if (tp->t_flags & TF_MORETOCOME) {
16189                         tp->t_flags |= TF_LASTIDLE;
16190                         idle = 0;
16191                 }
16192         }
16193         if ((tp->snd_una == tp->snd_max) &&
16194             rack->r_ctl.rc_went_idle_time &&
16195             TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16196                 idle = cts - rack->r_ctl.rc_went_idle_time;
16197                 if (idle > rack_min_probertt_hold) {
16198                         /* Count as a probe rtt */
16199                         if (rack->in_probe_rtt == 0) {
16200                                 rack->r_ctl.rc_lower_rtt_us_cts = cts;
16201                                 rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16202                                 rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16203                                 rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16204                         } else {
16205                                 rack_exit_probertt(rack, cts);
16206                         }
16207                 }
16208                 idle = 0;
16209         }
16210         if (rack_use_fsb && (rack->r_fsb_inited == 0))
16211                 rack_init_fsb_block(tp, rack);
16212 again:
16213         /*
16214          * If we've recently taken a timeout, snd_max will be greater than
16215          * snd_nxt.  There may be SACK information that allows us to avoid
16216          * resending already delivered data.  Adjust snd_nxt accordingly.
16217          */
16218         sendalot = 0;
16219         cts = tcp_get_usecs(&tv);
16220         ms_cts = tcp_tv_to_mssectick(&tv);
16221         tso = 0;
16222         mtu = 0;
16223         segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16224         minseg = segsiz;
16225         if (rack->r_ctl.rc_pace_max_segs == 0)
16226                 pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16227         else
16228                 pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16229         sb_offset = tp->snd_max - tp->snd_una;
16230         cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16231         flags = tcp_outflags[tp->t_state];
16232         while (rack->rc_free_cnt < rack_free_cache) {
16233                 rsm = rack_alloc(rack);
16234                 if (rsm == NULL) {
16235                         if (inp->inp_hpts_calls)
16236                                 /* Retry in a ms */
16237                                 slot = (1 * HPTS_USEC_IN_MSEC);
16238                         so = inp->inp_socket;
16239                         sb = &so->so_snd;
16240                         goto just_return_nolock;
16241                 }
16242                 TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16243                 rack->rc_free_cnt++;
16244                 rsm = NULL;
16245         }
16246         if (inp->inp_hpts_calls)
16247                 inp->inp_hpts_calls = 0;
16248         sack_rxmit = 0;
16249         len = 0;
16250         rsm = NULL;
16251         if (flags & TH_RST) {
16252                 SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16253                 so = inp->inp_socket;
16254                 sb = &so->so_snd;
16255                 goto send;
16256         }
16257         if (rack->r_ctl.rc_resend) {
16258                 /* Retransmit timer */
16259                 rsm = rack->r_ctl.rc_resend;
16260                 rack->r_ctl.rc_resend = NULL;
16261                 rsm->r_flags &= ~RACK_TLP;
16262                 len = rsm->r_end - rsm->r_start;
16263                 sack_rxmit = 1;
16264                 sendalot = 0;
16265                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16266                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16267                          __func__, __LINE__,
16268                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16269                 sb_offset = rsm->r_start - tp->snd_una;
16270                 if (len >= segsiz)
16271                         len = segsiz;
16272         } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16273                 /* We have a retransmit that takes precedence */
16274                 rsm->r_flags &= ~RACK_TLP;
16275                 if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16276                     ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16277                         /* Enter recovery if not induced by a time-out */
16278                         rack->r_ctl.rc_rsm_start = rsm->r_start;
16279                         rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
16280                         rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
16281                         rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
16282                 }
16283 #ifdef INVARIANTS
16284                 if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16285                         panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16286                               tp, rack, rsm, rsm->r_start, tp->snd_una);
16287                 }
16288 #endif
16289                 len = rsm->r_end - rsm->r_start;
16290                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16291                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16292                          __func__, __LINE__,
16293                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16294                 sb_offset = rsm->r_start - tp->snd_una;
16295                 sendalot = 0;
16296                 if (len >= segsiz)
16297                         len = segsiz;
16298                 if (len > 0) {
16299                         sack_rxmit = 1;
16300                         KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16301                         KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16302                             min(len, segsiz));
16303                         counter_u64_add(rack_rtm_prr_retran, 1);
16304                 }
16305         } else if (rack->r_ctl.rc_tlpsend) {
16306                 /* Tail loss probe */
16307                 long cwin;
16308                 long tlen;
16309
16310                 doing_tlp = 1;
16311                 /*
16312                  * Check if we can do a TLP with a RACK'd packet
16313                  * this can happen if we are not doing the rack
16314                  * cheat and we skipped to a TLP and it
16315                  * went off.
16316                  */
16317                 rsm = rack->r_ctl.rc_tlpsend;
16318                 rsm->r_flags |= RACK_TLP;
16319
16320                 rack->r_ctl.rc_tlpsend = NULL;
16321                 sack_rxmit = 1;
16322                 tlen = rsm->r_end - rsm->r_start;
16323                 if (tlen > segsiz)
16324                         tlen = segsiz;
16325                 tp->t_sndtlppack++;
16326                 tp->t_sndtlpbyte += tlen;
16327                 KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16328                         ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16329                          __func__, __LINE__,
16330                          rsm->r_start, tp->snd_una, tp, rack, rsm));
16331                 sb_offset = rsm->r_start - tp->snd_una;
16332                 cwin = min(tp->snd_wnd, tlen);
16333                 len = cwin;
16334         }
16335         if (rack->r_must_retran &&
16336             (rsm == NULL)) {
16337                 /*
16338                  * Non-Sack and we had a RTO or MTU change, we
16339                  * need to retransmit until we reach
16340                  * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
16341                  */
16342                 if (SEQ_GT(tp->snd_max, tp->snd_una)) {
16343                         int sendwin, flight;
16344
16345                         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
16346                         flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
16347                         if (flight >= sendwin) {
16348                                 so = inp->inp_socket;
16349                                 sb = &so->so_snd;
16350                                 goto just_return_nolock;
16351                         }
16352                         rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
16353                         KASSERT(rsm != NULL, ("rsm is NULL rack:%p r_must_retran set", rack));
16354                         if (rsm == NULL) {
16355                                 /* TSNH */
16356                                 rack->r_must_retran = 0;
16357                                 rack->r_ctl.rc_out_at_rto = 0;
16358                                 rack->r_must_retran = 0;
16359                                 so = inp->inp_socket;
16360                                 sb = &so->so_snd;
16361                                 goto just_return_nolock;
16362                         }
16363                         sack_rxmit = 1;
16364                         len = rsm->r_end - rsm->r_start;
16365                         sendalot = 0;
16366                         sb_offset = rsm->r_start - tp->snd_una;
16367                         if (len >= segsiz)
16368                                 len = segsiz;
16369                 } else {
16370                         /* We must be done if there is nothing outstanding */
16371                         rack->r_must_retran = 0;
16372                         rack->r_ctl.rc_out_at_rto = 0;
16373                 }
16374         }
16375         /*
16376          * Enforce a connection sendmap count limit if set
16377          * as long as we are not retransmiting.
16378          */
16379         if ((rsm == NULL) &&
16380             (rack->do_detection == 0) &&
16381             (V_tcp_map_entries_limit > 0) &&
16382             (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
16383                 counter_u64_add(rack_to_alloc_limited, 1);
16384                 if (!rack->alloc_limit_reported) {
16385                         rack->alloc_limit_reported = 1;
16386                         counter_u64_add(rack_alloc_limited_conns, 1);
16387                 }
16388                 so = inp->inp_socket;
16389                 sb = &so->so_snd;
16390                 goto just_return_nolock;
16391         }
16392         if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
16393                 /* we are retransmitting the fin */
16394                 len--;
16395                 if (len) {
16396                         /*
16397                          * When retransmitting data do *not* include the
16398                          * FIN. This could happen from a TLP probe.
16399                          */
16400                         flags &= ~TH_FIN;
16401                 }
16402         }
16403 #ifdef INVARIANTS
16404         /* For debugging */
16405         rack->r_ctl.rc_rsm_at_retran = rsm;
16406 #endif
16407         if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
16408             ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
16409                 int ret;
16410
16411                 ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len);
16412                 if (ret == 0)
16413                         return (0);
16414         }
16415         so = inp->inp_socket;
16416         sb = &so->so_snd;
16417         if (do_a_prefetch == 0) {
16418                 kern_prefetch(sb, &do_a_prefetch);
16419                 do_a_prefetch = 1;
16420         }
16421 #ifdef NETFLIX_SHARED_CWND
16422         if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
16423             rack->rack_enable_scwnd) {
16424                 /* We are doing cwnd sharing */
16425                 if (rack->gp_ready &&
16426                     (rack->rack_attempted_scwnd == 0) &&
16427                     (rack->r_ctl.rc_scw == NULL) &&
16428                     tp->t_lib) {
16429                         /* The pcbid is in, lets make an attempt */
16430                         counter_u64_add(rack_try_scwnd, 1);
16431                         rack->rack_attempted_scwnd = 1;
16432                         rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
16433                                                                    &rack->r_ctl.rc_scw_index,
16434                                                                    segsiz);
16435                 }
16436                 if (rack->r_ctl.rc_scw &&
16437                     (rack->rack_scwnd_is_idle == 1) &&
16438                     sbavail(&so->so_snd)) {
16439                         /* we are no longer out of data */
16440                         tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
16441                         rack->rack_scwnd_is_idle = 0;
16442                 }
16443                 if (rack->r_ctl.rc_scw) {
16444                         /* First lets update and get the cwnd */
16445                         rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
16446                                                                     rack->r_ctl.rc_scw_index,
16447                                                                     tp->snd_cwnd, tp->snd_wnd, segsiz);
16448                 }
16449         }
16450 #endif
16451         /*
16452          * Get standard flags, and add SYN or FIN if requested by 'hidden'
16453          * state flags.
16454          */
16455         if (tp->t_flags & TF_NEEDFIN)
16456                 flags |= TH_FIN;
16457         if (tp->t_flags & TF_NEEDSYN)
16458                 flags |= TH_SYN;
16459         if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
16460                 void *end_rsm;
16461                 end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
16462                 if (end_rsm)
16463                         kern_prefetch(end_rsm, &prefetch_rsm);
16464                 prefetch_rsm = 1;
16465         }
16466         SOCKBUF_LOCK(sb);
16467         /*
16468          * If snd_nxt == snd_max and we have transmitted a FIN, the
16469          * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
16470          * negative length.  This can also occur when TCP opens up its
16471          * congestion window while receiving additional duplicate acks after
16472          * fast-retransmit because TCP will reset snd_nxt to snd_max after
16473          * the fast-retransmit.
16474          *
16475          * In the normal retransmit-FIN-only case, however, snd_nxt will be
16476          * set to snd_una, the sb_offset will be 0, and the length may wind
16477          * up 0.
16478          *
16479          * If sack_rxmit is true we are retransmitting from the scoreboard
16480          * in which case len is already set.
16481          */
16482         if ((sack_rxmit == 0) &&
16483             (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
16484                 uint32_t avail;
16485
16486                 avail = sbavail(sb);
16487                 if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
16488                         sb_offset = tp->snd_nxt - tp->snd_una;
16489                 else
16490                         sb_offset = 0;
16491                 if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
16492                         if (rack->r_ctl.rc_tlp_new_data) {
16493                                 /* TLP is forcing out new data */
16494                                 if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
16495                                         rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
16496                                 }
16497                                 if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
16498                                         if (tp->snd_wnd > sb_offset)
16499                                                 len = tp->snd_wnd - sb_offset;
16500                                         else
16501                                                 len = 0;
16502                                 } else {
16503                                         len = rack->r_ctl.rc_tlp_new_data;
16504                                 }
16505                                 rack->r_ctl.rc_tlp_new_data = 0;
16506                                 doing_tlp = 1;
16507                         }  else {
16508                                 len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
16509                         }
16510                         if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
16511                                 /*
16512                                  * For prr=off, we need to send only 1 MSS
16513                                  * at a time. We do this because another sack could
16514                                  * be arriving that causes us to send retransmits and
16515                                  * we don't want to be on a long pace due to a larger send
16516                                  * that keeps us from sending out the retransmit.
16517                                  */
16518                                 len = segsiz;
16519                         }
16520                 } else {
16521                         uint32_t outstanding;
16522                         /*
16523                          * We are inside of a Fast recovery episode, this
16524                          * is caused by a SACK or 3 dup acks. At this point
16525                          * we have sent all the retransmissions and we rely
16526                          * on PRR to dictate what we will send in the form of
16527                          * new data.
16528                          */
16529
16530                         outstanding = tp->snd_max - tp->snd_una;
16531                         if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
16532                                 if (tp->snd_wnd > outstanding) {
16533                                         len = tp->snd_wnd - outstanding;
16534                                         /* Check to see if we have the data */
16535                                         if ((sb_offset + len) > avail) {
16536                                                 /* It does not all fit */
16537                                                 if (avail > sb_offset)
16538                                                         len = avail - sb_offset;
16539                                                 else
16540                                                         len = 0;
16541                                         }
16542                                 } else {
16543                                         len = 0;
16544                                 }
16545                         } else if (avail > sb_offset) {
16546                                 len = avail - sb_offset;
16547                         } else {
16548                                 len = 0;
16549                         }
16550                         if (len > 0) {
16551                                 if (len > rack->r_ctl.rc_prr_sndcnt) {
16552                                         len = rack->r_ctl.rc_prr_sndcnt;
16553                                 }
16554                                 if (len > 0) {
16555                                         sub_from_prr = 1;
16556                                         counter_u64_add(rack_rtm_prr_newdata, 1);
16557                                 }
16558                         }
16559                         if (len > segsiz) {
16560                                 /*
16561                                  * We should never send more than a MSS when
16562                                  * retransmitting or sending new data in prr
16563                                  * mode unless the override flag is on. Most
16564                                  * likely the PRR algorithm is not going to
16565                                  * let us send a lot as well :-)
16566                                  */
16567                                 if (rack->r_ctl.rc_prr_sendalot == 0) {
16568                                         len = segsiz;
16569                                 }
16570                         } else if (len < segsiz) {
16571                                 /*
16572                                  * Do we send any? The idea here is if the
16573                                  * send empty's the socket buffer we want to
16574                                  * do it. However if not then lets just wait
16575                                  * for our prr_sndcnt to get bigger.
16576                                  */
16577                                 long leftinsb;
16578
16579                                 leftinsb = sbavail(sb) - sb_offset;
16580                                 if (leftinsb > len) {
16581                                         /* This send does not empty the sb */
16582                                         len = 0;
16583                                 }
16584                         }
16585                 }
16586         } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
16587                 /*
16588                  * If you have not established
16589                  * and are not doing FAST OPEN
16590                  * no data please.
16591                  */
16592                 if ((sack_rxmit == 0) &&
16593                     (!IS_FASTOPEN(tp->t_flags))){
16594                         len = 0;
16595                         sb_offset = 0;
16596                 }
16597         }
16598         if (prefetch_so_done == 0) {
16599                 kern_prefetch(so, &prefetch_so_done);
16600                 prefetch_so_done = 1;
16601         }
16602         /*
16603          * Lop off SYN bit if it has already been sent.  However, if this is
16604          * SYN-SENT state and if segment contains data and if we don't know
16605          * that foreign host supports TAO, suppress sending segment.
16606          */
16607         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
16608             ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
16609                 /*
16610                  * When sending additional segments following a TFO SYN|ACK,
16611                  * do not include the SYN bit.
16612                  */
16613                 if (IS_FASTOPEN(tp->t_flags) &&
16614                     (tp->t_state == TCPS_SYN_RECEIVED))
16615                         flags &= ~TH_SYN;
16616         }
16617         /*
16618          * Be careful not to send data and/or FIN on SYN segments. This
16619          * measure is needed to prevent interoperability problems with not
16620          * fully conformant TCP implementations.
16621          */
16622         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
16623                 len = 0;
16624                 flags &= ~TH_FIN;
16625         }
16626         /*
16627          * On TFO sockets, ensure no data is sent in the following cases:
16628          *
16629          *  - When retransmitting SYN|ACK on a passively-created socket
16630          *
16631          *  - When retransmitting SYN on an actively created socket
16632          *
16633          *  - When sending a zero-length cookie (cookie request) on an
16634          *    actively created socket
16635          *
16636          *  - When the socket is in the CLOSED state (RST is being sent)
16637          */
16638         if (IS_FASTOPEN(tp->t_flags) &&
16639             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
16640              ((tp->t_state == TCPS_SYN_SENT) &&
16641               (tp->t_tfo_client_cookie_len == 0)) ||
16642              (flags & TH_RST))) {
16643                 sack_rxmit = 0;
16644                 len = 0;
16645         }
16646         /* Without fast-open there should never be data sent on a SYN */
16647         if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
16648                 tp->snd_nxt = tp->iss;
16649                 len = 0;
16650         }
16651         if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
16652                 /* We only send 1 MSS if we have a DSACK block */
16653                 add_flag |= RACK_SENT_W_DSACK;
16654                 len = segsiz;
16655         }
16656         orig_len = len;
16657         if (len <= 0) {
16658                 /*
16659                  * If FIN has been sent but not acked, but we haven't been
16660                  * called to retransmit, len will be < 0.  Otherwise, window
16661                  * shrank after we sent into it.  If window shrank to 0,
16662                  * cancel pending retransmit, pull snd_nxt back to (closed)
16663                  * window, and set the persist timer if it isn't already
16664                  * going.  If the window didn't close completely, just wait
16665                  * for an ACK.
16666                  *
16667                  * We also do a general check here to ensure that we will
16668                  * set the persist timer when we have data to send, but a
16669                  * 0-byte window. This makes sure the persist timer is set
16670                  * even if the packet hits one of the "goto send" lines
16671                  * below.
16672                  */
16673                 len = 0;
16674                 if ((tp->snd_wnd == 0) &&
16675                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16676                     (tp->snd_una == tp->snd_max) &&
16677                     (sb_offset < (int)sbavail(sb))) {
16678                         rack_enter_persist(tp, rack, cts);
16679                 }
16680         } else if ((rsm == NULL) &&
16681                    (doing_tlp == 0) &&
16682                    (len < pace_max_seg)) {
16683                 /*
16684                  * We are not sending a maximum sized segment for
16685                  * some reason. Should we not send anything (think
16686                  * sws or persists)?
16687                  */
16688                 if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16689                     (TCPS_HAVEESTABLISHED(tp->t_state)) &&
16690                     (len < minseg) &&
16691                     (len < (int)(sbavail(sb) - sb_offset))) {
16692                         /*
16693                          * Here the rwnd is less than
16694                          * the minimum pacing size, this is not a retransmit,
16695                          * we are established and
16696                          * the send is not the last in the socket buffer
16697                          * we send nothing, and we may enter persists
16698                          * if nothing is outstanding.
16699                          */
16700                         len = 0;
16701                         if (tp->snd_max == tp->snd_una) {
16702                                 /*
16703                                  * Nothing out we can
16704                                  * go into persists.
16705                                  */
16706                                 rack_enter_persist(tp, rack, cts);
16707                         }
16708                      } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
16709                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16710                            (len < (int)(sbavail(sb) - sb_offset)) &&
16711                            (len < minseg)) {
16712                         /*
16713                          * Here we are not retransmitting, and
16714                          * the cwnd is not so small that we could
16715                          * not send at least a min size (rxt timer
16716                          * not having gone off), We have 2 segments or
16717                          * more already in flight, its not the tail end
16718                          * of the socket buffer  and the cwnd is blocking
16719                          * us from sending out a minimum pacing segment size.
16720                          * Lets not send anything.
16721                          */
16722                         len = 0;
16723                 } else if (((tp->snd_wnd - ctf_outstanding(tp)) <
16724                             min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
16725                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
16726                            (len < (int)(sbavail(sb) - sb_offset)) &&
16727                            (TCPS_HAVEESTABLISHED(tp->t_state))) {
16728                         /*
16729                          * Here we have a send window but we have
16730                          * filled it up and we can't send another pacing segment.
16731                          * We also have in flight more than 2 segments
16732                          * and we are not completing the sb i.e. we allow
16733                          * the last bytes of the sb to go out even if
16734                          * its not a full pacing segment.
16735                          */
16736                         len = 0;
16737                 } else if ((rack->r_ctl.crte != NULL) &&
16738                            (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
16739                            (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
16740                            (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
16741                            (len < (int)(sbavail(sb) - sb_offset))) {
16742                         /*
16743                          * Here we are doing hardware pacing, this is not a TLP,
16744                          * we are not sending a pace max segment size, there is rwnd
16745                          * room to send at least N pace_max_seg, the cwnd is greater
16746                          * than or equal to a full pacing segments plus 4 mss and we have 2 or
16747                          * more segments in flight and its not the tail of the socket buffer.
16748                          *
16749                          * We don't want to send instead we need to get more ack's in to
16750                          * allow us to send a full pacing segment. Normally, if we are pacing
16751                          * about the right speed, we should have finished our pacing
16752                          * send as most of the acks have come back if we are at the
16753                          * right rate. This is a bit fuzzy since return path delay
16754                          * can delay the acks, which is why we want to make sure we
16755                          * have cwnd space to have a bit more than a max pace segments in flight.
16756                          *
16757                          * If we have not gotten our acks back we are pacing at too high a
16758                          * rate delaying will not hurt and will bring our GP estimate down by
16759                          * injecting the delay. If we don't do this we will send
16760                          * 2 MSS out in response to the acks being clocked in which
16761                          * defeats the point of hw-pacing (i.e. to help us get
16762                          * larger TSO's out).
16763                          */
16764                         len = 0;
16765
16766                 }
16767
16768         }
16769         /* len will be >= 0 after this point. */
16770         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
16771         rack_sndbuf_autoscale(rack);
16772         /*
16773          * Decide if we can use TCP Segmentation Offloading (if supported by
16774          * hardware).
16775          *
16776          * TSO may only be used if we are in a pure bulk sending state.  The
16777          * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
16778          * options prevent using TSO.  With TSO the TCP header is the same
16779          * (except for the sequence number) for all generated packets.  This
16780          * makes it impossible to transmit any options which vary per
16781          * generated segment or packet.
16782          *
16783          * IPv4 handling has a clear separation of ip options and ip header
16784          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
16785          * the right thing below to provide length of just ip options and thus
16786          * checking for ipoptlen is enough to decide if ip options are present.
16787          */
16788         ipoptlen = 0;
16789 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16790         /*
16791          * Pre-calculate here as we save another lookup into the darknesses
16792          * of IPsec that way and can actually decide if TSO is ok.
16793          */
16794 #ifdef INET6
16795         if (isipv6 && IPSEC_ENABLED(ipv6))
16796                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
16797 #ifdef INET
16798         else
16799 #endif
16800 #endif                          /* INET6 */
16801 #ifdef INET
16802                 if (IPSEC_ENABLED(ipv4))
16803                         ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
16804 #endif                          /* INET */
16805 #endif
16806
16807 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16808         ipoptlen += ipsec_optlen;
16809 #endif
16810         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
16811             (tp->t_port == 0) &&
16812             ((tp->t_flags & TF_SIGNATURE) == 0) &&
16813             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
16814             ipoptlen == 0)
16815                 tso = 1;
16816         {
16817                 uint32_t outstanding;
16818
16819                 outstanding = tp->snd_max - tp->snd_una;
16820                 if (tp->t_flags & TF_SENTFIN) {
16821                         /*
16822                          * If we sent a fin, snd_max is 1 higher than
16823                          * snd_una
16824                          */
16825                         outstanding--;
16826                 }
16827                 if (sack_rxmit) {
16828                         if ((rsm->r_flags & RACK_HAS_FIN) == 0)
16829                                 flags &= ~TH_FIN;
16830                 } else {
16831                         if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
16832                                    sbused(sb)))
16833                                 flags &= ~TH_FIN;
16834                 }
16835         }
16836         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
16837             (long)TCP_MAXWIN << tp->rcv_scale);
16838
16839         /*
16840          * Sender silly window avoidance.   We transmit under the following
16841          * conditions when len is non-zero:
16842          *
16843          * - We have a full segment (or more with TSO) - This is the last
16844          * buffer in a write()/send() and we are either idle or running
16845          * NODELAY - we've timed out (e.g. persist timer) - we have more
16846          * then 1/2 the maximum send window's worth of data (receiver may be
16847          * limited the window size) - we need to retransmit
16848          */
16849         if (len) {
16850                 if (len >= segsiz) {
16851                         goto send;
16852                 }
16853                 /*
16854                  * NOTE! on localhost connections an 'ack' from the remote
16855                  * end may occur synchronously with the output and cause us
16856                  * to flush a buffer queued with moretocome.  XXX
16857                  *
16858                  */
16859                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
16860                     (idle || (tp->t_flags & TF_NODELAY)) &&
16861                     ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
16862                     (tp->t_flags & TF_NOPUSH) == 0) {
16863                         pass = 2;
16864                         goto send;
16865                 }
16866                 if ((tp->snd_una == tp->snd_max) && len) {      /* Nothing outstanding */
16867                         pass = 22;
16868                         goto send;
16869                 }
16870                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
16871                         pass = 4;
16872                         goto send;
16873                 }
16874                 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */
16875                         pass = 5;
16876                         goto send;
16877                 }
16878                 if (sack_rxmit) {
16879                         pass = 6;
16880                         goto send;
16881                 }
16882                 if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
16883                     (ctf_outstanding(tp) < (segsiz * 2))) {
16884                         /*
16885                          * We have less than two MSS outstanding (delayed ack)
16886                          * and our rwnd will not let us send a full sized
16887                          * MSS. Lets go ahead and let this small segment
16888                          * out because we want to try to have at least two
16889                          * packets inflight to not be caught by delayed ack.
16890                          */
16891                         pass = 12;
16892                         goto send;
16893                 }
16894         }
16895         /*
16896          * Sending of standalone window updates.
16897          *
16898          * Window updates are important when we close our window due to a
16899          * full socket buffer and are opening it again after the application
16900          * reads data from it.  Once the window has opened again and the
16901          * remote end starts to send again the ACK clock takes over and
16902          * provides the most current window information.
16903          *
16904          * We must avoid the silly window syndrome whereas every read from
16905          * the receive buffer, no matter how small, causes a window update
16906          * to be sent.  We also should avoid sending a flurry of window
16907          * updates when the socket buffer had queued a lot of data and the
16908          * application is doing small reads.
16909          *
16910          * Prevent a flurry of pointless window updates by only sending an
16911          * update when we can increase the advertized window by more than
16912          * 1/4th of the socket buffer capacity.  When the buffer is getting
16913          * full or is very small be more aggressive and send an update
16914          * whenever we can increase by two mss sized segments. In all other
16915          * situations the ACK's to new incoming data will carry further
16916          * window increases.
16917          *
16918          * Don't send an independent window update if a delayed ACK is
16919          * pending (it will get piggy-backed on it) or the remote side
16920          * already has done a half-close and won't send more data.  Skip
16921          * this if the connection is in T/TCP half-open state.
16922          */
16923         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
16924             !(tp->t_flags & TF_DELACK) &&
16925             !TCPS_HAVERCVDFIN(tp->t_state)) {
16926                 /*
16927                  * "adv" is the amount we could increase the window, taking
16928                  * into account that we are limited by TCP_MAXWIN <<
16929                  * tp->rcv_scale.
16930                  */
16931                 int32_t adv;
16932                 int oldwin;
16933
16934                 adv = recwin;
16935                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
16936                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
16937                         if (adv > oldwin)
16938                             adv -= oldwin;
16939                         else {
16940                                 /* We can't increase the window */
16941                                 adv = 0;
16942                         }
16943                 } else
16944                         oldwin = 0;
16945
16946                 /*
16947                  * If the new window size ends up being the same as or less
16948                  * than the old size when it is scaled, then don't force
16949                  * a window update.
16950                  */
16951                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
16952                         goto dontupdate;
16953
16954                 if (adv >= (int32_t)(2 * segsiz) &&
16955                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
16956                      recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
16957                      so->so_rcv.sb_hiwat <= 8 * segsiz)) {
16958                         pass = 7;
16959                         goto send;
16960                 }
16961                 if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
16962                         pass = 23;
16963                         goto send;
16964                 }
16965         }
16966 dontupdate:
16967
16968         /*
16969          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
16970          * is also a catch-all for the retransmit timer timeout case.
16971          */
16972         if (tp->t_flags & TF_ACKNOW) {
16973                 pass = 8;
16974                 goto send;
16975         }
16976         if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
16977                 pass = 9;
16978                 goto send;
16979         }
16980         /*
16981          * If our state indicates that FIN should be sent and we have not
16982          * yet done so, then we need to send.
16983          */
16984         if ((flags & TH_FIN) &&
16985             (tp->snd_nxt == tp->snd_una)) {
16986                 pass = 11;
16987                 goto send;
16988         }
16989         /*
16990          * No reason to send a segment, just return.
16991          */
16992 just_return:
16993         SOCKBUF_UNLOCK(sb);
16994 just_return_nolock:
16995         {
16996                 int app_limited = CTF_JR_SENT_DATA;
16997
16998                 if (tot_len_this_send > 0) {
16999                         /* Make sure snd_nxt is up to max */
17000                         rack->r_ctl.fsb.recwin = recwin;
17001                         slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17002                         if ((error == 0) &&
17003                             rack_use_rfo &&
17004                             ((flags & (TH_SYN|TH_FIN)) == 0) &&
17005                             (ipoptlen == 0) &&
17006                             (tp->snd_nxt == tp->snd_max) &&
17007                             (tp->rcv_numsacks == 0) &&
17008                             rack->r_fsb_inited &&
17009                             TCPS_HAVEESTABLISHED(tp->t_state) &&
17010                             (rack->r_must_retran == 0) &&
17011                             ((tp->t_flags & TF_NEEDFIN) == 0) &&
17012                             (len > 0) && (orig_len > 0) &&
17013                             (orig_len > len) &&
17014                             ((orig_len - len) >= segsiz) &&
17015                             ((optlen == 0) ||
17016                              ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17017                                 /* We can send at least one more MSS using our fsb */
17018
17019                                 rack->r_fast_output = 1;
17020                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17021                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17022                                 rack->r_ctl.fsb.tcp_flags = flags;
17023                                 rack->r_ctl.fsb.left_to_send = orig_len - len;
17024                                 KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17025                                         ("rack:%p left_to_send:%u sbavail:%u out:%u",
17026                                         rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17027                                          (tp->snd_max - tp->snd_una)));
17028                                 if (rack->r_ctl.fsb.left_to_send < segsiz)
17029                                         rack->r_fast_output = 0;
17030                                 else {
17031                                         if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17032                                                 rack->r_ctl.fsb.rfo_apply_push = 1;
17033                                         else
17034                                                 rack->r_ctl.fsb.rfo_apply_push = 0;
17035                                 }
17036                         } else
17037                                 rack->r_fast_output = 0;
17038
17039
17040                         rack_log_fsb(rack, tp, so, flags,
17041                                      ipoptlen, orig_len, len, 0,
17042                                      1, optlen, __LINE__, 1);
17043                         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17044                                 tp->snd_nxt = tp->snd_max;
17045                 } else {
17046                         int end_window = 0;
17047                         uint32_t seq = tp->gput_ack;
17048
17049                         rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17050                         if (rsm) {
17051                                 /*
17052                                  * Mark the last sent that we just-returned (hinting
17053                                  * that delayed ack may play a role in any rtt measurement).
17054                                  */
17055                                 rsm->r_just_ret = 1;
17056                         }
17057                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17058                         rack->r_ctl.rc_agg_delayed = 0;
17059                         rack->r_early = 0;
17060                         rack->r_late = 0;
17061                         rack->r_ctl.rc_agg_early = 0;
17062                         if ((ctf_outstanding(tp) +
17063                              min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17064                                  minseg)) >= tp->snd_wnd) {
17065                                 /* We are limited by the rwnd */
17066                                 app_limited = CTF_JR_RWND_LIMITED;
17067                                 if (IN_FASTRECOVERY(tp->t_flags))
17068                                     rack->r_ctl.rc_prr_sndcnt = 0;
17069                         } else if (ctf_outstanding(tp) >= sbavail(sb)) {
17070                                 /* We are limited by whats available -- app limited */
17071                                 app_limited = CTF_JR_APP_LIMITED;
17072                                 if (IN_FASTRECOVERY(tp->t_flags))
17073                                     rack->r_ctl.rc_prr_sndcnt = 0;
17074                         } else if ((idle == 0) &&
17075                                    ((tp->t_flags & TF_NODELAY) == 0) &&
17076                                    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17077                                    (len < segsiz)) {
17078                                 /*
17079                                  * No delay is not on and the
17080                                  * user is sending less than 1MSS. This
17081                                  * brings out SWS avoidance so we
17082                                  * don't send. Another app-limited case.
17083                                  */
17084                                 app_limited = CTF_JR_APP_LIMITED;
17085                         } else if (tp->t_flags & TF_NOPUSH) {
17086                                 /*
17087                                  * The user has requested no push of
17088                                  * the last segment and we are
17089                                  * at the last segment. Another app
17090                                  * limited case.
17091                                  */
17092                                 app_limited = CTF_JR_APP_LIMITED;
17093                         } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17094                                 /* Its the cwnd */
17095                                 app_limited = CTF_JR_CWND_LIMITED;
17096                         } else if (IN_FASTRECOVERY(tp->t_flags) &&
17097                                    (rack->rack_no_prr == 0) &&
17098                                    (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17099                                 app_limited = CTF_JR_PRR;
17100                         } else {
17101                                 /* Now why here are we not sending? */
17102 #ifdef NOW
17103 #ifdef INVARIANTS
17104                                 panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17105 #endif
17106 #endif
17107                                 app_limited = CTF_JR_ASSESSING;
17108                         }
17109                         /*
17110                          * App limited in some fashion, for our pacing GP
17111                          * measurements we don't want any gap (even cwnd).
17112                          * Close  down the measurement window.
17113                          */
17114                         if (rack_cwnd_block_ends_measure &&
17115                             ((app_limited == CTF_JR_CWND_LIMITED) ||
17116                              (app_limited == CTF_JR_PRR))) {
17117                                 /*
17118                                  * The reason we are not sending is
17119                                  * the cwnd (or prr). We have been configured
17120                                  * to end the measurement window in
17121                                  * this case.
17122                                  */
17123                                 end_window = 1;
17124                         } else if (rack_rwnd_block_ends_measure &&
17125                                    (app_limited == CTF_JR_RWND_LIMITED)) {
17126                                 /*
17127                                  * We are rwnd limited and have been
17128                                  * configured to end the measurement
17129                                  * window in this case.
17130                                  */
17131                                 end_window = 1;
17132                         } else if (app_limited == CTF_JR_APP_LIMITED) {
17133                                 /*
17134                                  * A true application limited period, we have
17135                                  * ran out of data.
17136                                  */
17137                                 end_window = 1;
17138                         } else if (app_limited == CTF_JR_ASSESSING) {
17139                                 /*
17140                                  * In the assessing case we hit the end of
17141                                  * the if/else and had no known reason
17142                                  * This will panic us under invariants..
17143                                  *
17144                                  * If we get this out in logs we need to
17145                                  * investagate which reason we missed.
17146                                  */
17147                                 end_window = 1;
17148                         }
17149                         if (end_window) {
17150                                 uint8_t log = 0;
17151
17152                                 if ((tp->t_flags & TF_GPUTINPROG) &&
17153                                     SEQ_GT(tp->gput_ack, tp->snd_max)) {
17154                                         /* Mark the last packet has app limited */
17155                                         tp->gput_ack = tp->snd_max;
17156                                         log = 1;
17157                                 }
17158                                 rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17159                                 if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17160                                         if (rack->r_ctl.rc_app_limited_cnt == 0)
17161                                                 rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17162                                         else {
17163                                                 /*
17164                                                  * Go out to the end app limited and mark
17165                                                  * this new one as next and move the end_appl up
17166                                                  * to this guy.
17167                                                  */
17168                                                 if (rack->r_ctl.rc_end_appl)
17169                                                         rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17170                                                 rack->r_ctl.rc_end_appl = rsm;
17171                                         }
17172                                         rsm->r_flags |= RACK_APP_LIMITED;
17173                                         rack->r_ctl.rc_app_limited_cnt++;
17174                                 }
17175                                 if (log)
17176                                         rack_log_pacing_delay_calc(rack,
17177                                                                    rack->r_ctl.rc_app_limited_cnt, seq,
17178                                                                    tp->gput_ack, 0, 0, 4, __LINE__, NULL);
17179                         }
17180                 }
17181                 if (slot) {
17182                         /* set the rack tcb into the slot N */
17183                         counter_u64_add(rack_paced_segments, 1);
17184                 } else if (tot_len_this_send) {
17185                         counter_u64_add(rack_unpaced_segments, 1);
17186                 }
17187                 /* Check if we need to go into persists or not */
17188                 if ((tp->snd_max == tp->snd_una) &&
17189                     TCPS_HAVEESTABLISHED(tp->t_state) &&
17190                     sbavail(sb) &&
17191                     (sbavail(sb) > tp->snd_wnd) &&
17192                     (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17193                         /* Yes lets make sure to move to persist before timer-start */
17194                         rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17195                 }
17196                 rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17197                 rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17198         }
17199 #ifdef NETFLIX_SHARED_CWND
17200         if ((sbavail(sb) == 0) &&
17201             rack->r_ctl.rc_scw) {
17202                 tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17203                 rack->rack_scwnd_is_idle = 1;
17204         }
17205 #endif
17206 #ifdef TCP_ACCOUNTING
17207         if (tot_len_this_send > 0) {
17208                 crtsc = get_cyclecount();
17209                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17210                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
17211                 }
17212                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17213                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17214                         tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17215                 }
17216                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17217                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17218                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17219                 }
17220                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17221         } else {
17222                 crtsc = get_cyclecount();
17223                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17224                         tp->tcp_cnt_counters[SND_LIMITED]++;
17225                 }
17226                 counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17227                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17228                         tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17229                 }
17230                 counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17231         }
17232         sched_unpin();
17233 #endif
17234         return (0);
17235
17236 send:
17237         if (rsm || sack_rxmit)
17238                 counter_u64_add(rack_nfto_resend, 1);
17239         else
17240                 counter_u64_add(rack_non_fto_send, 1);
17241         if ((flags & TH_FIN) &&
17242             sbavail(sb)) {
17243                 /*
17244                  * We do not transmit a FIN
17245                  * with data outstanding. We
17246                  * need to make it so all data
17247                  * is acked first.
17248                  */
17249                 flags &= ~TH_FIN;
17250         }
17251         /* Enforce stack imposed max seg size if we have one */
17252         if (rack->r_ctl.rc_pace_max_segs &&
17253             (len > rack->r_ctl.rc_pace_max_segs)) {
17254                 mark = 1;
17255                 len = rack->r_ctl.rc_pace_max_segs;
17256         }
17257         SOCKBUF_LOCK_ASSERT(sb);
17258         if (len > 0) {
17259                 if (len >= segsiz)
17260                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17261                 else
17262                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17263         }
17264         /*
17265          * Before ESTABLISHED, force sending of initial options unless TCP
17266          * set not to do any options. NOTE: we assume that the IP/TCP header
17267          * plus TCP options always fit in a single mbuf, leaving room for a
17268          * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17269          * + optlen <= MCLBYTES
17270          */
17271         optlen = 0;
17272 #ifdef INET6
17273         if (isipv6)
17274                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17275         else
17276 #endif
17277                 hdrlen = sizeof(struct tcpiphdr);
17278
17279         /*
17280          * Compute options for segment. We only have to care about SYN and
17281          * established connection segments.  Options for SYN-ACK segments
17282          * are handled in TCP syncache.
17283          */
17284         to.to_flags = 0;
17285         if ((tp->t_flags & TF_NOOPT) == 0) {
17286                 /* Maximum segment size. */
17287                 if (flags & TH_SYN) {
17288                         tp->snd_nxt = tp->iss;
17289                         to.to_mss = tcp_mssopt(&inp->inp_inc);
17290                         if (tp->t_port)
17291                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
17292                         to.to_flags |= TOF_MSS;
17293
17294                         /*
17295                          * On SYN or SYN|ACK transmits on TFO connections,
17296                          * only include the TFO option if it is not a
17297                          * retransmit, as the presence of the TFO option may
17298                          * have caused the original SYN or SYN|ACK to have
17299                          * been dropped by a middlebox.
17300                          */
17301                         if (IS_FASTOPEN(tp->t_flags) &&
17302                             (tp->t_rxtshift == 0)) {
17303                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
17304                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17305                                         to.to_tfo_cookie =
17306                                                 (u_int8_t *)&tp->t_tfo_cookie.server;
17307                                         to.to_flags |= TOF_FASTOPEN;
17308                                         wanted_cookie = 1;
17309                                 } else if (tp->t_state == TCPS_SYN_SENT) {
17310                                         to.to_tfo_len =
17311                                                 tp->t_tfo_client_cookie_len;
17312                                         to.to_tfo_cookie =
17313                                                 tp->t_tfo_cookie.client;
17314                                         to.to_flags |= TOF_FASTOPEN;
17315                                         wanted_cookie = 1;
17316                                         /*
17317                                          * If we wind up having more data to
17318                                          * send with the SYN than can fit in
17319                                          * one segment, don't send any more
17320                                          * until the SYN|ACK comes back from
17321                                          * the other end.
17322                                          */
17323                                         sendalot = 0;
17324                                 }
17325                         }
17326                 }
17327                 /* Window scaling. */
17328                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
17329                         to.to_wscale = tp->request_r_scale;
17330                         to.to_flags |= TOF_SCALE;
17331                 }
17332                 /* Timestamps. */
17333                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
17334                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
17335                         to.to_tsval = ms_cts + tp->ts_offset;
17336                         to.to_tsecr = tp->ts_recent;
17337                         to.to_flags |= TOF_TS;
17338                 }
17339                 /* Set receive buffer autosizing timestamp. */
17340                 if (tp->rfbuf_ts == 0 &&
17341                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
17342                         tp->rfbuf_ts = tcp_ts_getticks();
17343                 /* Selective ACK's. */
17344                 if (tp->t_flags & TF_SACK_PERMIT) {
17345                         if (flags & TH_SYN)
17346                                 to.to_flags |= TOF_SACKPERM;
17347                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17348                                  tp->rcv_numsacks > 0) {
17349                                 to.to_flags |= TOF_SACK;
17350                                 to.to_nsacks = tp->rcv_numsacks;
17351                                 to.to_sacks = (u_char *)tp->sackblks;
17352                         }
17353                 }
17354 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17355                 /* TCP-MD5 (RFC2385). */
17356                 if (tp->t_flags & TF_SIGNATURE)
17357                         to.to_flags |= TOF_SIGNATURE;
17358 #endif                          /* TCP_SIGNATURE */
17359
17360                 /* Processing the options. */
17361                 hdrlen += optlen = tcp_addoptions(&to, opt);
17362                 /*
17363                  * If we wanted a TFO option to be added, but it was unable
17364                  * to fit, ensure no data is sent.
17365                  */
17366                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
17367                     !(to.to_flags & TOF_FASTOPEN))
17368                         len = 0;
17369         }
17370         if (tp->t_port) {
17371                 if (V_tcp_udp_tunneling_port == 0) {
17372                         /* The port was removed?? */
17373                         SOCKBUF_UNLOCK(&so->so_snd);
17374 #ifdef TCP_ACCOUNTING
17375                         crtsc = get_cyclecount();
17376                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17377                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
17378                         }
17379                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
17380                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17381                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
17382                         }
17383                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
17384                         sched_unpin();
17385 #endif
17386                         return (EHOSTUNREACH);
17387                 }
17388                 hdrlen += sizeof(struct udphdr);
17389         }
17390 #ifdef INET6
17391         if (isipv6)
17392                 ipoptlen = ip6_optlen(tp->t_inpcb);
17393         else
17394 #endif
17395                 if (tp->t_inpcb->inp_options)
17396                         ipoptlen = tp->t_inpcb->inp_options->m_len -
17397                                 offsetof(struct ipoption, ipopt_list);
17398                 else
17399                         ipoptlen = 0;
17400 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17401         ipoptlen += ipsec_optlen;
17402 #endif
17403
17404         /*
17405          * Adjust data length if insertion of options will bump the packet
17406          * length beyond the t_maxseg length. Clear the FIN bit because we
17407          * cut off the tail of the segment.
17408          */
17409         if (len + optlen + ipoptlen > tp->t_maxseg) {
17410                 if (tso) {
17411                         uint32_t if_hw_tsomax;
17412                         uint32_t moff;
17413                         int32_t max_len;
17414
17415                         /* extract TSO information */
17416                         if_hw_tsomax = tp->t_tsomax;
17417                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
17418                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
17419                         KASSERT(ipoptlen == 0,
17420                                 ("%s: TSO can't do IP options", __func__));
17421
17422                         /*
17423                          * Check if we should limit by maximum payload
17424                          * length:
17425                          */
17426                         if (if_hw_tsomax != 0) {
17427                                 /* compute maximum TSO length */
17428                                 max_len = (if_hw_tsomax - hdrlen -
17429                                            max_linkhdr);
17430                                 if (max_len <= 0) {
17431                                         len = 0;
17432                                 } else if (len > max_len) {
17433                                         sendalot = 1;
17434                                         len = max_len;
17435                                         mark = 2;
17436                                 }
17437                         }
17438                         /*
17439                          * Prevent the last segment from being fractional
17440                          * unless the send sockbuf can be emptied:
17441                          */
17442                         max_len = (tp->t_maxseg - optlen);
17443                         if ((sb_offset + len) < sbavail(sb)) {
17444                                 moff = len % (u_int)max_len;
17445                                 if (moff != 0) {
17446                                         mark = 3;
17447                                         len -= moff;
17448                                 }
17449                         }
17450                         /*
17451                          * In case there are too many small fragments don't
17452                          * use TSO:
17453                          */
17454                         if (len <= segsiz) {
17455                                 mark = 4;
17456                                 tso = 0;
17457                         }
17458                         /*
17459                          * Send the FIN in a separate segment after the bulk
17460                          * sending is done. We don't trust the TSO
17461                          * implementations to clear the FIN flag on all but
17462                          * the last segment.
17463                          */
17464                         if (tp->t_flags & TF_NEEDFIN) {
17465                                 sendalot = 4;
17466                         }
17467                 } else {
17468                         mark = 5;
17469                         if (optlen + ipoptlen >= tp->t_maxseg) {
17470                                 /*
17471                                  * Since we don't have enough space to put
17472                                  * the IP header chain and the TCP header in
17473                                  * one packet as required by RFC 7112, don't
17474                                  * send it. Also ensure that at least one
17475                                  * byte of the payload can be put into the
17476                                  * TCP segment.
17477                                  */
17478                                 SOCKBUF_UNLOCK(&so->so_snd);
17479                                 error = EMSGSIZE;
17480                                 sack_rxmit = 0;
17481                                 goto out;
17482                         }
17483                         len = tp->t_maxseg - optlen - ipoptlen;
17484                         sendalot = 5;
17485                 }
17486         } else {
17487                 tso = 0;
17488                 mark = 6;
17489         }
17490         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
17491                 ("%s: len > IP_MAXPACKET", __func__));
17492 #ifdef DIAGNOSTIC
17493 #ifdef INET6
17494         if (max_linkhdr + hdrlen > MCLBYTES)
17495 #else
17496                 if (max_linkhdr + hdrlen > MHLEN)
17497 #endif
17498                         panic("tcphdr too big");
17499 #endif
17500
17501         /*
17502          * This KASSERT is here to catch edge cases at a well defined place.
17503          * Before, those had triggered (random) panic conditions further
17504          * down.
17505          */
17506         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17507         if ((len == 0) &&
17508             (flags & TH_FIN) &&
17509             (sbused(sb))) {
17510                 /*
17511                  * We have outstanding data, don't send a fin by itself!.
17512                  */
17513                 goto just_return;
17514         }
17515         /*
17516          * Grab a header mbuf, attaching a copy of data to be transmitted,
17517          * and initialize the header from the template for sends on this
17518          * connection.
17519          */
17520         hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
17521         if (len) {
17522                 uint32_t max_val;
17523                 uint32_t moff;
17524
17525                 if (rack->r_ctl.rc_pace_max_segs)
17526                         max_val = rack->r_ctl.rc_pace_max_segs;
17527                 else if (rack->rc_user_set_max_segs)
17528                         max_val = rack->rc_user_set_max_segs * segsiz;
17529                 else
17530                         max_val = len;
17531                 /*
17532                  * We allow a limit on sending with hptsi.
17533                  */
17534                 if (len > max_val) {
17535                         mark = 7;
17536                         len = max_val;
17537                 }
17538 #ifdef INET6
17539                 if (MHLEN < hdrlen + max_linkhdr)
17540                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
17541                 else
17542 #endif
17543                         m = m_gethdr(M_NOWAIT, MT_DATA);
17544
17545                 if (m == NULL) {
17546                         SOCKBUF_UNLOCK(sb);
17547                         error = ENOBUFS;
17548                         sack_rxmit = 0;
17549                         goto out;
17550                 }
17551                 m->m_data += max_linkhdr;
17552                 m->m_len = hdrlen;
17553
17554                 /*
17555                  * Start the m_copy functions from the closest mbuf to the
17556                  * sb_offset in the socket buffer chain.
17557                  */
17558                 mb = sbsndptr_noadv(sb, sb_offset, &moff);
17559                 s_mb = mb;
17560                 s_moff = moff;
17561                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
17562                         m_copydata(mb, moff, (int)len,
17563                                    mtod(m, caddr_t)+hdrlen);
17564                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17565                                 sbsndptr_adv(sb, mb, len);
17566                         m->m_len += len;
17567                 } else {
17568                         struct sockbuf *msb;
17569
17570                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
17571                                 msb = NULL;
17572                         else
17573                                 msb = sb;
17574                         m->m_next = tcp_m_copym(
17575                                 mb, moff, &len,
17576                                 if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
17577                                 ((rsm == NULL) ? hw_tls : 0)
17578 #ifdef NETFLIX_COPY_ARGS
17579                                 , &filled_all
17580 #endif
17581                                 );
17582                         if (len <= (tp->t_maxseg - optlen)) {
17583                                 /*
17584                                  * Must have ran out of mbufs for the copy
17585                                  * shorten it to no longer need tso. Lets
17586                                  * not put on sendalot since we are low on
17587                                  * mbufs.
17588                                  */
17589                                 tso = 0;
17590                         }
17591                         if (m->m_next == NULL) {
17592                                 SOCKBUF_UNLOCK(sb);
17593                                 (void)m_free(m);
17594                                 error = ENOBUFS;
17595                                 sack_rxmit = 0;
17596                                 goto out;
17597                         }
17598                 }
17599                 if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
17600                         if (rsm && (rsm->r_flags & RACK_TLP)) {
17601                                 /*
17602                                  * TLP should not count in retran count, but
17603                                  * in its own bin
17604                                  */
17605                                 counter_u64_add(rack_tlp_retran, 1);
17606                                 counter_u64_add(rack_tlp_retran_bytes, len);
17607                         } else {
17608                                 tp->t_sndrexmitpack++;
17609                                 KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
17610                                 KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
17611                         }
17612 #ifdef STATS
17613                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
17614                                                  len);
17615 #endif
17616                 } else {
17617                         KMOD_TCPSTAT_INC(tcps_sndpack);
17618                         KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
17619 #ifdef STATS
17620                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
17621                                                  len);
17622 #endif
17623                 }
17624                 /*
17625                  * If we're sending everything we've got, set PUSH. (This
17626                  * will keep happy those implementations which only give
17627                  * data to the user when a buffer fills or a PUSH comes in.)
17628                  */
17629                 if (sb_offset + len == sbused(sb) &&
17630                     sbused(sb) &&
17631                     !(flags & TH_SYN)) {
17632                         flags |= TH_PUSH;
17633                         add_flag |= RACK_HAD_PUSH;
17634                 }
17635
17636                 SOCKBUF_UNLOCK(sb);
17637         } else {
17638                 SOCKBUF_UNLOCK(sb);
17639                 if (tp->t_flags & TF_ACKNOW)
17640                         KMOD_TCPSTAT_INC(tcps_sndacks);
17641                 else if (flags & (TH_SYN | TH_FIN | TH_RST))
17642                         KMOD_TCPSTAT_INC(tcps_sndctrl);
17643                 else
17644                         KMOD_TCPSTAT_INC(tcps_sndwinup);
17645
17646                 m = m_gethdr(M_NOWAIT, MT_DATA);
17647                 if (m == NULL) {
17648                         error = ENOBUFS;
17649                         sack_rxmit = 0;
17650                         goto out;
17651                 }
17652 #ifdef INET6
17653                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
17654                     MHLEN >= hdrlen) {
17655                         M_ALIGN(m, hdrlen);
17656                 } else
17657 #endif
17658                         m->m_data += max_linkhdr;
17659                 m->m_len = hdrlen;
17660         }
17661         SOCKBUF_UNLOCK_ASSERT(sb);
17662         m->m_pkthdr.rcvif = (struct ifnet *)0;
17663 #ifdef MAC
17664         mac_inpcb_create_mbuf(inp, m);
17665 #endif
17666         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
17667 #ifdef INET6
17668                 if (isipv6)
17669                         ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
17670                 else
17671 #endif                          /* INET6 */
17672                         ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
17673                 th = rack->r_ctl.fsb.th;
17674                 udp = rack->r_ctl.fsb.udp;
17675                 if (udp) {
17676                         if (isipv6)
17677                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17678                         else
17679                                 ulen = hdrlen + len - sizeof(struct ip);
17680                         udp->uh_ulen = htons(ulen);
17681                 }
17682         } else {
17683 #ifdef INET6
17684                 if (isipv6) {
17685                         ip6 = mtod(m, struct ip6_hdr *);
17686                         if (tp->t_port) {
17687                                 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
17688                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17689                                 udp->uh_dport = tp->t_port;
17690                                 ulen = hdrlen + len - sizeof(struct ip6_hdr);
17691                                 udp->uh_ulen = htons(ulen);
17692                                 th = (struct tcphdr *)(udp + 1);
17693                         } else
17694                                 th = (struct tcphdr *)(ip6 + 1);
17695                         tcpip_fillheaders(inp, tp->t_port, ip6, th);
17696                 } else
17697 #endif                          /* INET6 */
17698                 {
17699                         ip = mtod(m, struct ip *);
17700 #ifdef TCPDEBUG
17701                         ipov = (struct ipovly *)ip;
17702 #endif
17703                         if (tp->t_port) {
17704                                 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
17705                                 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
17706                                 udp->uh_dport = tp->t_port;
17707                                 ulen = hdrlen + len - sizeof(struct ip);
17708                                 udp->uh_ulen = htons(ulen);
17709                                 th = (struct tcphdr *)(udp + 1);
17710                         } else
17711                                 th = (struct tcphdr *)(ip + 1);
17712                         tcpip_fillheaders(inp, tp->t_port, ip, th);
17713                 }
17714         }
17715         /*
17716          * Fill in fields, remembering maximum advertised window for use in
17717          * delaying messages about window sizes. If resending a FIN, be sure
17718          * not to use a new sequence number.
17719          */
17720         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
17721             tp->snd_nxt == tp->snd_max)
17722                 tp->snd_nxt--;
17723         /*
17724          * If we are starting a connection, send ECN setup SYN packet. If we
17725          * are on a retransmit, we may resend those bits a number of times
17726          * as per RFC 3168.
17727          */
17728         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
17729                 if (tp->t_rxtshift >= 1) {
17730                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
17731                                 flags |= TH_ECE | TH_CWR;
17732                 } else
17733                         flags |= TH_ECE | TH_CWR;
17734         }
17735         /* Handle parallel SYN for ECN */
17736         if ((tp->t_state == TCPS_SYN_RECEIVED) &&
17737             (tp->t_flags2 & TF2_ECN_SND_ECE)) {
17738                 flags |= TH_ECE;
17739                 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
17740         }
17741         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
17742             (tp->t_flags2 & TF2_ECN_PERMIT)) {
17743                 /*
17744                  * If the peer has ECN, mark data packets with ECN capable
17745                  * transmission (ECT). Ignore pure ack packets,
17746                  * retransmissions.
17747                  */
17748                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
17749                     (sack_rxmit == 0)) {
17750 #ifdef INET6
17751                         if (isipv6)
17752                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
17753                         else
17754 #endif
17755                                 ip->ip_tos |= IPTOS_ECN_ECT0;
17756                         KMOD_TCPSTAT_INC(tcps_ecn_ect0);
17757                         /*
17758                          * Reply with proper ECN notifications.
17759                          * Only set CWR on new data segments.
17760                          */
17761                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
17762                                 flags |= TH_CWR;
17763                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
17764                         }
17765                 }
17766                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
17767                         flags |= TH_ECE;
17768         }
17769         /*
17770          * If we are doing retransmissions, then snd_nxt will not reflect
17771          * the first unsent octet.  For ACK only packets, we do not want the
17772          * sequence number of the retransmitted packet, we want the sequence
17773          * number of the next unsent octet.  So, if there is no data (and no
17774          * SYN or FIN), use snd_max instead of snd_nxt when filling in
17775          * ti_seq.  But if we are in persist state, snd_max might reflect
17776          * one byte beyond the right edge of the window, so use snd_nxt in
17777          * that case, since we know we aren't doing a retransmission.
17778          * (retransmit and persist are mutually exclusive...)
17779          */
17780         if (sack_rxmit == 0) {
17781                 if (len || (flags & (TH_SYN | TH_FIN))) {
17782                         th->th_seq = htonl(tp->snd_nxt);
17783                         rack_seq = tp->snd_nxt;
17784                 } else {
17785                         th->th_seq = htonl(tp->snd_max);
17786                         rack_seq = tp->snd_max;
17787                 }
17788         } else {
17789                 th->th_seq = htonl(rsm->r_start);
17790                 rack_seq = rsm->r_start;
17791         }
17792         th->th_ack = htonl(tp->rcv_nxt);
17793         th->th_flags = flags;
17794         /*
17795          * Calculate receive window.  Don't shrink window, but avoid silly
17796          * window syndrome.
17797          * If a RST segment is sent, advertise a window of zero.
17798          */
17799         if (flags & TH_RST) {
17800                 recwin = 0;
17801         } else {
17802                 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
17803                     recwin < (long)segsiz) {
17804                         recwin = 0;
17805                 }
17806                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
17807                     recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
17808                         recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
17809         }
17810
17811         /*
17812          * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
17813          * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
17814          * handled in syncache.
17815          */
17816         if (flags & TH_SYN)
17817                 th->th_win = htons((u_short)
17818                                    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
17819         else {
17820                 /* Avoid shrinking window with window scaling. */
17821                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
17822                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
17823         }
17824         /*
17825          * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
17826          * window.  This may cause the remote transmitter to stall.  This
17827          * flag tells soreceive() to disable delayed acknowledgements when
17828          * draining the buffer.  This can occur if the receiver is
17829          * attempting to read more data than can be buffered prior to
17830          * transmitting on the connection.
17831          */
17832         if (th->th_win == 0) {
17833                 tp->t_sndzerowin++;
17834                 tp->t_flags |= TF_RXWIN0SENT;
17835         } else
17836                 tp->t_flags &= ~TF_RXWIN0SENT;
17837         tp->snd_up = tp->snd_una;       /* drag it along, its deprecated */
17838         /* Now are we using fsb?, if so copy the template data to the mbuf */
17839         if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
17840                 uint8_t *cpto;
17841
17842                 cpto = mtod(m, uint8_t *);
17843                 memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
17844                 /*
17845                  * We have just copied in:
17846                  * IP/IP6
17847                  * <optional udphdr>
17848                  * tcphdr (no options)
17849                  *
17850                  * We need to grab the correct pointers into the mbuf
17851                  * for both the tcp header, and possibly the udp header (if tunneling).
17852                  * We do this by using the offset in the copy buffer and adding it
17853                  * to the mbuf base pointer (cpto).
17854                  */
17855 #ifdef INET6
17856                 if (isipv6)
17857                         ip6 = mtod(m, struct ip6_hdr *);
17858                 else
17859 #endif                          /* INET6 */
17860                         ip = mtod(m, struct ip *);
17861                 th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
17862                 /* If we have a udp header lets set it into the mbuf as well */
17863                 if (udp)
17864                         udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
17865         }
17866 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
17867         if (to.to_flags & TOF_SIGNATURE) {
17868                 /*
17869                  * Calculate MD5 signature and put it into the place
17870                  * determined before.
17871                  * NOTE: since TCP options buffer doesn't point into
17872                  * mbuf's data, calculate offset and use it.
17873                  */
17874                 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
17875                                                        (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
17876                         /*
17877                          * Do not send segment if the calculation of MD5
17878                          * digest has failed.
17879                          */
17880                         goto out;
17881                 }
17882         }
17883 #endif
17884         if (optlen) {
17885                 bcopy(opt, th + 1, optlen);
17886                 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
17887         }
17888         /*
17889          * Put TCP length in extended header, and then checksum extended
17890          * header and data.
17891          */
17892         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
17893 #ifdef INET6
17894         if (isipv6) {
17895                 /*
17896                  * ip6_plen is not need to be filled now, and will be filled
17897                  * in ip6_output.
17898                  */
17899                 if (tp->t_port) {
17900                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
17901                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17902                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
17903                         th->th_sum = htons(0);
17904                         UDPSTAT_INC(udps_opackets);
17905                 } else {
17906                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
17907                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17908                         th->th_sum = in6_cksum_pseudo(ip6,
17909                                                       sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
17910                                                       0);
17911                 }
17912         }
17913 #endif
17914 #if defined(INET6) && defined(INET)
17915         else
17916 #endif
17917 #ifdef INET
17918         {
17919                 if (tp->t_port) {
17920                         m->m_pkthdr.csum_flags = CSUM_UDP;
17921                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
17922                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
17923                                                 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
17924                         th->th_sum = htons(0);
17925                         UDPSTAT_INC(udps_opackets);
17926                 } else {
17927                         m->m_pkthdr.csum_flags = CSUM_TCP;
17928                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
17929                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
17930                                                ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
17931                                                                         IPPROTO_TCP + len + optlen));
17932                 }
17933                 /* IP version must be set here for ipv4/ipv6 checking later */
17934                 KASSERT(ip->ip_v == IPVERSION,
17935                         ("%s: IP version incorrect: %d", __func__, ip->ip_v));
17936         }
17937 #endif
17938         /*
17939          * Enable TSO and specify the size of the segments. The TCP pseudo
17940          * header checksum is always provided. XXX: Fixme: This is currently
17941          * not the case for IPv6.
17942          */
17943         if (tso) {
17944                 KASSERT(len > tp->t_maxseg - optlen,
17945                         ("%s: len <= tso_segsz", __func__));
17946                 m->m_pkthdr.csum_flags |= CSUM_TSO;
17947                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
17948         }
17949         KASSERT(len + hdrlen == m_length(m, NULL),
17950                 ("%s: mbuf chain different than expected: %d + %u != %u",
17951                  __func__, len, hdrlen, m_length(m, NULL)));
17952
17953 #ifdef TCP_HHOOK
17954         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
17955         hhook_run_tcp_est_out(tp, th, &to, len, tso);
17956 #endif
17957         /* We're getting ready to send; log now. */
17958         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
17959                 union tcp_log_stackspecific log;
17960
17961                 memset(&log.u_bbr, 0, sizeof(log.u_bbr));
17962                 log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
17963                 log.u_bbr.ininput = rack->rc_inp->inp_in_input;
17964                 if (rack->rack_no_prr)
17965                         log.u_bbr.flex1 = 0;
17966                 else
17967                         log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
17968                 log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
17969                 log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
17970                 log.u_bbr.flex4 = orig_len;
17971                 if (filled_all)
17972                         log.u_bbr.flex5 = 0x80000000;
17973                 else
17974                         log.u_bbr.flex5 = 0;
17975                 /* Save off the early/late values */
17976                 log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
17977                 log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
17978                 log.u_bbr.bw_inuse = rack_get_bw(rack);
17979                 if (rsm || sack_rxmit) {
17980                         if (doing_tlp)
17981                                 log.u_bbr.flex8 = 2;
17982                         else
17983                                 log.u_bbr.flex8 = 1;
17984                 } else {
17985                         log.u_bbr.flex8 = 0;
17986                 }
17987                 log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17988                 log.u_bbr.flex7 = mark;
17989                 log.u_bbr.flex7 <<= 8;
17990                 log.u_bbr.flex7 |= pass;
17991                 log.u_bbr.pkts_out = tp->t_maxseg;
17992                 log.u_bbr.timeStamp = cts;
17993                 log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17994                 log.u_bbr.lt_epoch = cwnd_to_use;
17995                 log.u_bbr.delivered = sendalot;
17996                 lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
17997                                      len, &log, false, NULL, NULL, 0, &tv);
17998         } else
17999                 lgb = NULL;
18000
18001         /*
18002          * Fill in IP length and desired time to live and send to IP level.
18003          * There should be a better way to handle ttl and tos; we could keep
18004          * them in the template, but need a way to checksum without them.
18005          */
18006         /*
18007          * m->m_pkthdr.len should have been set before cksum calcuration,
18008          * because in6_cksum() need it.
18009          */
18010 #ifdef INET6
18011         if (isipv6) {
18012                 /*
18013                  * we separately set hoplimit for every segment, since the
18014                  * user might want to change the value via setsockopt. Also,
18015                  * desired default hop limit might be changed via Neighbor
18016                  * Discovery.
18017                  */
18018                 rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18019
18020                 /*
18021                  * Set the packet size here for the benefit of DTrace
18022                  * probes. ip6_output() will set it properly; it's supposed
18023                  * to include the option header lengths as well.
18024                  */
18025                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18026
18027                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18028                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18029                 else
18030                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18031
18032                 if (tp->t_state == TCPS_SYN_SENT)
18033                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18034
18035                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18036                 /* TODO: IPv6 IP6TOS_ECT bit on */
18037                 error = ip6_output(m,
18038 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18039                                    inp->in6p_outputopts,
18040 #else
18041                                    NULL,
18042 #endif
18043                                    &inp->inp_route6,
18044                                    ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18045                                    NULL, NULL, inp);
18046
18047                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18048                         mtu = inp->inp_route6.ro_nh->nh_mtu;
18049         }
18050 #endif                          /* INET6 */
18051 #if defined(INET) && defined(INET6)
18052         else
18053 #endif
18054 #ifdef INET
18055         {
18056                 ip->ip_len = htons(m->m_pkthdr.len);
18057 #ifdef INET6
18058                 if (inp->inp_vflag & INP_IPV6PROTO)
18059                         ip->ip_ttl = in6_selecthlim(inp, NULL);
18060 #endif                          /* INET6 */
18061                 rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18062                 /*
18063                  * If we do path MTU discovery, then we set DF on every
18064                  * packet. This might not be the best thing to do according
18065                  * to RFC3390 Section 2. However the tcp hostcache migitates
18066                  * the problem so it affects only the first tcp connection
18067                  * with a host.
18068                  *
18069                  * NB: Don't set DF on small MTU/MSS to have a safe
18070                  * fallback.
18071                  */
18072                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18073                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18074                         if (tp->t_port == 0 || len < V_tcp_minmss) {
18075                                 ip->ip_off |= htons(IP_DF);
18076                         }
18077                 } else {
18078                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18079                 }
18080
18081                 if (tp->t_state == TCPS_SYN_SENT)
18082                         TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18083
18084                 TCP_PROBE5(send, NULL, tp, ip, tp, th);
18085
18086                 error = ip_output(m,
18087 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18088                                   inp->inp_options,
18089 #else
18090                                   NULL,
18091 #endif
18092                                   &inp->inp_route,
18093                                   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18094                                   inp);
18095                 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18096                         mtu = inp->inp_route.ro_nh->nh_mtu;
18097         }
18098 #endif                          /* INET */
18099
18100 out:
18101         if (lgb) {
18102                 lgb->tlb_errno = error;
18103                 lgb = NULL;
18104         }
18105         /*
18106          * In transmit state, time the transmission and arrange for the
18107          * retransmit.  In persist state, just set snd_max.
18108          */
18109         if (error == 0) {
18110                 rack->forced_ack = 0;   /* If we send something zap the FA flag */
18111                 if (rsm && (doing_tlp == 0)) {
18112                         /* Set we retransmitted */
18113                         rack->rc_gp_saw_rec = 1;
18114                 } else {
18115                         if (cwnd_to_use > tp->snd_ssthresh) {
18116                                 /* Set we sent in CA */
18117                                 rack->rc_gp_saw_ca = 1;
18118                         } else {
18119                                 /* Set we sent in SS */
18120                                 rack->rc_gp_saw_ss = 1;
18121                         }
18122                 }
18123                 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18124                     (tp->t_flags & TF_SACK_PERMIT) &&
18125                     tp->rcv_numsacks > 0)
18126                         tcp_clean_dsack_blocks(tp);
18127                 tot_len_this_send += len;
18128                 if (len == 0)
18129                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18130                 else if (len == 1) {
18131                         counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18132                 } else if (len > 1) {
18133                         int idx;
18134
18135                         idx = (len / segsiz) + 3;
18136                         if (idx >= TCP_MSS_ACCT_ATIMER)
18137                                 counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18138                         else
18139                                 counter_u64_add(rack_out_size[idx], 1);
18140                 }
18141         }
18142         if ((rack->rack_no_prr == 0) &&
18143             sub_from_prr &&
18144             (error == 0)) {
18145                 if (rack->r_ctl.rc_prr_sndcnt >= len)
18146                         rack->r_ctl.rc_prr_sndcnt -= len;
18147                 else
18148                         rack->r_ctl.rc_prr_sndcnt = 0;
18149         }
18150         sub_from_prr = 0;
18151         if (doing_tlp && (rsm == NULL)) {
18152                 /* New send doing a TLP */
18153                 add_flag |= RACK_TLP;
18154                 tp->t_sndtlppack++;
18155                 tp->t_sndtlpbyte += len;
18156         }
18157         rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18158                         rack_to_usec_ts(&tv),
18159                         rsm, add_flag, s_mb, s_moff);
18160
18161
18162         if ((error == 0) &&
18163             (len > 0) &&
18164             (tp->snd_una == tp->snd_max))
18165                 rack->r_ctl.rc_tlp_rxt_last_time = cts;
18166         {
18167                 tcp_seq startseq = tp->snd_nxt;
18168
18169                 /* Track our lost count */
18170                 if (rsm && (doing_tlp == 0))
18171                         rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18172                 /*
18173                  * Advance snd_nxt over sequence space of this segment.
18174                  */
18175                 if (error)
18176                         /* We don't log or do anything with errors */
18177                         goto nomore;
18178                 if (doing_tlp == 0) {
18179                         if (rsm == NULL) {
18180                                 /*
18181                                  * Not a retransmission of some
18182                                  * sort, new data is going out so
18183                                  * clear our TLP count and flag.
18184                                  */
18185                                 rack->rc_tlp_in_progress = 0;
18186                                 rack->r_ctl.rc_tlp_cnt_out = 0;
18187                         }
18188                 } else {
18189                         /*
18190                          * We have just sent a TLP, mark that it is true
18191                          * and make sure our in progress is set so we
18192                          * continue to check the count.
18193                          */
18194                         rack->rc_tlp_in_progress = 1;
18195                         rack->r_ctl.rc_tlp_cnt_out++;
18196                 }
18197                 if (flags & (TH_SYN | TH_FIN)) {
18198                         if (flags & TH_SYN)
18199                                 tp->snd_nxt++;
18200                         if (flags & TH_FIN) {
18201                                 tp->snd_nxt++;
18202                                 tp->t_flags |= TF_SENTFIN;
18203                         }
18204                 }
18205                 /* In the ENOBUFS case we do *not* update snd_max */
18206                 if (sack_rxmit)
18207                         goto nomore;
18208
18209                 tp->snd_nxt += len;
18210                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18211                         if (tp->snd_una == tp->snd_max) {
18212                                 /*
18213                                  * Update the time we just added data since
18214                                  * none was outstanding.
18215                                  */
18216                                 rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18217                                 tp->t_acktime = ticks;
18218                         }
18219                         tp->snd_max = tp->snd_nxt;
18220                         /*
18221                          * Time this transmission if not a retransmission and
18222                          * not currently timing anything.
18223                          * This is only relevant in case of switching back to
18224                          * the base stack.
18225                          */
18226                         if (tp->t_rtttime == 0) {
18227                                 tp->t_rtttime = ticks;
18228                                 tp->t_rtseq = startseq;
18229                                 KMOD_TCPSTAT_INC(tcps_segstimed);
18230                         }
18231                         if (len &&
18232                             ((tp->t_flags & TF_GPUTINPROG) == 0))
18233                                 rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18234                 }
18235                 /*
18236                  * If we are doing FO we need to update the mbuf position and subtract
18237                  * this happens when the peer sends us duplicate information and
18238                  * we thus want to send a DSACK.
18239                  *
18240                  * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18241                  * turned off? If not then we are going to echo multiple DSACK blocks
18242                  * out (with the TSO), which we should not be doing.
18243                  */
18244                 if (rack->r_fast_output && len) {
18245                         if (rack->r_ctl.fsb.left_to_send > len)
18246                                 rack->r_ctl.fsb.left_to_send -= len;
18247                         else
18248                                 rack->r_ctl.fsb.left_to_send = 0;
18249                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18250                                 rack->r_fast_output = 0;
18251                         if (rack->r_fast_output) {
18252                                 rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18253                                 rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18254                         }
18255                 }
18256         }
18257 nomore:
18258         if (error) {
18259                 rack->r_ctl.rc_agg_delayed = 0;
18260                 rack->r_early = 0;
18261                 rack->r_late = 0;
18262                 rack->r_ctl.rc_agg_early = 0;
18263                 SOCKBUF_UNLOCK_ASSERT(sb);      /* Check gotos. */
18264                 /*
18265                  * Failures do not advance the seq counter above. For the
18266                  * case of ENOBUFS we will fall out and retry in 1ms with
18267                  * the hpts. Everything else will just have to retransmit
18268                  * with the timer.
18269                  *
18270                  * In any case, we do not want to loop around for another
18271                  * send without a good reason.
18272                  */
18273                 sendalot = 0;
18274                 switch (error) {
18275                 case EPERM:
18276                         tp->t_softerror = error;
18277 #ifdef TCP_ACCOUNTING
18278                         crtsc = get_cyclecount();
18279                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18280                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18281                         }
18282                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18283                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18284                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18285                         }
18286                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18287                         sched_unpin();
18288 #endif
18289                         return (error);
18290                 case ENOBUFS:
18291                         /*
18292                          * Pace us right away to retry in a some
18293                          * time
18294                          */
18295                         slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18296                         if (rack->rc_enobuf < 0x7f)
18297                                 rack->rc_enobuf++;
18298                         if (slot < (10 * HPTS_USEC_IN_MSEC))
18299                                 slot = 10 * HPTS_USEC_IN_MSEC;
18300                         if (rack->r_ctl.crte != NULL) {
18301                                 counter_u64_add(rack_saw_enobuf_hw, 1);
18302                                 tcp_rl_log_enobuf(rack->r_ctl.crte);
18303                         }
18304                         counter_u64_add(rack_saw_enobuf, 1);
18305                         goto enobufs;
18306                 case EMSGSIZE:
18307                         /*
18308                          * For some reason the interface we used initially
18309                          * to send segments changed to another or lowered
18310                          * its MTU. If TSO was active we either got an
18311                          * interface without TSO capabilits or TSO was
18312                          * turned off. If we obtained mtu from ip_output()
18313                          * then update it and try again.
18314                          */
18315                         if (tso)
18316                                 tp->t_flags &= ~TF_TSO;
18317                         if (mtu != 0) {
18318                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
18319                                 goto again;
18320                         }
18321                         slot = 10 * HPTS_USEC_IN_MSEC;
18322                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18323 #ifdef TCP_ACCOUNTING
18324                         crtsc = get_cyclecount();
18325                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18326                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18327                         }
18328                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18329                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18330                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18331                         }
18332                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18333                         sched_unpin();
18334 #endif
18335                         return (error);
18336                 case ENETUNREACH:
18337                         counter_u64_add(rack_saw_enetunreach, 1);
18338                 case EHOSTDOWN:
18339                 case EHOSTUNREACH:
18340                 case ENETDOWN:
18341                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
18342                                 tp->t_softerror = error;
18343                         }
18344                         /* FALLTHROUGH */
18345                 default:
18346                         slot = 10 * HPTS_USEC_IN_MSEC;
18347                         rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18348 #ifdef TCP_ACCOUNTING
18349                         crtsc = get_cyclecount();
18350                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18351                                 tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18352                         }
18353                         counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18354                         if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18355                                 tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18356                         }
18357                         counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18358                         sched_unpin();
18359 #endif
18360                         return (error);
18361                 }
18362         } else {
18363                 rack->rc_enobuf = 0;
18364                 if (IN_FASTRECOVERY(tp->t_flags) && rsm)
18365                         rack->r_ctl.retran_during_recovery += len;
18366         }
18367         KMOD_TCPSTAT_INC(tcps_sndtotal);
18368
18369         /*
18370          * Data sent (as far as we can tell). If this advertises a larger
18371          * window than any other segment, then remember the size of the
18372          * advertised window. Any pending ACK has now been sent.
18373          */
18374         if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
18375                 tp->rcv_adv = tp->rcv_nxt + recwin;
18376
18377         tp->last_ack_sent = tp->rcv_nxt;
18378         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
18379 enobufs:
18380         if (sendalot) {
18381                 /* Do we need to turn off sendalot? */
18382                 if (rack->r_ctl.rc_pace_max_segs &&
18383                     (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
18384                         /* We hit our max. */
18385                         sendalot = 0;
18386                 } else if ((rack->rc_user_set_max_segs) &&
18387                            (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
18388                         /* We hit the user defined max */
18389                         sendalot = 0;
18390                 }
18391         }
18392         if ((error == 0) && (flags & TH_FIN))
18393                 tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
18394         if (flags & TH_RST) {
18395                 /*
18396                  * We don't send again after sending a RST.
18397                  */
18398                 slot = 0;
18399                 sendalot = 0;
18400                 if (error == 0)
18401                         tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
18402         } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
18403                 /*
18404                  * Get our pacing rate, if an error
18405                  * occurred in sending (ENOBUF) we would
18406                  * hit the else if with slot preset. Other
18407                  * errors return.
18408                  */
18409                 slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
18410         }
18411         if (rsm &&
18412             (rsm->r_flags & RACK_HAS_SYN) == 0 &&
18413             rack->use_rack_rr) {
18414                 /* Its a retransmit and we use the rack cheat? */
18415                 if ((slot == 0) ||
18416                     (rack->rc_always_pace == 0) ||
18417                     (rack->r_rr_config == 1)) {
18418                         /*
18419                          * We have no pacing set or we
18420                          * are using old-style rack or
18421                          * we are overriden to use the old 1ms pacing.
18422                          */
18423                         slot = rack->r_ctl.rc_min_to;
18424                 }
18425         }
18426         /* We have sent clear the flag */
18427         rack->r_ent_rec_ns = 0;
18428         if (rack->r_must_retran) {
18429                 if (rsm) {
18430                         rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
18431                         if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
18432                                 /*
18433                                  * We have retransmitted all.
18434                                  */
18435                                 rack->r_must_retran = 0;
18436                                 rack->r_ctl.rc_out_at_rto = 0;
18437                         }
18438                 } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18439                         /*
18440                          * Sending new data will also kill
18441                          * the loop.
18442                          */
18443                         rack->r_must_retran = 0;
18444                         rack->r_ctl.rc_out_at_rto = 0;
18445                 }
18446         }
18447         rack->r_ctl.fsb.recwin = recwin;
18448         if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
18449             SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
18450                 /*
18451                  * We hit an RTO and now have past snd_max at the RTO
18452                  * clear all the WAS flags.
18453                  */
18454                 tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
18455         }
18456         if (slot) {
18457                 /* set the rack tcb into the slot N */
18458                 counter_u64_add(rack_paced_segments, 1);
18459                 if ((error == 0) &&
18460                     rack_use_rfo &&
18461                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18462                     (rsm == NULL) &&
18463                     (tp->snd_nxt == tp->snd_max) &&
18464                     (ipoptlen == 0) &&
18465                     (tp->rcv_numsacks == 0) &&
18466                     rack->r_fsb_inited &&
18467                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18468                     (rack->r_must_retran == 0) &&
18469                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18470                     (len > 0) && (orig_len > 0) &&
18471                     (orig_len > len) &&
18472                     ((orig_len - len) >= segsiz) &&
18473                     ((optlen == 0) ||
18474                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18475                         /* We can send at least one more MSS using our fsb */
18476
18477                         rack->r_fast_output = 1;
18478                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18479                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18480                         rack->r_ctl.fsb.tcp_flags = flags;
18481                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18482                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18483                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18484                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18485                                  (tp->snd_max - tp->snd_una)));
18486                         if (rack->r_ctl.fsb.left_to_send < segsiz)
18487                                 rack->r_fast_output = 0;
18488                         else {
18489                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18490                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18491                                 else
18492                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18493                         }
18494                 } else
18495                         rack->r_fast_output = 0;
18496                 rack_log_fsb(rack, tp, so, flags,
18497                              ipoptlen, orig_len, len, error,
18498                              (rsm == NULL), optlen, __LINE__, 2);
18499         } else if (sendalot) {
18500                 int ret;
18501
18502                 if (len)
18503                         counter_u64_add(rack_unpaced_segments, 1);
18504                 sack_rxmit = 0;
18505                 if ((error == 0) &&
18506                     rack_use_rfo &&
18507                     ((flags & (TH_SYN|TH_FIN)) == 0) &&
18508                     (rsm == NULL) &&
18509                     (ipoptlen == 0) &&
18510                     (tp->rcv_numsacks == 0) &&
18511                     (tp->snd_nxt == tp->snd_max) &&
18512                     (rack->r_must_retran == 0) &&
18513                     rack->r_fsb_inited &&
18514                     TCPS_HAVEESTABLISHED(tp->t_state) &&
18515                     ((tp->t_flags & TF_NEEDFIN) == 0) &&
18516                     (len > 0) && (orig_len > 0) &&
18517                     (orig_len > len) &&
18518                     ((orig_len - len) >= segsiz) &&
18519                     ((optlen == 0) ||
18520                      ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
18521                         /* we can use fast_output for more */
18522
18523                         rack->r_fast_output = 1;
18524                         rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18525                         rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18526                         rack->r_ctl.fsb.tcp_flags = flags;
18527                         rack->r_ctl.fsb.left_to_send = orig_len - len;
18528                         KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
18529                                 ("rack:%p left_to_send:%u sbavail:%u out:%u",
18530                                  rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
18531                                  (tp->snd_max - tp->snd_una)));
18532                         if (rack->r_ctl.fsb.left_to_send < segsiz) {
18533                                 rack->r_fast_output = 0;
18534                         }
18535                         if (rack->r_fast_output) {
18536                                 if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
18537                                         rack->r_ctl.fsb.rfo_apply_push = 1;
18538                                 else
18539                                         rack->r_ctl.fsb.rfo_apply_push = 0;
18540                                 rack_log_fsb(rack, tp, so, flags,
18541                                              ipoptlen, orig_len, len, error,
18542                                              (rsm == NULL), optlen, __LINE__, 3);
18543                                 error = 0;
18544                                 ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
18545                                 if (ret >= 0)
18546                                         return (ret);
18547                                 else if (error)
18548                                         goto nomore;
18549
18550                         }
18551                 }
18552                 goto again;
18553         } else if (len) {
18554                 counter_u64_add(rack_unpaced_segments, 1);
18555         }
18556         /* Assure when we leave that snd_nxt will point to top */
18557         if (SEQ_GT(tp->snd_max, tp->snd_nxt))
18558                 tp->snd_nxt = tp->snd_max;
18559         rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
18560 #ifdef TCP_ACCOUNTING
18561         crtsc = get_cyclecount() - ts_val;
18562         if (tot_len_this_send) {
18563                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18564                         tp->tcp_cnt_counters[SND_OUT_DATA]++;
18565                 }
18566                 counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
18567                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18568                         tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
18569                 }
18570                 counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
18571                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18572                         tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
18573                 }
18574                 counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
18575         } else {
18576                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18577                         tp->tcp_cnt_counters[SND_OUT_ACK]++;
18578                 }
18579                 counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
18580                 if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18581                         tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
18582                 }
18583                 counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
18584         }
18585         sched_unpin();
18586 #endif
18587         if (error == ENOBUFS)
18588                 error = 0;
18589         return (error);
18590 }
18591
18592 static void
18593 rack_update_seg(struct tcp_rack *rack)
18594 {
18595         uint32_t orig_val;
18596
18597         orig_val = rack->r_ctl.rc_pace_max_segs;
18598         rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
18599         if (orig_val != rack->r_ctl.rc_pace_max_segs)
18600                 rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL);
18601 }
18602
18603 static void
18604 rack_mtu_change(struct tcpcb *tp)
18605 {
18606         /*
18607          * The MSS may have changed
18608          */
18609         struct tcp_rack *rack;
18610
18611         rack = (struct tcp_rack *)tp->t_fb_ptr;
18612         if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
18613                 /*
18614                  * The MTU has changed we need to resend everything
18615                  * since all we have sent is lost. We first fix
18616                  * up the mtu though.
18617                  */
18618                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
18619                 /* We treat this like a full retransmit timeout without the cwnd adjustment */
18620                 rack_remxt_tmr(tp);
18621                 rack->r_fast_output = 0;
18622                 rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
18623                                                 rack->r_ctl.rc_sacked);
18624                 rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
18625                 rack->r_must_retran = 1;
18626
18627         }
18628         sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
18629         /* We don't use snd_nxt to retransmit */
18630         tp->snd_nxt = tp->snd_max;
18631 }
18632
18633 static int
18634 rack_set_profile(struct tcp_rack *rack, int prof)
18635 {
18636         int err = EINVAL;
18637         if (prof == 1) {
18638                 /* pace_always=1 */
18639                 if (rack->rc_always_pace == 0) {
18640                         if (tcp_can_enable_pacing() == 0)
18641                                 return (EBUSY);
18642                 }
18643                 rack->rc_always_pace = 1;
18644                 if (rack->use_fixed_rate || rack->gp_ready)
18645                         rack_set_cc_pacing(rack);
18646                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18647                 rack->rack_attempt_hdwr_pace = 0;
18648                 /* cmpack=1 */
18649                 if (rack_use_cmp_acks)
18650                         rack->r_use_cmp_ack = 1;
18651                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18652                     rack->r_use_cmp_ack)
18653                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18654                 /* scwnd=1 */
18655                 rack->rack_enable_scwnd = 1;
18656                 /* dynamic=100 */
18657                 rack->rc_gp_dyn_mul = 1;
18658                 /* gp_inc_ca */
18659                 rack->r_ctl.rack_per_of_gp_ca = 100;
18660                 /* rrr_conf=3 */
18661                 rack->r_rr_config = 3;
18662                 /* npush=2 */
18663                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18664                 /* fillcw=1 */
18665                 rack->rc_pace_to_cwnd = 1;
18666                 rack->rc_pace_fill_if_rttin_range = 0;
18667                 rack->rtt_limit_mul = 0;
18668                 /* noprr=1 */
18669                 rack->rack_no_prr = 1;
18670                 /* lscwnd=1 */
18671                 rack->r_limit_scw = 1;
18672                 /* gp_inc_rec */
18673                 rack->r_ctl.rack_per_of_gp_rec = 90;
18674                 err = 0;
18675
18676         } else if (prof == 3) {
18677                 /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
18678                 /* pace_always=1 */
18679                 if (rack->rc_always_pace == 0) {
18680                         if (tcp_can_enable_pacing() == 0)
18681                                 return (EBUSY);
18682                 }
18683                 rack->rc_always_pace = 1;
18684                 if (rack->use_fixed_rate || rack->gp_ready)
18685                         rack_set_cc_pacing(rack);
18686                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18687                 rack->rack_attempt_hdwr_pace = 0;
18688                 /* cmpack=1 */
18689                 if (rack_use_cmp_acks)
18690                         rack->r_use_cmp_ack = 1;
18691                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
18692                     rack->r_use_cmp_ack)
18693                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18694                 /* scwnd=1 */
18695                 rack->rack_enable_scwnd = 1;
18696                 /* dynamic=100 */
18697                 rack->rc_gp_dyn_mul = 1;
18698                 /* gp_inc_ca */
18699                 rack->r_ctl.rack_per_of_gp_ca = 100;
18700                 /* rrr_conf=3 */
18701                 rack->r_rr_config = 3;
18702                 /* npush=2 */
18703                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18704                 /* fillcw=2 */
18705                 rack->rc_pace_to_cwnd = 1;
18706                 rack->r_fill_less_agg = 1;
18707                 rack->rc_pace_fill_if_rttin_range = 0;
18708                 rack->rtt_limit_mul = 0;
18709                 /* noprr=1 */
18710                 rack->rack_no_prr = 1;
18711                 /* lscwnd=1 */
18712                 rack->r_limit_scw = 1;
18713                 /* gp_inc_rec */
18714                 rack->r_ctl.rack_per_of_gp_rec = 90;
18715                 err = 0;
18716
18717
18718         } else if (prof == 2) {
18719                 /* cmpack=1 */
18720                 if (rack->rc_always_pace == 0) {
18721                         if (tcp_can_enable_pacing() == 0)
18722                                 return (EBUSY);
18723                 }
18724                 rack->rc_always_pace = 1;
18725                 if (rack->use_fixed_rate || rack->gp_ready)
18726                         rack_set_cc_pacing(rack);
18727                 rack->r_use_cmp_ack = 1;
18728                 if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18729                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18730                 /* pace_always=1 */
18731                 rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18732                 /* scwnd=1 */
18733                 rack->rack_enable_scwnd = 1;
18734                 /* dynamic=100 */
18735                 rack->rc_gp_dyn_mul = 1;
18736                 rack->r_ctl.rack_per_of_gp_ca = 100;
18737                 /* rrr_conf=3 */
18738                 rack->r_rr_config = 3;
18739                 /* npush=2 */
18740                 rack->r_ctl.rc_no_push_at_mrtt = 2;
18741                 /* fillcw=1 */
18742                 rack->rc_pace_to_cwnd = 1;
18743                 rack->rc_pace_fill_if_rttin_range = 0;
18744                 rack->rtt_limit_mul = 0;
18745                 /* noprr=1 */
18746                 rack->rack_no_prr = 1;
18747                 /* lscwnd=0 */
18748                 rack->r_limit_scw = 0;
18749                 err = 0;
18750         } else if (prof == 0) {
18751                 /* This changes things back to the default settings */
18752                 err = 0;
18753                 if (rack->rc_always_pace) {
18754                         tcp_decrement_paced_conn();
18755                         rack_undo_cc_pacing(rack);
18756                         rack->rc_always_pace = 0;
18757                 }
18758                 if (rack_pace_every_seg && tcp_can_enable_pacing()) {
18759                         rack->rc_always_pace = 1;
18760                         if (rack->use_fixed_rate || rack->gp_ready)
18761                                 rack_set_cc_pacing(rack);
18762                 } else
18763                         rack->rc_always_pace = 0;
18764                 if (rack_use_cmp_acks)
18765                         rack->r_use_cmp_ack = 1;
18766                 else
18767                         rack->r_use_cmp_ack = 0;
18768                 if (rack_disable_prr)
18769                         rack->rack_no_prr = 1;
18770                 else
18771                         rack->rack_no_prr = 0;
18772                 if (rack_gp_no_rec_chg)
18773                         rack->rc_gp_no_rec_chg = 1;
18774                 else
18775                         rack->rc_gp_no_rec_chg = 0;
18776                 if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
18777                         rack->r_mbuf_queue = 1;
18778                         if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
18779                                 rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18780                         rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18781                 } else {
18782                         rack->r_mbuf_queue = 0;
18783                         rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
18784                 }
18785                 if (rack_enable_shared_cwnd)
18786                         rack->rack_enable_scwnd = 1;
18787                 else
18788                         rack->rack_enable_scwnd = 0;
18789                 if (rack_do_dyn_mul) {
18790                         /* When dynamic adjustment is on CA needs to start at 100% */
18791                         rack->rc_gp_dyn_mul = 1;
18792                         if (rack_do_dyn_mul >= 100)
18793                                 rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
18794                 } else {
18795                         rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
18796                         rack->rc_gp_dyn_mul = 0;
18797                 }
18798                 rack->r_rr_config = 0;
18799                 rack->r_ctl.rc_no_push_at_mrtt = 0;
18800                 rack->rc_pace_to_cwnd = 0;
18801                 rack->rc_pace_fill_if_rttin_range = 0;
18802                 rack->rtt_limit_mul = 0;
18803
18804                 if (rack_enable_hw_pacing)
18805                         rack->rack_hdw_pace_ena = 1;
18806                 else
18807                         rack->rack_hdw_pace_ena = 0;
18808                 if (rack_disable_prr)
18809                         rack->rack_no_prr = 1;
18810                 else
18811                         rack->rack_no_prr = 0;
18812                 if (rack_limits_scwnd)
18813                         rack->r_limit_scw  = 1;
18814                 else
18815                         rack->r_limit_scw  = 0;
18816                 err = 0;
18817         }
18818         return (err);
18819 }
18820
18821 static int
18822 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
18823 {
18824         struct deferred_opt_list *dol;
18825
18826         dol = malloc(sizeof(struct deferred_opt_list),
18827                      M_TCPFSB, M_NOWAIT|M_ZERO);
18828         if (dol == NULL) {
18829                 /*
18830                  * No space yikes -- fail out..
18831                  */
18832                 return (0);
18833         }
18834         dol->optname = sopt_name;
18835         dol->optval = loptval;
18836         TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
18837         return (1);
18838 }
18839
18840 static int
18841 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
18842                     uint32_t optval, uint64_t loptval)
18843 {
18844         struct epoch_tracker et;
18845         struct sockopt sopt;
18846         struct cc_newreno_opts opt;
18847         uint64_t val;
18848         int error = 0;
18849         uint16_t ca, ss;
18850
18851         switch (sopt_name) {
18852
18853         case TCP_RACK_PACING_BETA:
18854                 RACK_OPTS_INC(tcp_rack_beta);
18855                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18856                         /* This only works for newreno. */
18857                         error = EINVAL;
18858                         break;
18859                 }
18860                 if (rack->rc_pacing_cc_set) {
18861                         /*
18862                          * Set them into the real CC module
18863                          * whats in the rack pcb is the old values
18864                          * to be used on restoral/
18865                          */
18866                         sopt.sopt_dir = SOPT_SET;
18867                         opt.name = CC_NEWRENO_BETA;
18868                         opt.val = optval;
18869                         if (CC_ALGO(tp)->ctl_output != NULL)
18870                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18871                         else {
18872                                 error = ENOENT;
18873                                 break;
18874                         }
18875                 } else {
18876                         /*
18877                          * Not pacing yet so set it into our local
18878                          * rack pcb storage.
18879                          */
18880                         rack->r_ctl.rc_saved_beta.beta = optval;
18881                 }
18882                 break;
18883         case TCP_RACK_TIMER_SLOP:
18884                 RACK_OPTS_INC(tcp_rack_timer_slop);
18885                 rack->r_ctl.timer_slop = optval;
18886                 if (rack->rc_tp->t_srtt) {
18887                         /*
18888                          * If we have an SRTT lets update t_rxtcur
18889                          * to have the new slop.
18890                          */
18891                         RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
18892                                            rack_rto_min, rack_rto_max,
18893                                            rack->r_ctl.timer_slop);
18894                 }
18895                 break;
18896         case TCP_RACK_PACING_BETA_ECN:
18897                 RACK_OPTS_INC(tcp_rack_beta_ecn);
18898                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
18899                         /* This only works for newreno. */
18900                         error = EINVAL;
18901                         break;
18902                 }
18903                 if (rack->rc_pacing_cc_set) {
18904                         /*
18905                          * Set them into the real CC module
18906                          * whats in the rack pcb is the old values
18907                          * to be used on restoral/
18908                          */
18909                         sopt.sopt_dir = SOPT_SET;
18910                         opt.name = CC_NEWRENO_BETA_ECN;
18911                         opt.val = optval;
18912                         if (CC_ALGO(tp)->ctl_output != NULL)
18913                                 error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
18914                         else
18915                                 error = ENOENT;
18916                 } else {
18917                         /*
18918                          * Not pacing yet so set it into our local
18919                          * rack pcb storage.
18920                          */
18921                         rack->r_ctl.rc_saved_beta.beta_ecn = optval;
18922                         rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN;
18923                 }
18924                 break;
18925         case TCP_DEFER_OPTIONS:
18926                 RACK_OPTS_INC(tcp_defer_opt);
18927                 if (optval) {
18928                         if (rack->gp_ready) {
18929                                 /* Too late */
18930                                 error = EINVAL;
18931                                 break;
18932                         }
18933                         rack->defer_options = 1;
18934                 } else
18935                         rack->defer_options = 0;
18936                 break;
18937         case TCP_RACK_MEASURE_CNT:
18938                 RACK_OPTS_INC(tcp_rack_measure_cnt);
18939                 if (optval && (optval <= 0xff)) {
18940                         rack->r_ctl.req_measurements = optval;
18941                 } else
18942                         error = EINVAL;
18943                 break;
18944         case TCP_REC_ABC_VAL:
18945                 RACK_OPTS_INC(tcp_rec_abc_val);
18946                 if (optval > 0)
18947                         rack->r_use_labc_for_rec = 1;
18948                 else
18949                         rack->r_use_labc_for_rec = 0;
18950                 break;
18951         case TCP_RACK_ABC_VAL:
18952                 RACK_OPTS_INC(tcp_rack_abc_val);
18953                 if ((optval > 0) && (optval < 255))
18954                         rack->rc_labc = optval;
18955                 else
18956                         error = EINVAL;
18957                 break;
18958         case TCP_HDWR_UP_ONLY:
18959                 RACK_OPTS_INC(tcp_pacing_up_only);
18960                 if (optval)
18961                         rack->r_up_only = 1;
18962                 else
18963                         rack->r_up_only = 0;
18964                 break;
18965         case TCP_PACING_RATE_CAP:
18966                 RACK_OPTS_INC(tcp_pacing_rate_cap);
18967                 rack->r_ctl.bw_rate_cap = loptval;
18968                 break;
18969         case TCP_RACK_PROFILE:
18970                 RACK_OPTS_INC(tcp_profile);
18971                 error = rack_set_profile(rack, optval);
18972                 break;
18973         case TCP_USE_CMP_ACKS:
18974                 RACK_OPTS_INC(tcp_use_cmp_acks);
18975                 if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
18976                         /* You can't turn it off once its on! */
18977                         error = EINVAL;
18978                 } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
18979                         rack->r_use_cmp_ack = 1;
18980                         rack->r_mbuf_queue = 1;
18981                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
18982                 }
18983                 if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
18984                         rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
18985                 break;
18986         case TCP_SHARED_CWND_TIME_LIMIT:
18987                 RACK_OPTS_INC(tcp_lscwnd);
18988                 if (optval)
18989                         rack->r_limit_scw = 1;
18990                 else
18991                         rack->r_limit_scw = 0;
18992                 break;
18993         case TCP_RACK_PACE_TO_FILL:
18994                 RACK_OPTS_INC(tcp_fillcw);
18995                 if (optval == 0)
18996                         rack->rc_pace_to_cwnd = 0;
18997                 else {
18998                         rack->rc_pace_to_cwnd = 1;
18999                         if (optval > 1)
19000                                 rack->r_fill_less_agg = 1;
19001                 }
19002                 if ((optval >= rack_gp_rtt_maxmul) &&
19003                     rack_gp_rtt_maxmul &&
19004                     (optval < 0xf)) {
19005                         rack->rc_pace_fill_if_rttin_range = 1;
19006                         rack->rtt_limit_mul = optval;
19007                 } else {
19008                         rack->rc_pace_fill_if_rttin_range = 0;
19009                         rack->rtt_limit_mul = 0;
19010                 }
19011                 break;
19012         case TCP_RACK_NO_PUSH_AT_MAX:
19013                 RACK_OPTS_INC(tcp_npush);
19014                 if (optval == 0)
19015                         rack->r_ctl.rc_no_push_at_mrtt = 0;
19016                 else if (optval < 0xff)
19017                         rack->r_ctl.rc_no_push_at_mrtt = optval;
19018                 else
19019                         error = EINVAL;
19020                 break;
19021         case TCP_SHARED_CWND_ENABLE:
19022                 RACK_OPTS_INC(tcp_rack_scwnd);
19023                 if (optval == 0)
19024                         rack->rack_enable_scwnd = 0;
19025                 else
19026                         rack->rack_enable_scwnd = 1;
19027                 break;
19028         case TCP_RACK_MBUF_QUEUE:
19029                 /* Now do we use the LRO mbuf-queue feature */
19030                 RACK_OPTS_INC(tcp_rack_mbufq);
19031                 if (optval || rack->r_use_cmp_ack)
19032                         rack->r_mbuf_queue = 1;
19033                 else
19034                         rack->r_mbuf_queue = 0;
19035                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19036                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19037                 else
19038                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19039                 break;
19040         case TCP_RACK_NONRXT_CFG_RATE:
19041                 RACK_OPTS_INC(tcp_rack_cfg_rate);
19042                 if (optval == 0)
19043                         rack->rack_rec_nonrxt_use_cr = 0;
19044                 else
19045                         rack->rack_rec_nonrxt_use_cr = 1;
19046                 break;
19047         case TCP_NO_PRR:
19048                 RACK_OPTS_INC(tcp_rack_noprr);
19049                 if (optval == 0)
19050                         rack->rack_no_prr = 0;
19051                 else if (optval == 1)
19052                         rack->rack_no_prr = 1;
19053                 else if (optval == 2)
19054                         rack->no_prr_addback = 1;
19055                 else
19056                         error = EINVAL;
19057                 break;
19058         case TCP_TIMELY_DYN_ADJ:
19059                 RACK_OPTS_INC(tcp_timely_dyn);
19060                 if (optval == 0)
19061                         rack->rc_gp_dyn_mul = 0;
19062                 else {
19063                         rack->rc_gp_dyn_mul = 1;
19064                         if (optval >= 100) {
19065                                 /*
19066                                  * If the user sets something 100 or more
19067                                  * its the gp_ca value.
19068                                  */
19069                                 rack->r_ctl.rack_per_of_gp_ca  = optval;
19070                         }
19071                 }
19072                 break;
19073         case TCP_RACK_DO_DETECTION:
19074                 RACK_OPTS_INC(tcp_rack_do_detection);
19075                 if (optval == 0)
19076                         rack->do_detection = 0;
19077                 else
19078                         rack->do_detection = 1;
19079                 break;
19080         case TCP_RACK_TLP_USE:
19081                 if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19082                         error = EINVAL;
19083                         break;
19084                 }
19085                 RACK_OPTS_INC(tcp_tlp_use);
19086                 rack->rack_tlp_threshold_use = optval;
19087                 break;
19088         case TCP_RACK_TLP_REDUCE:
19089                 /* RACK TLP cwnd reduction (bool) */
19090                 RACK_OPTS_INC(tcp_rack_tlp_reduce);
19091                 rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19092                 break;
19093         /*  Pacing related ones */
19094         case TCP_RACK_PACE_ALWAYS:
19095                 /*
19096                  * zero is old rack method, 1 is new
19097                  * method using a pacing rate.
19098                  */
19099                 RACK_OPTS_INC(tcp_rack_pace_always);
19100                 if (optval > 0) {
19101                         if (rack->rc_always_pace) {
19102                                 error = EALREADY;
19103                                 break;
19104                         } else if (tcp_can_enable_pacing()) {
19105                                 rack->rc_always_pace = 1;
19106                                 if (rack->use_fixed_rate || rack->gp_ready)
19107                                         rack_set_cc_pacing(rack);
19108                         }
19109                         else {
19110                                 error = ENOSPC;
19111                                 break;
19112                         }
19113                 } else {
19114                         if (rack->rc_always_pace) {
19115                                 tcp_decrement_paced_conn();
19116                                 rack->rc_always_pace = 0;
19117                                 rack_undo_cc_pacing(rack);
19118                         }
19119                 }
19120                 if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19121                         tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19122                 else
19123                         tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19124                 /* A rate may be set irate or other, if so set seg size */
19125                 rack_update_seg(rack);
19126                 break;
19127         case TCP_BBR_RACK_INIT_RATE:
19128                 RACK_OPTS_INC(tcp_initial_rate);
19129                 val = optval;
19130                 /* Change from kbits per second to bytes per second */
19131                 val *= 1000;
19132                 val /= 8;
19133                 rack->r_ctl.init_rate = val;
19134                 if (rack->rc_init_win != rack_default_init_window) {
19135                         uint32_t win, snt;
19136
19137                         /*
19138                          * Options don't always get applied
19139                          * in the order you think. So in order
19140                          * to assure we update a cwnd we need
19141                          * to check and see if we are still
19142                          * where we should raise the cwnd.
19143                          */
19144                         win = rc_init_window(rack);
19145                         if (SEQ_GT(tp->snd_max, tp->iss))
19146                                 snt = tp->snd_max - tp->iss;
19147                         else
19148                                 snt = 0;
19149                         if ((snt < win) &&
19150                             (tp->snd_cwnd < win))
19151                                 tp->snd_cwnd = win;
19152                 }
19153                 if (rack->rc_always_pace)
19154                         rack_update_seg(rack);
19155                 break;
19156         case TCP_BBR_IWINTSO:
19157                 RACK_OPTS_INC(tcp_initial_win);
19158                 if (optval && (optval <= 0xff)) {
19159                         uint32_t win, snt;
19160
19161                         rack->rc_init_win = optval;
19162                         win = rc_init_window(rack);
19163                         if (SEQ_GT(tp->snd_max, tp->iss))
19164                                 snt = tp->snd_max - tp->iss;
19165                         else
19166                                 snt = 0;
19167                         if ((snt < win) &&
19168                             (tp->t_srtt |
19169 #ifdef NETFLIX_PEAKRATE
19170                              tp->t_maxpeakrate |
19171 #endif
19172                              rack->r_ctl.init_rate)) {
19173                                 /*
19174                                  * We are not past the initial window
19175                                  * and we have some bases for pacing,
19176                                  * so we need to possibly adjust up
19177                                  * the cwnd. Note even if we don't set
19178                                  * the cwnd, its still ok to raise the rc_init_win
19179                                  * which can be used coming out of idle when we
19180                                  * would have a rate.
19181                                  */
19182                                 if (tp->snd_cwnd < win)
19183                                         tp->snd_cwnd = win;
19184                         }
19185                         if (rack->rc_always_pace)
19186                                 rack_update_seg(rack);
19187                 } else
19188                         error = EINVAL;
19189                 break;
19190         case TCP_RACK_FORCE_MSEG:
19191                 RACK_OPTS_INC(tcp_rack_force_max_seg);
19192                 if (optval)
19193                         rack->rc_force_max_seg = 1;
19194                 else
19195                         rack->rc_force_max_seg = 0;
19196                 break;
19197         case TCP_RACK_PACE_MAX_SEG:
19198                 /* Max segments size in a pace in bytes */
19199                 RACK_OPTS_INC(tcp_rack_max_seg);
19200                 rack->rc_user_set_max_segs = optval;
19201                 rack_set_pace_segments(tp, rack, __LINE__, NULL);
19202                 break;
19203         case TCP_RACK_PACE_RATE_REC:
19204                 /* Set the fixed pacing rate in Bytes per second ca */
19205                 RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19206                 rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19207                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19208                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19209                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19210                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19211                 rack->use_fixed_rate = 1;
19212                 if (rack->rc_always_pace)
19213                         rack_set_cc_pacing(rack);
19214                 rack_log_pacing_delay_calc(rack,
19215                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19216                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19217                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19218                                            __LINE__, NULL);
19219                 break;
19220
19221         case TCP_RACK_PACE_RATE_SS:
19222                 /* Set the fixed pacing rate in Bytes per second ca */
19223                 RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19224                 rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19225                 if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19226                         rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19227                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19228                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19229                 rack->use_fixed_rate = 1;
19230                 if (rack->rc_always_pace)
19231                         rack_set_cc_pacing(rack);
19232                 rack_log_pacing_delay_calc(rack,
19233                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19234                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19235                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19236                                            __LINE__, NULL);
19237                 break;
19238
19239         case TCP_RACK_PACE_RATE_CA:
19240                 /* Set the fixed pacing rate in Bytes per second ca */
19241                 RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19242                 rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19243                 if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19244                         rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19245                 if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19246                         rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19247                 rack->use_fixed_rate = 1;
19248                 if (rack->rc_always_pace)
19249                         rack_set_cc_pacing(rack);
19250                 rack_log_pacing_delay_calc(rack,
19251                                            rack->r_ctl.rc_fixed_pacing_rate_ss,
19252                                            rack->r_ctl.rc_fixed_pacing_rate_ca,
19253                                            rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19254                                            __LINE__, NULL);
19255                 break;
19256         case TCP_RACK_GP_INCREASE_REC:
19257                 RACK_OPTS_INC(tcp_gp_inc_rec);
19258                 rack->r_ctl.rack_per_of_gp_rec = optval;
19259                 rack_log_pacing_delay_calc(rack,
19260                                            rack->r_ctl.rack_per_of_gp_ss,
19261                                            rack->r_ctl.rack_per_of_gp_ca,
19262                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19263                                            __LINE__, NULL);
19264                 break;
19265         case TCP_RACK_GP_INCREASE_CA:
19266                 RACK_OPTS_INC(tcp_gp_inc_ca);
19267                 ca = optval;
19268                 if (ca < 100) {
19269                         /*
19270                          * We don't allow any reduction
19271                          * over the GP b/w.
19272                          */
19273                         error = EINVAL;
19274                         break;
19275                 }
19276                 rack->r_ctl.rack_per_of_gp_ca = ca;
19277                 rack_log_pacing_delay_calc(rack,
19278                                            rack->r_ctl.rack_per_of_gp_ss,
19279                                            rack->r_ctl.rack_per_of_gp_ca,
19280                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19281                                            __LINE__, NULL);
19282                 break;
19283         case TCP_RACK_GP_INCREASE_SS:
19284                 RACK_OPTS_INC(tcp_gp_inc_ss);
19285                 ss = optval;
19286                 if (ss < 100) {
19287                         /*
19288                          * We don't allow any reduction
19289                          * over the GP b/w.
19290                          */
19291                         error = EINVAL;
19292                         break;
19293                 }
19294                 rack->r_ctl.rack_per_of_gp_ss = ss;
19295                 rack_log_pacing_delay_calc(rack,
19296                                            rack->r_ctl.rack_per_of_gp_ss,
19297                                            rack->r_ctl.rack_per_of_gp_ca,
19298                                            rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19299                                            __LINE__, NULL);
19300                 break;
19301         case TCP_RACK_RR_CONF:
19302                 RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
19303                 if (optval && optval <= 3)
19304                         rack->r_rr_config = optval;
19305                 else
19306                         rack->r_rr_config = 0;
19307                 break;
19308         case TCP_HDWR_RATE_CAP:
19309                 RACK_OPTS_INC(tcp_hdwr_rate_cap);
19310                 if (optval) {
19311                         if (rack->r_rack_hw_rate_caps == 0)
19312                                 rack->r_rack_hw_rate_caps = 1;
19313                         else
19314                                 error = EALREADY;
19315                 } else {
19316                         rack->r_rack_hw_rate_caps = 0;
19317                 }
19318                 break;
19319         case TCP_BBR_HDWR_PACE:
19320                 RACK_OPTS_INC(tcp_hdwr_pacing);
19321                 if (optval){
19322                         if (rack->rack_hdrw_pacing == 0) {
19323                                 rack->rack_hdw_pace_ena = 1;
19324                                 rack->rack_attempt_hdwr_pace = 0;
19325                         } else
19326                                 error = EALREADY;
19327                 } else {
19328                         rack->rack_hdw_pace_ena = 0;
19329 #ifdef RATELIMIT
19330                         if (rack->r_ctl.crte != NULL) {
19331                                 rack->rack_hdrw_pacing = 0;
19332                                 rack->rack_attempt_hdwr_pace = 0;
19333                                 tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
19334                                 rack->r_ctl.crte = NULL;
19335                         }
19336 #endif
19337                 }
19338                 break;
19339         /*  End Pacing related ones */
19340         case TCP_RACK_PRR_SENDALOT:
19341                 /* Allow PRR to send more than one seg */
19342                 RACK_OPTS_INC(tcp_rack_prr_sendalot);
19343                 rack->r_ctl.rc_prr_sendalot = optval;
19344                 break;
19345         case TCP_RACK_MIN_TO:
19346                 /* Minimum time between rack t-o's in ms */
19347                 RACK_OPTS_INC(tcp_rack_min_to);
19348                 rack->r_ctl.rc_min_to = optval;
19349                 break;
19350         case TCP_RACK_EARLY_SEG:
19351                 /* If early recovery max segments */
19352                 RACK_OPTS_INC(tcp_rack_early_seg);
19353                 rack->r_ctl.rc_early_recovery_segs = optval;
19354                 break;
19355         case TCP_RACK_REORD_THRESH:
19356                 /* RACK reorder threshold (shift amount) */
19357                 RACK_OPTS_INC(tcp_rack_reord_thresh);
19358                 if ((optval > 0) && (optval < 31))
19359                         rack->r_ctl.rc_reorder_shift = optval;
19360                 else
19361                         error = EINVAL;
19362                 break;
19363         case TCP_RACK_REORD_FADE:
19364                 /* Does reordering fade after ms time */
19365                 RACK_OPTS_INC(tcp_rack_reord_fade);
19366                 rack->r_ctl.rc_reorder_fade = optval;
19367                 break;
19368         case TCP_RACK_TLP_THRESH:
19369                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19370                 RACK_OPTS_INC(tcp_rack_tlp_thresh);
19371                 if (optval)
19372                         rack->r_ctl.rc_tlp_threshold = optval;
19373                 else
19374                         error = EINVAL;
19375                 break;
19376         case TCP_BBR_USE_RACK_RR:
19377                 RACK_OPTS_INC(tcp_rack_rr);
19378                 if (optval)
19379                         rack->use_rack_rr = 1;
19380                 else
19381                         rack->use_rack_rr = 0;
19382                 break;
19383         case TCP_FAST_RSM_HACK:
19384                 RACK_OPTS_INC(tcp_rack_fastrsm_hack);
19385                 if (optval)
19386                         rack->fast_rsm_hack = 1;
19387                 else
19388                         rack->fast_rsm_hack = 0;
19389                 break;
19390         case TCP_RACK_PKT_DELAY:
19391                 /* RACK added ms i.e. rack-rtt + reord + N */
19392                 RACK_OPTS_INC(tcp_rack_pkt_delay);
19393                 rack->r_ctl.rc_pkt_delay = optval;
19394                 break;
19395         case TCP_DELACK:
19396                 RACK_OPTS_INC(tcp_rack_delayed_ack);
19397                 if (optval == 0)
19398                         tp->t_delayed_ack = 0;
19399                 else
19400                         tp->t_delayed_ack = 1;
19401                 if (tp->t_flags & TF_DELACK) {
19402                         tp->t_flags &= ~TF_DELACK;
19403                         tp->t_flags |= TF_ACKNOW;
19404                         NET_EPOCH_ENTER(et);
19405                         rack_output(tp);
19406                         NET_EPOCH_EXIT(et);
19407                 }
19408                 break;
19409
19410         case TCP_BBR_RACK_RTT_USE:
19411                 RACK_OPTS_INC(tcp_rack_rtt_use);
19412                 if ((optval != USE_RTT_HIGH) &&
19413                     (optval != USE_RTT_LOW) &&
19414                     (optval != USE_RTT_AVG))
19415                         error = EINVAL;
19416                 else
19417                         rack->r_ctl.rc_rate_sample_method = optval;
19418                 break;
19419         case TCP_DATA_AFTER_CLOSE:
19420                 RACK_OPTS_INC(tcp_data_after_close);
19421                 if (optval)
19422                         rack->rc_allow_data_af_clo = 1;
19423                 else
19424                         rack->rc_allow_data_af_clo = 0;
19425                 break;
19426         default:
19427                 break;
19428         }
19429 #ifdef NETFLIX_STATS
19430         tcp_log_socket_option(tp, sopt_name, optval, error);
19431 #endif
19432         return (error);
19433 }
19434
19435
19436 static void
19437 rack_apply_deferred_options(struct tcp_rack *rack)
19438 {
19439         struct deferred_opt_list *dol, *sdol;
19440         uint32_t s_optval;
19441
19442         TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
19443                 TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
19444                 /* Disadvantage of deferal is you loose the error return */
19445                 s_optval = (uint32_t)dol->optval;
19446                 (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
19447                 free(dol, M_TCPDO);
19448         }
19449 }
19450
19451 static int
19452 rack_pru_options(struct tcpcb *tp, int flags)
19453 {
19454         if (flags & PRUS_OOB)
19455                 return (EOPNOTSUPP);
19456         return (0);
19457 }
19458
19459 static struct tcp_function_block __tcp_rack = {
19460         .tfb_tcp_block_name = __XSTRING(STACKNAME),
19461         .tfb_tcp_output = rack_output,
19462         .tfb_do_queued_segments = ctf_do_queued_segments,
19463         .tfb_do_segment_nounlock = rack_do_segment_nounlock,
19464         .tfb_tcp_do_segment = rack_do_segment,
19465         .tfb_tcp_ctloutput = rack_ctloutput,
19466         .tfb_tcp_fb_init = rack_init,
19467         .tfb_tcp_fb_fini = rack_fini,
19468         .tfb_tcp_timer_stop_all = rack_stopall,
19469         .tfb_tcp_timer_activate = rack_timer_activate,
19470         .tfb_tcp_timer_active = rack_timer_active,
19471         .tfb_tcp_timer_stop = rack_timer_stop,
19472         .tfb_tcp_rexmit_tmr = rack_remxt_tmr,
19473         .tfb_tcp_handoff_ok = rack_handoff_ok,
19474         .tfb_tcp_mtu_chg = rack_mtu_change,
19475         .tfb_pru_options = rack_pru_options,
19476
19477 };
19478
19479 /*
19480  * rack_ctloutput() must drop the inpcb lock before performing copyin on
19481  * socket option arguments.  When it re-acquires the lock after the copy, it
19482  * has to revalidate that the connection is still valid for the socket
19483  * option.
19484  */
19485 static int
19486 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
19487     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19488 {
19489         uint64_t loptval;
19490         int32_t error = 0, optval;
19491
19492         switch (sopt->sopt_name) {
19493         case TCP_RACK_TLP_REDUCE:               /*  URL:tlp_reduce */
19494         /*  Pacing related ones */
19495         case TCP_RACK_PACE_ALWAYS:              /*  URL:pace_always */
19496         case TCP_BBR_RACK_INIT_RATE:            /*  URL:irate */
19497         case TCP_BBR_IWINTSO:                   /*  URL:tso_iwin */
19498         case TCP_RACK_PACE_MAX_SEG:             /*  URL:pace_max_seg */
19499         case TCP_RACK_FORCE_MSEG:               /*  URL:force_max_seg */
19500         case TCP_RACK_PACE_RATE_CA:             /*  URL:pr_ca */
19501         case TCP_RACK_PACE_RATE_SS:             /*  URL:pr_ss*/
19502         case TCP_RACK_PACE_RATE_REC:            /*  URL:pr_rec */
19503         case TCP_RACK_GP_INCREASE_CA:           /*  URL:gp_inc_ca */
19504         case TCP_RACK_GP_INCREASE_SS:           /*  URL:gp_inc_ss */
19505         case TCP_RACK_GP_INCREASE_REC:          /*  URL:gp_inc_rec */
19506         case TCP_RACK_RR_CONF:                  /*  URL:rrr_conf */
19507         case TCP_BBR_HDWR_PACE:                 /*  URL:hdwrpace */
19508         case TCP_HDWR_RATE_CAP:                 /*  URL: hdwrcap boolean */
19509         case TCP_PACING_RATE_CAP:               /*  URL:cap-- used by side-channel */
19510         case TCP_HDWR_UP_ONLY:                  /*  URL:uponly -- hardware pacing  boolean */
19511        /* End pacing related */
19512         case TCP_FAST_RSM_HACK:                 /*  URL:frsm_hack */
19513         case TCP_DELACK:                        /*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
19514         case TCP_RACK_PRR_SENDALOT:             /*  URL:prr_sendalot */
19515         case TCP_RACK_MIN_TO:                   /*  URL:min_to */
19516         case TCP_RACK_EARLY_SEG:                /*  URL:early_seg */
19517         case TCP_RACK_REORD_THRESH:             /*  URL:reord_thresh */
19518         case TCP_RACK_REORD_FADE:               /*  URL:reord_fade */
19519         case TCP_RACK_TLP_THRESH:               /*  URL:tlp_thresh */
19520         case TCP_RACK_PKT_DELAY:                /*  URL:pkt_delay */
19521         case TCP_RACK_TLP_USE:                  /*  URL:tlp_use */
19522         case TCP_BBR_RACK_RTT_USE:              /*  URL:rttuse */
19523         case TCP_BBR_USE_RACK_RR:               /*  URL:rackrr */
19524         case TCP_RACK_DO_DETECTION:             /*  URL:detect */
19525         case TCP_NO_PRR:                        /*  URL:noprr */
19526         case TCP_TIMELY_DYN_ADJ:                /*  URL:dynamic */
19527         case TCP_DATA_AFTER_CLOSE:              /*  no URL */
19528         case TCP_RACK_NONRXT_CFG_RATE:          /*  URL:nonrxtcr */
19529         case TCP_SHARED_CWND_ENABLE:            /*  URL:scwnd */
19530         case TCP_RACK_MBUF_QUEUE:               /*  URL:mqueue */
19531         case TCP_RACK_NO_PUSH_AT_MAX:           /*  URL:npush */
19532         case TCP_RACK_PACE_TO_FILL:             /*  URL:fillcw */
19533         case TCP_SHARED_CWND_TIME_LIMIT:        /*  URL:lscwnd */
19534         case TCP_RACK_PROFILE:                  /*  URL:profile */
19535         case TCP_USE_CMP_ACKS:                  /*  URL:cmpack */
19536         case TCP_RACK_ABC_VAL:                  /*  URL:labc */
19537         case TCP_REC_ABC_VAL:                   /*  URL:reclabc */
19538         case TCP_RACK_MEASURE_CNT:              /*  URL:measurecnt */
19539         case TCP_DEFER_OPTIONS:                 /*  URL:defer */
19540         case TCP_RACK_PACING_BETA:              /*  URL:pacing_beta */
19541         case TCP_RACK_PACING_BETA_ECN:          /*  URL:pacing_beta_ecn */
19542         case TCP_RACK_TIMER_SLOP:               /*  URL:timer_slop */
19543                 break;
19544         default:
19545                 /* Filter off all unknown options to the base stack */
19546                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19547                 break;
19548         }
19549         INP_WUNLOCK(inp);
19550         if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
19551                 error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
19552                 /*
19553                  * We truncate it down to 32 bits for the socket-option trace this
19554                  * means rates > 34Gbps won't show right, but thats probably ok.
19555                  */
19556                 optval = (uint32_t)loptval;
19557         } else {
19558                 error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
19559                 /* Save it in 64 bit form too */
19560                 loptval = optval;
19561         }
19562         if (error)
19563                 return (error);
19564         INP_WLOCK(inp);
19565         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
19566                 INP_WUNLOCK(inp);
19567                 return (ECONNRESET);
19568         }
19569         if (tp->t_fb != &__tcp_rack) {
19570                 INP_WUNLOCK(inp);
19571                 return (ENOPROTOOPT);
19572         }
19573         if (rack->defer_options && (rack->gp_ready == 0) &&
19574             (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
19575             (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
19576             (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
19577             (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
19578                 /* Options are beind deferred */
19579                 if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
19580                         INP_WUNLOCK(inp);
19581                         return (0);
19582                 } else {
19583                         /* No memory to defer, fail */
19584                         INP_WUNLOCK(inp);
19585                         return (ENOMEM);
19586                 }
19587         }
19588         error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
19589         INP_WUNLOCK(inp);
19590         return (error);
19591 }
19592
19593 static void
19594 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
19595 {
19596
19597         INP_WLOCK_ASSERT(tp->t_inpcb);
19598         bzero(ti, sizeof(*ti));
19599
19600         ti->tcpi_state = tp->t_state;
19601         if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
19602                 ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
19603         if (tp->t_flags & TF_SACK_PERMIT)
19604                 ti->tcpi_options |= TCPI_OPT_SACK;
19605         if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
19606                 ti->tcpi_options |= TCPI_OPT_WSCALE;
19607                 ti->tcpi_snd_wscale = tp->snd_scale;
19608                 ti->tcpi_rcv_wscale = tp->rcv_scale;
19609         }
19610         if (tp->t_flags2 & TF2_ECN_PERMIT)
19611                 ti->tcpi_options |= TCPI_OPT_ECN;
19612         if (tp->t_flags & TF_FASTOPEN)
19613                 ti->tcpi_options |= TCPI_OPT_TFO;
19614         /* still kept in ticks is t_rcvtime */
19615         ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
19616         /* Since we hold everything in precise useconds this is easy */
19617         ti->tcpi_rtt = tp->t_srtt;
19618         ti->tcpi_rttvar = tp->t_rttvar;
19619         ti->tcpi_rto = tp->t_rxtcur;
19620         ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
19621         ti->tcpi_snd_cwnd = tp->snd_cwnd;
19622         /*
19623          * FreeBSD-specific extension fields for tcp_info.
19624          */
19625         ti->tcpi_rcv_space = tp->rcv_wnd;
19626         ti->tcpi_rcv_nxt = tp->rcv_nxt;
19627         ti->tcpi_snd_wnd = tp->snd_wnd;
19628         ti->tcpi_snd_bwnd = 0;          /* Unused, kept for compat. */
19629         ti->tcpi_snd_nxt = tp->snd_nxt;
19630         ti->tcpi_snd_mss = tp->t_maxseg;
19631         ti->tcpi_rcv_mss = tp->t_maxseg;
19632         ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
19633         ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
19634         ti->tcpi_snd_zerowin = tp->t_sndzerowin;
19635 #ifdef NETFLIX_STATS
19636         ti->tcpi_total_tlp = tp->t_sndtlppack;
19637         ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
19638         memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
19639 #endif
19640 #ifdef TCP_OFFLOAD
19641         if (tp->t_flags & TF_TOE) {
19642                 ti->tcpi_options |= TCPI_OPT_TOE;
19643                 tcp_offload_tcp_info(tp, ti);
19644         }
19645 #endif
19646 }
19647
19648 static int
19649 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
19650     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
19651 {
19652         int32_t error, optval;
19653         uint64_t val, loptval;
19654         struct  tcp_info ti;
19655         /*
19656          * Because all our options are either boolean or an int, we can just
19657          * pull everything into optval and then unlock and copy. If we ever
19658          * add a option that is not a int, then this will have quite an
19659          * impact to this routine.
19660          */
19661         error = 0;
19662         switch (sopt->sopt_name) {
19663         case TCP_INFO:
19664                 /* First get the info filled */
19665                 rack_fill_info(tp, &ti);
19666                 /* Fix up the rtt related fields if needed */
19667                 INP_WUNLOCK(inp);
19668                 error = sooptcopyout(sopt, &ti, sizeof ti);
19669                 return (error);
19670         /*
19671          * Beta is the congestion control value for NewReno that influences how
19672          * much of a backoff happens when loss is detected. It is normally set
19673          * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
19674          * when you exit recovery.
19675          */
19676         case TCP_RACK_PACING_BETA:
19677                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19678                         error = EINVAL;
19679                 else if (rack->rc_pacing_cc_set == 0)
19680                         optval = rack->r_ctl.rc_saved_beta.beta;
19681                 else {
19682                         /*
19683                          * Reach out into the CC data and report back what
19684                          * I have previously set. Yeah it looks hackish but
19685                          * we don't want to report the saved values.
19686                          */
19687                         if (tp->ccv->cc_data)
19688                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta;
19689                         else
19690                                 error = EINVAL;
19691                 }
19692                 break;
19693                 /*
19694                  * Beta_ecn is the congestion control value for NewReno that influences how
19695                  * much of a backoff happens when a ECN mark is detected. It is normally set
19696                  * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
19697                  * you exit recovery. Note that classic ECN has a beta of 50, it is only
19698                  * ABE Ecn that uses this "less" value, but we do too with pacing :)
19699                  */
19700
19701         case TCP_RACK_PACING_BETA_ECN:
19702                 if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
19703                         error = EINVAL;
19704                 else if (rack->rc_pacing_cc_set == 0)
19705                         optval = rack->r_ctl.rc_saved_beta.beta_ecn;
19706                 else {
19707                         /*
19708                          * Reach out into the CC data and report back what
19709                          * I have previously set. Yeah it looks hackish but
19710                          * we don't want to report the saved values.
19711                          */
19712                         if (tp->ccv->cc_data)
19713                                 optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
19714                         else
19715                                 error = EINVAL;
19716                 }
19717                 break;
19718         case TCP_FAST_RSM_HACK:
19719                 optval = rack->fast_rsm_hack;
19720                 break;
19721         case TCP_DEFER_OPTIONS:
19722                 optval = rack->defer_options;
19723                 break;
19724         case TCP_RACK_MEASURE_CNT:
19725                 optval = rack->r_ctl.req_measurements;
19726                 break;
19727         case TCP_REC_ABC_VAL:
19728                 optval = rack->r_use_labc_for_rec;
19729                 break;
19730         case TCP_RACK_ABC_VAL:
19731                 optval = rack->rc_labc;
19732                 break;
19733         case TCP_HDWR_UP_ONLY:
19734                 optval= rack->r_up_only;
19735                 break;
19736         case TCP_PACING_RATE_CAP:
19737                 loptval = rack->r_ctl.bw_rate_cap;
19738                 break;
19739         case TCP_RACK_PROFILE:
19740                 /* You cannot retrieve a profile, its write only */
19741                 error = EINVAL;
19742                 break;
19743         case TCP_USE_CMP_ACKS:
19744                 optval = rack->r_use_cmp_ack;
19745                 break;
19746         case TCP_RACK_PACE_TO_FILL:
19747                 optval = rack->rc_pace_to_cwnd;
19748                 if (optval && rack->r_fill_less_agg)
19749                         optval++;
19750                 break;
19751         case TCP_RACK_NO_PUSH_AT_MAX:
19752                 optval = rack->r_ctl.rc_no_push_at_mrtt;
19753                 break;
19754         case TCP_SHARED_CWND_ENABLE:
19755                 optval = rack->rack_enable_scwnd;
19756                 break;
19757         case TCP_RACK_NONRXT_CFG_RATE:
19758                 optval = rack->rack_rec_nonrxt_use_cr;
19759                 break;
19760         case TCP_NO_PRR:
19761                 if (rack->rack_no_prr  == 1)
19762                         optval = 1;
19763                 else if (rack->no_prr_addback == 1)
19764                         optval = 2;
19765                 else
19766                         optval = 0;
19767                 break;
19768         case TCP_RACK_DO_DETECTION:
19769                 optval = rack->do_detection;
19770                 break;
19771         case TCP_RACK_MBUF_QUEUE:
19772                 /* Now do we use the LRO mbuf-queue feature */
19773                 optval = rack->r_mbuf_queue;
19774                 break;
19775         case TCP_TIMELY_DYN_ADJ:
19776                 optval = rack->rc_gp_dyn_mul;
19777                 break;
19778         case TCP_BBR_IWINTSO:
19779                 optval = rack->rc_init_win;
19780                 break;
19781         case TCP_RACK_TLP_REDUCE:
19782                 /* RACK TLP cwnd reduction (bool) */
19783                 optval = rack->r_ctl.rc_tlp_cwnd_reduce;
19784                 break;
19785         case TCP_BBR_RACK_INIT_RATE:
19786                 val = rack->r_ctl.init_rate;
19787                 /* convert to kbits per sec */
19788                 val *= 8;
19789                 val /= 1000;
19790                 optval = (uint32_t)val;
19791                 break;
19792         case TCP_RACK_FORCE_MSEG:
19793                 optval = rack->rc_force_max_seg;
19794                 break;
19795         case TCP_RACK_PACE_MAX_SEG:
19796                 /* Max segments in a pace */
19797                 optval = rack->rc_user_set_max_segs;
19798                 break;
19799         case TCP_RACK_PACE_ALWAYS:
19800                 /* Use the always pace method */
19801                 optval = rack->rc_always_pace;
19802                 break;
19803         case TCP_RACK_PRR_SENDALOT:
19804                 /* Allow PRR to send more than one seg */
19805                 optval = rack->r_ctl.rc_prr_sendalot;
19806                 break;
19807         case TCP_RACK_MIN_TO:
19808                 /* Minimum time between rack t-o's in ms */
19809                 optval = rack->r_ctl.rc_min_to;
19810                 break;
19811         case TCP_RACK_EARLY_SEG:
19812                 /* If early recovery max segments */
19813                 optval = rack->r_ctl.rc_early_recovery_segs;
19814                 break;
19815         case TCP_RACK_REORD_THRESH:
19816                 /* RACK reorder threshold (shift amount) */
19817                 optval = rack->r_ctl.rc_reorder_shift;
19818                 break;
19819         case TCP_RACK_REORD_FADE:
19820                 /* Does reordering fade after ms time */
19821                 optval = rack->r_ctl.rc_reorder_fade;
19822                 break;
19823         case TCP_BBR_USE_RACK_RR:
19824                 /* Do we use the rack cheat for rxt */
19825                 optval = rack->use_rack_rr;
19826                 break;
19827         case TCP_RACK_RR_CONF:
19828                 optval = rack->r_rr_config;
19829                 break;
19830         case TCP_HDWR_RATE_CAP:
19831                 optval = rack->r_rack_hw_rate_caps;
19832                 break;
19833         case TCP_BBR_HDWR_PACE:
19834                 optval = rack->rack_hdw_pace_ena;
19835                 break;
19836         case TCP_RACK_TLP_THRESH:
19837                 /* RACK TLP theshold i.e. srtt+(srtt/N) */
19838                 optval = rack->r_ctl.rc_tlp_threshold;
19839                 break;
19840         case TCP_RACK_PKT_DELAY:
19841                 /* RACK added ms i.e. rack-rtt + reord + N */
19842                 optval = rack->r_ctl.rc_pkt_delay;
19843                 break;
19844         case TCP_RACK_TLP_USE:
19845                 optval = rack->rack_tlp_threshold_use;
19846                 break;
19847         case TCP_RACK_PACE_RATE_CA:
19848                 optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
19849                 break;
19850         case TCP_RACK_PACE_RATE_SS:
19851                 optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
19852                 break;
19853         case TCP_RACK_PACE_RATE_REC:
19854                 optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
19855                 break;
19856         case TCP_RACK_GP_INCREASE_SS:
19857                 optval = rack->r_ctl.rack_per_of_gp_ca;
19858                 break;
19859         case TCP_RACK_GP_INCREASE_CA:
19860                 optval = rack->r_ctl.rack_per_of_gp_ss;
19861                 break;
19862         case TCP_BBR_RACK_RTT_USE:
19863                 optval = rack->r_ctl.rc_rate_sample_method;
19864                 break;
19865         case TCP_DELACK:
19866                 optval = tp->t_delayed_ack;
19867                 break;
19868         case TCP_DATA_AFTER_CLOSE:
19869                 optval = rack->rc_allow_data_af_clo;
19870                 break;
19871         case TCP_SHARED_CWND_TIME_LIMIT:
19872                 optval = rack->r_limit_scw;
19873                 break;
19874         case TCP_RACK_TIMER_SLOP:
19875                 optval = rack->r_ctl.timer_slop;
19876                 break;
19877         default:
19878                 return (tcp_default_ctloutput(so, sopt, inp, tp));
19879                 break;
19880         }
19881         INP_WUNLOCK(inp);
19882         if (error == 0) {
19883                 if (TCP_PACING_RATE_CAP)
19884                         error = sooptcopyout(sopt, &loptval, sizeof loptval);
19885                 else
19886                         error = sooptcopyout(sopt, &optval, sizeof optval);
19887         }
19888         return (error);
19889 }
19890
19891 static int
19892 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
19893 {
19894         int32_t error = EINVAL;
19895         struct tcp_rack *rack;
19896
19897         rack = (struct tcp_rack *)tp->t_fb_ptr;
19898         if (rack == NULL) {
19899                 /* Huh? */
19900                 goto out;
19901         }
19902         if (sopt->sopt_dir == SOPT_SET) {
19903                 return (rack_set_sockopt(so, sopt, inp, tp, rack));
19904         } else if (sopt->sopt_dir == SOPT_GET) {
19905                 return (rack_get_sockopt(so, sopt, inp, tp, rack));
19906         }
19907 out:
19908         INP_WUNLOCK(inp);
19909         return (error);
19910 }
19911
19912 static const char *rack_stack_names[] = {
19913         __XSTRING(STACKNAME),
19914 #ifdef STACKALIAS
19915         __XSTRING(STACKALIAS),
19916 #endif
19917 };
19918
19919 static int
19920 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
19921 {
19922         memset(mem, 0, size);
19923         return (0);
19924 }
19925
19926 static void
19927 rack_dtor(void *mem, int32_t size, void *arg)
19928 {
19929
19930 }
19931
19932 static bool rack_mod_inited = false;
19933
19934 static int
19935 tcp_addrack(module_t mod, int32_t type, void *data)
19936 {
19937         int32_t err = 0;
19938         int num_stacks;
19939
19940         switch (type) {
19941         case MOD_LOAD:
19942                 rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
19943                     sizeof(struct rack_sendmap),
19944                     rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
19945
19946                 rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
19947                     sizeof(struct tcp_rack),
19948                     rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
19949
19950                 sysctl_ctx_init(&rack_sysctl_ctx);
19951                 rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
19952                     SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
19953                     OID_AUTO,
19954 #ifdef STACKALIAS
19955                     __XSTRING(STACKALIAS),
19956 #else
19957                     __XSTRING(STACKNAME),
19958 #endif
19959                     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
19960                     "");
19961                 if (rack_sysctl_root == NULL) {
19962                         printf("Failed to add sysctl node\n");
19963                         err = EFAULT;
19964                         goto free_uma;
19965                 }
19966                 rack_init_sysctls();
19967                 num_stacks = nitems(rack_stack_names);
19968                 err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
19969                     rack_stack_names, &num_stacks);
19970                 if (err) {
19971                         printf("Failed to register %s stack name for "
19972                             "%s module\n", rack_stack_names[num_stacks],
19973                             __XSTRING(MODNAME));
19974                         sysctl_ctx_free(&rack_sysctl_ctx);
19975 free_uma:
19976                         uma_zdestroy(rack_zone);
19977                         uma_zdestroy(rack_pcb_zone);
19978                         rack_counter_destroy();
19979                         printf("Failed to register rack module -- err:%d\n", err);
19980                         return (err);
19981                 }
19982                 tcp_lro_reg_mbufq();
19983                 rack_mod_inited = true;
19984                 break;
19985         case MOD_QUIESCE:
19986                 err = deregister_tcp_functions(&__tcp_rack, true, false);
19987                 break;
19988         case MOD_UNLOAD:
19989                 err = deregister_tcp_functions(&__tcp_rack, false, true);
19990                 if (err == EBUSY)
19991                         break;
19992                 if (rack_mod_inited) {
19993                         uma_zdestroy(rack_zone);
19994                         uma_zdestroy(rack_pcb_zone);
19995                         sysctl_ctx_free(&rack_sysctl_ctx);
19996                         rack_counter_destroy();
19997                         rack_mod_inited = false;
19998                 }
19999                 tcp_lro_dereg_mbufq();
20000                 err = 0;
20001                 break;
20002         default:
20003                 return (EOPNOTSUPP);
20004         }
20005         return (err);
20006 }
20007
20008 static moduledata_t tcp_rack = {
20009         .name = __XSTRING(MODNAME),
20010         .evhand = tcp_addrack,
20011         .priv = 0
20012 };
20013
20014 MODULE_VERSION(MODNAME, 1);
20015 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20016 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);